Knowledge base

Tags: paper-2, Technology Illusion, Capability Source: Bain & Company — Technology Report 2026 (Syed Ali, Frank Ford, Matthew Leybold, Alexandra Juegelt, Dana Aulanier) URL: https://www.bain.com/insights/cybersecuritys-new-ai-imperative-attacking-the-backlog-of-vulnerability-alerts-technology-report-2026/ Date published: 2026-09-29 Date flagged: 2026-11-07

Instrument (recorded before the finding, per CLAUDE.md)

Data source: undisclosed. The piece refers once to "The CISOs we surveyed" and otherwise attributes figures to "one leader," "the CISO at one large enterprise," and "Bain analysis." No survey name, no sample size, no field dates, no selection rule, and no respondent counts are given anywhere — Figure 2 states that bubble size reflects how many CISOs named each risk and then prints no counts.

Inference method: none stated. The headline quantities are reported as upper bounds ("as much as 20% to 25%"), as a typical direction ("typically by a double-digit percentage"), or as single unnamed instances ("one leader tripled the size of its remediation team"). Figure 1's attack timings are labelled "estimated."

Observation window: not stated. Published 2026-09-29 as part of Bain's Technology Report 2026.

The limit is the entry, so it is stated once and plainly: this is a high-credibility publisher carrying numbers with no disclosed instrument behind them. Bain sits on the high-credibility publisher list in both CLAUDE.md files, and that list is about editorial standards, not about whether a given figure has a method. These figures do not. They must never be cited as measurement of how much capacity organizations are spending to absorb AI output, in either direction, and the reason is Brandon's own precedent — Paper 3 retired the bare "70% of transformations fail" figure because nobody could check its numerator. "As much as 20% to 25%" from an unnamed survey of unnamed size is the same object at an earlier stage of its life.

Summary

Bain argues that AI-powered scanning has broken the economics of vulnerability management from both ends. On the attacker side, Figure 1 ("Bain analysis," timings "estimated") puts the duration of an attack chain falling from an estimated four weeks to about 18 hours, with individual stages dropping from 60–168 hours pre-AI to 1–11 hours with AI; reconnaissance falls from 168 hours to one, lateral movement from 168 to three. On the defender side, AI-powered scanning can raise the volume of vulnerability alerts "by as much as eightfold," and "more than half" of an organization's vulnerabilities are found outside its critical applications.

The organizational response Bain reports is the part this base has been looking for. Leaders are redirecting "as much as 20% to 25%" of their cybersecurity human resources away from other work to remediate alerts coming out of AI-powered scans, using third-party partners to hold normal operations together. One leader "tripled the size of its remediation team." Remediation budgets have been raised "typically by a double-digit percentage." Separately, the CISO at one large enterprise is said to enforce an internal guideline that developers spend "around 30%" of their time on proactive vulnerability management — which the piece presents as general DevOps practice rather than an AI-driven reallocation.

What the piece does not supply is any account of where these numbers come from. There is no stated methodology for any of them, and the strongest-sounding figures are the ones attributed to the fewest identified sources.

Five Breakpoints Intersection

Technology Illusion: Bain reports that AI-powered scanning can increase alert volume eightfold and that the organizational response is to redirect up to a quarter of the security function and raise remediation budgets by double digits — which is a description of capability deployed onto an operating model that cannot absorb its output at the rate the capability produces it, with the gap closed by buying and reallocating people rather than by the technology delivering a net reduction in work.

A flow mapping was considered and is held rather than tagged. An alert backlog is, on its face, structural friction blocking execution, and the second breakpoint would be the obvious label. It is held because every quantity that would make the mapping concrete — backlog size, time-to-remediate, queue length — is absent from the piece, which gives no backlog size and no remediation-time metric at all. A mapping whose only supporting number is an undisclosed-method upper bound is exactly the speculative tag both CLAUDE.md files prohibit.

The remaining breakpoints have no mapping here. Nothing in the piece speaks to outcome clarity, to internal stakeholder incentives, or to reinforcement decay.

Four Forces Intersection

Capability — as flow, and asserted rather than measured. The claim that a fifth to a quarter of a security function is being redirected to absorb AI-generated alert volume is a Capability-as-flow claim in its purest form: the organization's capacity to act on what the technology produces is the binding constraint, and it is being met by reallocation. The force is implicated by the claim; it is not evidenced by an instrument, and this entry exists partly to keep those two things apart.

The other three forces are not tagged. There is no construct in the piece bearing on outcome clarity, on incentive alignment under pressure, or on sustained reinforcement. The source was read, so these are judgments rather than honest nulls.

Relevance to Brandon's Positioning

Its value to Brandon is as a specimen, and the specimen is of a failure mode he has already made his name retiring. The 2026-11-06 scan recorded that the market has begun asserting the review-capacity problem with numbers that have no primary behind them — a ServiceNow "33-to-1 production-to-validation gap," a "20–40% needing review" attributed to Gartner 2024, a "73% fewer errors" attributed to MIT Sloan 2025, not one of them verifiable. This piece is the fourth member of that set and the most dangerous one, precisely because Bain's name will carry it where the other three could not go. The absorber's cost is being quantified in public faster than anyone is measuring it, and the figures in circulation are converging on a plausible-sounding band that no instrument produced.

So the routing is a prohibition rather than a citation. Brandon's argument that organizations do not price what AI deployment costs the functions downstream is correct and is better served by the two honest data points in the base — Bank of America's named CEO putting 1,000 contractors on a 3,000-person function on the record, and Microsoft publishing install telemetry while publishing no capacity figure at all — than by a consultancy band with no method. Citing Bain's 20–25% would hand a hostile reader the exact move Paper 3 earned credibility by refusing, and the McAdam standard in knowledge/thesis-challenges.md is the challenge that would collect the damage.

It does one further useful thing: it bounds the story thread Does anyone design for verification cost? On that thread's criterion (c) — an instrument measuring the absorbing function's headcount, spend or utilization — this is not a hit. It is evidence that the quantity is now being published without being measured, which converts the thread's target from find the instrument to find the instrument before the asserted band hardens into a fact.

Sources & Confidence

• Bain & Company, Cybersecurity's New AI Imperative: Attacking the Backlog of Vulnerability Alerts, Technology Report 2026, 2026-09-29 — High as a publisher and as evidence of what the consultancy market is now claiming; Low as an instrument, on the undisclosed-methodology grounds set out above. Both ratings are intended.
• Cross-referenced: bofa-moynihan-1000-contractors-ai-found-flaws-2026-10.md (Moderate — named CEO, spoken estimate, bundled cost) and microsoft-vulnerability-management-ai-era-2026-11.md (High on first-party telemetry).

Filed by / Routing

Filed by: Stafford (judgment layer, 2026-11-07), from its own exploratory query — not from the 12:00 scan's candidate pool. Routing: strategy — carries a standing do-not-cite on its headline figures.

Tags: paper-2, Technology Illusion, Process Friction, Capability Source: Microsoft — Windows Experience Blog (Nuno Costa, Partner Director, Windows Servicing and Delivery) URL: https://blogs.windows.com/windowsexperience/2026/11/05/a-new-standard-for-vulnerability-management-in-the-ai-era/ Date published: 2026-11-05 Date flagged: 2026-11-07

Instrument (recorded before the finding, per CLAUDE.md)

Data source: two first-party Microsoft figures in one vendor post. (1) Microsoft's own Windows diagnostic data on security-update installation across "a measured Windows enterprise sample," August 2026. (2) Microsoft's own count of CVEs addressed in its October 2026 monthly Windows security update. Neither is a survey and neither is self-report by a third party — the installation figure is telemetry the vendor collects from running devices, which is the genre this base is scarcest in.

Inference method: descriptive shares only. No comparison group, no pre/post design, no causal estimate of any AI effect. The post reports what share of devices had installed an update by a given elapsed time; it does not attribute that share to AI, to the CVE volume, or to anything else.

Observation window: telemetry August 2026; CVE count October 2026; post published 2026-11-05.

Five limits that must travel with every use. (1) No sample size is disclosed — "a measured Windows enterprise sample" is not defined, bounded or dated beyond the month, so the shares cannot be given a confidence interval or a population. (2) Patch-installation latency long predates AI and this post contains no counterfactual — nothing here establishes that AI-assisted discovery changed install behavior, and the entry must never be cited for that. (3) Vendor reporting on its own customers' patching carries a clear interest — Microsoft is arguing that customers should patch faster, so the incentive runs toward reporting adoption as worse; the direction of any bias is knowable, which is better than unknowable, but it is not neutral. (4) MDASH is Microsoft's own agentic scanner. The post does not place Microsoft in Project Glasswing and does not name Claude; Anthropic appears only in footnote 1, as a cited research article on N-day exploit risk, not as a tool Microsoft used. Any claim that Microsoft was in Mosk's treated cohort would be invention. (5) Read via automated page extraction with verbatim quotation of the load-bearing sentences, not a full manual read of the post; the quotes below are exact, the surrounding emphasis is this entry's.

Summary

Nuno Costa, Partner Director for Windows Servicing and Delivery, sets out how Microsoft now finds vulnerabilities in Windows and what it believes that changes. The discovery side is explicitly AI-assisted: the post describes MDASH, "an agentic code scanner that uses multiple AI models and specialized agents," which finds and validates potential vulnerabilities. The volume figure attached to the new regime is that Microsoft's October 2026 monthly security update addressed more than 1,100 CVEs for Windows.

Costa then states the constraint himself, and it is the sentence this entry exists for: "Finding vulnerabilities more quickly only helps if engineering processes can keep pace." On who remains accountable, the post is equally direct: "Engineers remain responsible for reviewing the work, making the engineering decisions, and owning the final output," and the scanner's role is framed as helping "engineers investigate and remediate issues more efficiently."

The post also publishes a figure about what happens after the fix exists. From August 2026 diagnostic data across a measured Windows enterprise sample, fewer than half of devices had installed the security update two weeks after release, and about one in four had still not installed it after four weeks. The post supplies no headcount, team size, remediation budget or reviewer-capacity figure anywhere — it names the pacing problem and quantifies the step after it, while leaving the capacity of the step in between entirely unquantified.

Five Breakpoints Intersection

Technology Illusion: Microsoft reports a working capability on the discovery side (an agentic scanner contributing to 1,100-plus CVEs addressed in a single month) and a working fix side (the update shipped), and then reports from its own telemetry that fewer than half of enterprise devices had installed that update two weeks later — so the security benefit the capability produced is realized only by the organizations that deploy it, and at the point of deployment, which is entirely organizational behavior rather than technical capability, more than half the estate had not moved. The technology was ready; the condition it landed on was not.

Process Friction: Costa's own sentence — "Finding vulnerabilities more quickly only helps if engineering processes can keep pace" — is a vendor naming structural friction between an accelerated upstream and an unchanged downstream, and the install telemetry puts a number on friction one step further down again, where the completed fix sits unapplied on a quarter of devices a month after release.

The fifth breakpoint was considered and is not tagged, and the reason is recorded because the temptation was real. A shipped-but-uninstalled patch looks like a textbook case of progress that has not actually happened. It is declined because that breakpoint's mechanism is the decay of transformation energy as novelty fades, and there is no transformation, no initiative and no time-series of effort in this data — only a static adoption share. Tagging it would be the loose tagging both CLAUDE.md files prohibit.

The remaining two breakpoints have no mapping here. Nothing in the post speaks to outcome clarity or to stakeholder incentives.

Four Forces Intersection

Capability — and the post measures both halves against the same deployment, which is rare. Readiness: the scanner was usable, with engineers retained as the accountable reviewers, so nothing here suggests the capability arrived unready. Flow: Microsoft states that discovery speed only helps if engineering processes keep pace, and then publishes telemetry showing that the furthest-downstream step in that chain — installation on the customer estate — is where the chain stalls, with over half of devices unpatched at two weeks.

The other three forces are not tagged. The unit of observation is a device-installation share and a CVE count. There is no construct here bearing on outcome clarity, on incentive alignment under pressure, or on sustained reinforcement, and these omissions are judgments from a read source rather than honest nulls.

Relevance to Brandon's Positioning

This is the step past where the base's best counterargument stops measuring, and it is the strongest single piece of evidence in this base for the thesis sentence as written. Mosk's Project Glasswing paper (mosk-glasswing-frontier-ai-cyber-defense-2026-11.md, flagged 2026-11-07) is the cleanest live challenge to the thesis: a frontier capability landed in a group of large US tech firms with no reported operating-model redesign, more than doubled vulnerability disclosures, and "vendors took no longer to fix reported flaws." The 2026-11-07 scan brief read that as a control case and proposed conceding a scope condition on Claim 5 — where AI's output is self-validating, capability is sufficient. Microsoft's post is the reason not to concede it. Mosk's right-hand side is the public disclosure record, which ends at the vendor's fix. Microsoft publishes the next link: the fix exists, and the estate does not take it. A patched vulnerability is not self-validating in outcome terms — its value is captured only when organizations whose change windows, reboot tolerance and testing cycles differ actually install it. So vulnerability management is not the domain where coordination stops mattering; it is a domain where the base had simply never looked past the vendor.

Against the base's live chain this supplies the missing third point. Discovery: AI roughly doubles output (Mosk, about +36 disclosures per firm-month; Microsoft, 1,100-plus CVEs in one month). Fix: no measured slowdown (Mosk), bought at a 33% capacity expansion in the one firm that said a number out loud (bofa-moynihan-1000-contractors-ai-found-flaws-2026-10.md), and asserted without an instrument by consultancy (bain-vulnerability-alert-backlog-2026-09.md, flagged today). Install: fewer than half within two weeks, about one in four still unpatched at four weeks — first-party telemetry, no survey, no self-classification. The compression is upstream and the binding constraint is downstream and organizational, which is the thesis in one chain.

It also sharpens the story thread Does anyone design for verification cost? in a way worth stating precisely. On that thread's criterion (c) — an instrument measuring the absorbing function's headcount, spend or utilization — this post is another NULL, and a louder one than the AWS specification the scan found on 11-07, because AWS was a design document with no deployment behind it while Microsoft has a live deployment at Windows scale, names the pacing problem in its own words, and still publishes no capacity number. The pattern is now that the absorber's cost goes unquantified by the parties best placed to quantify it.

Sources & Confidence

• Microsoft, Windows Experience Blog, A new standard for vulnerability management in the AI era, Nuno Costa, 2026-11-05 — High as to what Microsoft states and as a first-party telemetry report; Moderate as to the generalizability of the installation shares, because no sample size, population or selection rule is disclosed.
• Mosk, Does Frontier AI Strengthen Cyber Defense? Evidence from Project Glasswing, arXiv:2611.06629, 2026-11-04 — Moderate (non-refereed preprint, vendor-allocated treatment); cited here only for the fix-time null, which is verified in the abstract.
• Bank of America Q3 2026 earnings call via ITPro, 2026-10-15 — Low to Moderate (a CEO's spoken estimates, bundled cost).
• Bain, Cybersecurity's New AI Imperative, 2026-09-29 — High publisher, Low instrument (no methodology disclosed); see that entry's warning.

Filed by / Routing

Filed by: Stafford (judgment layer, 2026-11-07), from its own exploratory query — not from the 12:00 scan's candidate pool. Routing: strategy — it is the reason to hold Claim 5 rather than concede a scope condition to Mosk this week.

Tags: paper-2, practitioner, Strategic Disconnection, Incentive Fragmentation, COUNTERARGUMENT Source: Implementation Science 17:75 — Laura J. Damschroder, Caitlin M. Reardon, Marilla A. Opra Widerquist, Julie Lowery (VA Center for Clinical Management Research, Ann Arbor) URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC9617234/ Date published: 2022-10-29 Date flagged: 2026-11-05

Summary

Damschroder and colleagues revise the Consolidated Framework for Implementation Research using feedback from the people who had used it. Two inputs: a literature review that found 376 articles with the CFIR in the title or abstract, of which 59 contained feedback on the framework, and a survey emailed in August 2020 to the 334 unique corresponding authors of those articles. 157 (47%) responded and 134 (40%) completed the survey. Respondents were experienced users — 83.1% reported healthcare settings, over 65% had used the CFIR on at least two projects and nearly 20% on five or more.

The usability result is reported in the paper's own survey table. Asked whether the CFIR is "easy for implementation researchers to use," 64 of 127 respondents (50.4%) said yes. Asked whether it is "easy for non-implementation researchers to use," 20 of 126 (15.9%) said yes. The paper rounds these to 50% and 16%. Free-text responses include "CFIR is far too complicated and difficult to use" and, on the Culture construct, that it "ends up becoming my 'I don't know where else this fits' bucket." The authors also record a dissenting user: "Implementation research is challenging in itself, and I see that the complexity of CFIR gets blamed" for the field's broader difficulty. The authors note that respondents often attributed the complexity to the framework's size, but that many of the same respondents identified missing themes — the critique ran in both directions at once.

The authors' response to a documented usability complaint was to add content, not to reduce it. The number of constructs increased in every domain except the Innovation Domain. The updated CFIR carries 48 constructs and 19 subconstructs across five domains (Innovation, Outer Setting, Inner Setting, Individuals, Implementation Process), against the 2009 original's 37 constructs across five domains. Some constructs were removed or re-sited rather than dropped: Implementation Climate and Readiness for Implementation were reclassified as antecedent assessments, Patient Needs and Resources was split into three constructs, External Policies and Incentives was separated into several, and Formally Appointed Internal Implementation Leader and Champion were combined into Implementation Leads. Construct names verified against the authors' own construct guide at cfirguide.org/constructs.

Five Breakpoints Intersection

Strategic Disconnection: the updated prior art now has a construct named for it, which the 2009 version did not. CFIR 2.0's Inner Setting domain adds Mission Alignment and Relative Priority as named constructs; the 2009 original's nearest equivalent was goals and feedback under implementation climate, which the base's existing entry recorded as "adjacent to, but narrower than, Strategic Disconnection." The adjacency argument that entry rested on is weaker against the 2022 version than against the 2009 one.

Incentive Fragmentation: the construct survives the revision under a cleaner name. The 2009 organizational incentives and rewards is carried into CFIR 2.0's Inner Setting domain as Incentive Systems, so the prior art for this breakpoint is not merely a 1975 antecedent but a current, maintained construct in the field's most-used determinant framework.

On the sustainment breakpoint — considered, declined as a tag, and the absence is the finding, which survives the revision. Verified against the authors' own construct list: none of the 48 constructs or 19 subconstructs concerns sustainment, sustainability, maintenance or continuation after implementation. The nearest, Adapting, covers modification during implementation. This matters more than the 2009 absence did, for two reasons. First, the revision was explicitly driven by users telling the authors what was missing, and sustainment was not among the gaps they closed. Second, the authors cite the sustainment literature in the same paper — including "Evidence-based intervention sustainability strategies: a systematic review" and "Implementation and sustainment of diverse practices in a large integrated health system" — so the omission is a scope decision taken with the literature in hand, not an oversight.

Two further mappings considered and declined (structural flow, and technology-on-top-of-broken-conditions). Structural Characteristics and its subconstructs (Physical Infrastructure, Information Technology Infrastructure, Work Infrastructure) are resource-adequacy dimensions, not flow blockage; tagging them would repeat the altitude error the 2026-11-04 Burke–Litwin entry declined on Equipment.

Four Forces Intersection

No Force tagged, and the reason is the same one the 2009 CFIR entry gave. The 16%/50% usability split is a measurement of practitioners' facility with a research instrument, not of an organization's capacity to execute change, and tagging Capability on it would move the Force a level away from where this framework uses it. The sustainment Force is not tagged on that absence, following the standing discipline in the 2009 CFIR and 2026-11-04 Burke–Litwin entries: the absence is recorded as the finding and not filled with a guessed tag.

Relevance to Brandon's Positioning

This source does two opposite things to asset 12 (master-plan.md Tier 2.1, the gate) on the same day, and both matter.

It supplies the quantified usability anchor the asset's new spine needed. Brandon's 2026-11-05 amendment reframed asset 12 around assembly-and-usability rather than novelty, and conceded the claim is "currently unsupported." It now has support: 16% of 126 experienced users say the nearest integrated prior art is usable by a non-specialist, reported by the framework's own authors, in the field's own journal, on a sample of people who had published with it. That is a far harder number than the qualitative ratings the amendment currently cites, and it is the first quantified instance in the base of prior art failing on operator usability.

And it removes the asset's designated best exhibit, by showing the field does not share the premise. The amendment calls Martins & Coetzee's Table 1 "the asset's single best exhibit" on the reading that the genre rates none of its models usable. Given a documented usability complaint from 134 users and a full revision, CFIR's authors grew the framework from 37 to 48 constructs plus a new subconstruct layer. That is an expert panel looking straight at the complaint and concluding diagnostic coverage beats parsimony — a credentialed, dated, deliberate rejection of the simplification move. The asset cannot claim the field agrees that its frameworks are unusable; it has to argue against a field that was told and chose otherwise.

The net is a narrowing, and the narrowed claim is stronger. The confirmed absence of any sustainment construct across both CFIR generations gives Brandon the one content claim his 11-05 note says he has left, now verified against the updated prior art rather than the 2009 version. But it also bounds it: a sustainment literature demonstrably exists and CFIR's authors cite it. So the defensible claim is not nobody has a sustainment construct — it is the integrated transformation-level diagnostics have no sustainment construct, while the sustainment literature that exists has not been integrated into them. This partly discharges the sustainability-literature check dated 2026-11-11 and tells it exactly where to look: the three references named above.

Sources & Confidence

• Damschroder, L.J., Reardon, C.M., Opra Widerquist, M.A. & Lowery, J. (2022). The updated Consolidated Framework for Implementation Research based on user feedback. Implementation Science 17:75. DOI 10.1186/s13012-022-01245-0. PMID 36309746; PMCID PMC9617234. Received 2022-04-21, accepted 2022-10-06, published 2022-10-29. Confidence: High — peer-reviewed, open access, authors are the framework's originators.
• CFIR construct list, cfirguide.org/constructs (the authors' own guide). Confidence: High for the construct and subconstruct names and for the sustainment absence, which was verified against the full enumerated list rather than inferred from the article text.
• Martins, N. & Coetzee, M. (2009), SA Journal of Human Resource Management 7(1) a177 — the exhibit this entry bears on. Confidence: Moderate, and see the 2026-11-05 thesis-challenges update: its Table 1 was re-read in the primary today and does not say what this base recorded.
• Read bound, stated honestly: the PMC full text is ~159,000 characters and the first ~100,000 were read. The survey counts, usability percentages, verbatim quotes, construct totals and the reorganization list are all from the read portion. The sustainment absence does not rest on the unread remainder — it was verified against the authors' complete construct enumeration at cfirguide.org. The three cited sustainment references were located in the reference list and have not themselves been read.

Filed by / Routing

Filed by: Stafford (judgment layer, exploratory slot, 2026-11-05) Routing: strategy — load-bearing for master-plan.md Tier 2.1 (asset 12, the gate) and for the usability limb of the relabeling objection in knowledge/thesis-challenges.md

Tags: paper-2, practitioner, Strategic Disconnection, Incentive Fragmentation, Process Friction, Purpose, Capability, COUNTERARGUMENT Source: SA Journal of Human Resource Management 7(1), Article 177 (OpenJournals) — Nico Martins & Melinde Coetzee, Department of Industrial Psychology and People Management, University of South Africa URL: https://sajhrm.co.za/index.php/sajhrm/article/view/177 Date published: 2009-09-11 Date flagged: 2026-11-04

Summary

Martins and Coetzee set out to test whether the Burke–Litwin model of organisational performance and change is useful as a diagnostic framework for identifying the factors that affect organisational effectiveness. The study is explicitly exploratory and qualitative: purposive sampling at a single international hotel whose workforce represented more than 17 nationalities, with 147 employees in focus groups and 11 executive managers in semi-structured probing interviews. Themes from both data sets were then categorised against the model's variables, and the populated model was used to report results back to participants.

The article's description of the model is the material part for this base. It reproduces Burke and Litwin's twelve variables and their division into two classes. Transformational factors: external environment; vision, mission and strategy; leadership; organisational culture; and individual and organisational performance. Transactional factors: structure; management practices; systems (policies and procedures); departmental/work unit climate; task requirements and individual skills/abilities; individual needs and values; and motivation. The causal ordering is stated directly — "the external environment affects transformational factors," which "in turn, affect the transactional factors," and "Both types of factors reciprocate, and eventually impact on, individual and organisational performance and overall effectiveness." The definitions are specific: vision/mission/strategy is "What employees believe to be the central purpose of the organisation"; structure covers "levels of responsibility, decision-making authority and relationships"; systems "primarily manifest themselves in the organisation's reward and control systems"; motivation is "Aroused behavioural tendencies to move toward goals, take needed action and persist until satisfaction is attained." The authors add two transactional dimensions of their own after diagnosis — Equipment ("The tools to do the job and the quality of available technology") and Working environment.

The article also carries Table 1, "Comparing diagnostic models," a peer-reviewed comparison of seven organisational diagnostic models — Weisbord, Nadler–Tushman congruence, Freedman's swamp model of sociotechnical systems, McKinsey seven-S, Galbraith STAR, and Burke–Litwin — scored on when to apply, strengths, limitations and overall evaluation. Burke–Litwin's listed strengths are "Explains linkages," "Shows cause-and-effect relationships between organisation's internal and external environments," and "Distinguishes between the role of transformational and transactional dynamics in organisational behaviour and change"; its only listed limitation is "Complexity." Its evaluation column reads "Well-researched" and "Empirical evidence of addition of value to outcomes of an organisational diagnostic process." On the model's intended use the authors write that it "helps OD practitioners to size up the change situation, to determine the kind of change (whether transformational or transactional) required and then to target interventions toward factors of the organisation that will produce the desired change," and that prior research "suggests that the model performs as intended." The study's own stated limitations include the use of a single model and a single organisation.

Five Breakpoints Intersection

Strategic Disconnection: The model's transformational variable for purpose is defined not as the stated strategy but as "What employees believe to be the central purpose of the organisation" — the gap between declared direction and what the organisation actually believes it is pursuing, which is the condition Strategic Disconnection names, given a dedicated named variable in a 1992 causal model.

Incentive Fragmentation: Burke–Litwin's Systems variable is defined as standardised policies and mechanisms that "primarily manifest themselves in the organisation's reward and control systems," so misaligned reward structure is not an insight this framework adds — it is an existing box in the prior model, positioned as a transactional cause of climate and motivation.

Process Friction: Structure is defined in terms of "levels of responsibility, decision-making authority and relationships" and Systems as "mechanisms that facilitate work" — structural friction that blocks execution has named upstream variables here, which is a harder result for the framework's originality claim than CFIR produced, because CFIR had no construct for structural process friction at all.

Technology Illusion: considered and declined. The authors add Equipment — "The tools to do the job and the quality of available technology" — but as a resource-adequacy dimension surfaced by hotel staff, not as a claim that technology is deployed on top of broken organisational conditions. Tagging it would be the altitude error.

Momentum Mirage: considered and declined, and this is the finding that matters. The nearest variable is Motivation, defined as "Aroused behavioural tendencies to move toward goals, take needed action and persist until satisfaction is attained" — individual behavioural persistence inside a task, not a change effort losing force after launch while continuing to report progress. There is no sustainment or post-launch decay construct anywhere in the twelve variables. Momentum Mirage is outside this taxonomy, as it was outside CFIR.

Four Forces Intersection

Purpose. The model places purpose upstream of everything transactional and defines it as employee belief rather than executive declaration, which is the Purpose construct as this framework uses it, named and positioned causally seventeen years before Paper 2.

Capability. Task requirements and individual skills/abilities is a named variable covering "the behaviour, including specific skills and knowledge required for task effectiveness" and the job-person match; in the hotel data the two highest-frequency change themes were individual and organisational performance (178 respondents) and task requirements and individual skills (100), so capability deficits dominated what staff wanted changed.

Commitment and Momentum: not tagged. Nothing in the twelve variables observes follow-through under organisational tradeoff, and the motivation variable sits at individual rather than change-effort altitude (see above). The empirical leg is one hotel, qualitative and exploratory, so none of these tags rests on the study's own findings.

Relevance to Brandon's Positioning

This is the strongest single piece of prior art against the relabeling objection that the base has acquired, and it breaks the specific defence installed on 2026-10-11. When CFIR broke the "nobody integrated them" and "none of them diagnoses" limbs, the replacement defence was that CFIR "presents its constructs without depicting interrelationships" while Five Breakpoints is a claim about interaction and sequence. Burke–Litwin is titled A Causal Model; it specifies direction (environment → transformational → transactional → performance), claims reciprocity between the two classes, and a peer-reviewed comparative table credits it with "Shows cause-and-effect relationships." The interaction-and-sequence defence does not survive it. The 2026-10-11 run predicted exactly this shape — "a causal model with arrows between twelve variables is the one shape that would also threaten the interaction claim now carrying the section" — and named Burke–Litwin as the next item to run.

What survives, and it is narrower and better than what it replaces. Three things. First, Momentum Mirage is outside this taxonomy and outside CFIR's, on an altitude argument that is specific and checkable rather than a blanket originality claim. Second, no account of machine speed — this is 1992, and the diagnostic application is 2009. Third, and most usefully, the McAdam standard indicts Burke–Litwin on exactly the terms it indicts Five Breakpoints: the diagnostic application here is one hotel, purposive sampling, 147 focus-group participants and 11 interviews, no comparison group and no transformation outcome; the separate psychometric work (an Organisational Performance Questionnaire based on the model, n = 398, South Africa) validates a questionnaire's structure, not the causal model against measured transformation results. This extends the 2026-11-02 Garicano & Rayo concession move from organisational economics to the OD diagnostic genre: the prior art is older, integrated, causal, diagnostic — and equally untested against outcomes.

Direct consequence for master-plan.md Tier 2.1 (asset 12, the gate). The section cannot be written as "nobody integrated them, nobody diagnosed them, nobody specified the interaction." All three are false. It has to be written as the Paper 3 seventy-percent move: name Burke–Litwin (1992), CFIR (2009) and Garicano & Rayo (2016), concede integration, diagnosis and causal specification, and stake the contribution on sustainment altitude plus machine speed plus a measurement the prior art also lacks. Table 1 of this article is itself a drafting asset — a peer-reviewed comparison of seven diagnostic models that discharges the McKinsey 7-S candidate named on 2026-09-17 and adds four more (Weisbord, Nadler–Tushman, Freedman, Galbraith STAR) to the prior-art map at no further search cost.

Sources & Confidence

• Martins, N. & Coetzee, M. (2009). Applying the Burke–Litwin model as a diagnostic framework for assessing organisational effectiveness. SA Journal of Human Resource Management 7(1), a177. DOI 10.4102/sajhrm.v7i1.177. Open access, full text read via the journal's own PDF. Confidence: Moderate — peer-reviewed and the primary was read in full, but a regional HR journal and an exploratory single-organisation qualitative design.
• Burke, W.W. & Litwin, G.H. (1992). A Causal Model of Organizational Performance and Change. Journal of Management 18(3), 523–545. DOI 10.1177/014920639201800306. Confidence: High for the citation and for the model's structure as reported by Martins & Coetzee. The primary itself was NOT read — the accessible copy (a ProQuest scan) carries no text layer and this container has no OCR, so every statement about the model's content here is sourced to Martins & Coetzee's reproduction of it, not to Burke & Litwin directly. Reading the primary remains owed.
• Psychometric validation of an Organisational Performance Questionnaire based on the Burke–Litwin model (n = 398, South Africa). Confidence: Low — identified from search metadata only, not retrieved or read. Cited here as an unverified pointer, not as evidence.

Filed by / Routing

Filed by: Stafford (judgment layer, Wednesday falsification sweep, 2026-11-04) Routing: strategy — load-bearing for master-plan.md Tier 2.1 and for the relabeling objection in knowledge/thesis-challenges.md

Tags: paper-2, practitioner, Technology Illusion, Strategic Disconnection, COUNTERARGUMENT Source: arXiv preprint 2512.05533v2 (cs.SE), accepted to the Journal of Systems and Software — Theocharis Tavantzis, Stefano Lambiase, Daniel Russo, Robert Feldt URL: https://arxiv.org/abs/2512.05533 Date published: 2025-12-05 (v1); v2 last revised 2026-08-05 Date flagged: 2026-11-04

Summary

The authors argue that research on AI adoption in software engineering has concentrated on technical concerns and underplayed the human-centred ones, and they propose a framework, informed by Behavioral Software Engineering (BSE), to support software organisations during early-stage AI adoption. The framework has nine dimensions — AI Strategy Design, AI Strategy Evaluation, Collaboration, Communication, Governance and Ethics, Leadership, Organizational Culture, Organizational Dynamics, and Up-skilling — each carrying design principles and actionable steps, clustered into three overarching levers.

The methodologically significant part for this base is how the framework was built. The authors ran a literature review of organisational change management models and catalogued 43 of them (Table 1, S01–S43), including ADKAR, Armenakis et al., Beckhard & Harris, Bridges' Transition Model, Bullock & Batten, Burke & Litwin, Carnall, the Congruence Model, Cummings & Worley, GE's Change Acceleration, Hayes, Jick, Judson, Kanter et al., Knoster, Kotter, Leppitt, Lewin, the McKinsey 7-S Model, Mento et al., Tampoe and Worley & Mohrman, classified by domain (Business Excellence, Digital Transformation, Education, Generic, Healthcare) and by type (Processual vs Descriptive — Kotter is Generic/Processual, Burke & Litwin is Generic/Descriptive). They then mapped the steps and factors of those models onto previously identified BSE-related challenges of AI adoption, and report that "combining existing organizational change models can mitigate the substantial majority of the identified BSE-related challenges." Specifically the mapping "addressed all but three of the BSE-related sub-challenges, namely: 'Considerate use of external AI with company's data', 'Inter-generational Collaboration', and 'Lack of skills due to AI'." They note separately that the dominant frameworks in the literature, "such as those by Kotter's and Lewin's models, originated prior to 2000."

For practitioner feedback the authors ran a survey (N = 105) using a $100-method allocation — respondents distribute 100 points across the nine dimensions — plus two expert workshops (N = 4 total). Up-skilling (15.2%) and AI Strategy Design (15.1%) drew the highest allocations; the authors read this as organisations prioritising procedural aspects while "human-centered guardrails remain comparatively underdeveloped." Their threats-to-validity section is unusually candid and three admissions matter here. The qualitative dataset was collected at the end of 2024 from organisations "at the beginning of their AI integration strategy," so the framework is explicitly early-stage. The mapping "relied on the interpretive judgment of the first author," with no formal inter-rater reliability process — "this absence constitutes a limitation, and disagreements were instead resolved through consensus discussions." And on the survey, respondents "could disproportionately represent early adopters or individuals with strong interest in AI, potentially overstating the perceived importance of personal career development factors—such as Up-skilling—and not organizational necessities." They also concede a coverage gap in their own framework: "actions related to anchoring change, such as sustaining adoption and managing long-term organizational adaptation, are not yet represented, constituting a possible direction for future work."

Five Breakpoints Intersection

Technology Illusion: The paper's founding premise and its survey result are the breakpoint stated as a measured skew — organisations fund the procedural apparatus of AI strategy while, in the authors' words, "human-centered guardrails remain comparatively underdeveloped," which is deploying the technology ahead of the organisational conditions it depends on, reported by the people doing the deploying.

Strategic Disconnection: The highest-frequency design principle the authors extracted from their interview data, at frequency 7 and tied with "Communicate the change constantly," is "Recognize a business need before starting an AI integration project" — a principle only worth stating because practitioners were observed starting AI integration projects without an identified business need, which is vague purpose producing the appearance of direction.

Momentum Mirage: considered and declined, although the source hands over the gap. The authors concede that "anchoring change, such as sustaining adoption and managing long-term organizational adaptation," is absent from their framework. That is evidence about the coverage of the prior art, not evidence of a change effort producing the appearance of progress — so it belongs in the positioning section below, not as a breakpoint tag.

Incentive Fragmentation and Process Friction: declined. Nine dimensions include Organizational Dynamics and Governance, but nothing in the instrument observes a reward conflict or a structural blockage; the data are practitioner interviews about adoption challenges and a weighting exercise.

Four Forces Intersection

No Force tagged, deliberately, and the reason is Brandon's own rule. The only quantitative instrument here is a $100-method allocation of perceived importance across nine dimensions by 105 self-selected respondents. Under the self-classified-maturity ruling of 2026-10-28, a share produced by respondents placing weight on categories they are not asked to evidence cannot support a claim about organisational state in either direction — and the authors say so about their own data before anyone else can, warning that the Up-skilling weight may reflect "personal career development factors... and not organizational necessities." Up-skilling at 15.2% is therefore not evidence that Capability is the binding constraint, and the low governance share is not evidence that guardrails are absent. Tagging Capability off that number would be the exact error the ruling exists to prevent. The qualitative leg could in principle support a Force tag, but it is a 2024 interview dataset reused from a different study with a different research orientation, which the authors also flag.

Relevance to Brandon's Positioning

This is the paper that kills the blanket form of the last surviving defence of the relabeling objection, and it does it in print, in a peer-reviewed venue, with a count. After CFIR (2026-10-11) and Burke–Litwin (today) broke the integration, diagnosis and causal-specification limbs, the only general defence left was no account of speed — none of this prior art addresses AI. That defence is now unavailable as stated: 43 organisational change models were systematically catalogued and mapped onto AI-adoption challenges, and the authors report coverage of all but three. Worse for the blanket claim, none of the three residuals is a time-compression or machine-speed argument — they are external-AI data handling, inter-generational collaboration, and AI-driven skill gaps. A hostile reviewer now has a single citation for "classic change management already covers AI transformation, and someone checked."

Four things bound it, and they are where the remaining ground actually is. (1) Scope — software engineering organisations, a BSE lens, and explicitly early-stage adoption, built on interview data collected at the end of 2024 from organisations just beginning. (2) Instrument — the coverage claim is a conceptual mapping performed by one author with no inter-rater reliability, which is an analyst's judgment, not a measurement; it establishes that the mapping can be made, not that the models work on AI transformations. (3) No failure-mode diagnosis — this is nine dimensions of prescription with design principles and actions, not an account of how AI transformations break; it is Category 8 in the standing-sources sense, a design claim rather than evidence. (4) The authors concede the sustainment gap in their own words — anchoring change and managing long-term adaptation "are not yet represented."

That fourth point is the most valuable sentence in the day's work. Momentum Mirage is now the one breakpoint that has survived all three prior-art checks: outside CFIR's taxonomy (2026-10-11), outside Burke–Litwin's twelve variables on an altitude argument (today), and conceded absent by the authors of the most current AI-era change framework (today). That converts Brandon's originality claim from a blanket assertion into a narrow, cited, falsifiable one — and it is the claim Tier 2.1 should be built on rather than the integration claim it has been carrying.

Also a Category 8 acquisition in its own right, and a source-candidate signal. The nine dimensions are a competitor spec for prescriptive AI-native design, and what they assume solved is worth naming: Governance and Ethics and Organizational Dynamics appear as dimensions to be designed, with ownership, incentive alignment and decision rights nowhere in the framework.

Sources & Confidence

• Tavantzis, T., Lambiase, S., Russo, D. & Feldt, R. From Challenge to Change: Design Principles for AI Transformations. arXiv:2512.05533v2 [cs.SE], v1 2025-12-05, v2 2026-08-05; accepted to the Journal of Systems and Software. CC BY 4.0. Full text retrieved and read. Confidence: Moderate — accepted at a reputable venue and the primary was read in full, but a preprint copy, a narrow domain, self-selected respondents, and a central mapping with no inter-rater reliability. The acceptance statement is the arXiv page's own and was not independently confirmed at the publisher.
• The 43-model catalogue (Table 1) is reported as the authors list it; none of those 43 primaries was read for this entry, including Burke & Litwin, Kotter and Lewin. Their classification here is Tavantzis et al.'s, not this base's. Confidence: Moderate for the catalogue's existence and composition, Low for any claim about what an individual listed model says.

Filed by / Routing

Filed by: Stafford (judgment layer, Wednesday falsification sweep, 2026-11-04) Routing: strategy — rewrites the surviving defence in knowledge/thesis-challenges.md (relabeling objection) and sets what master-plan.md Tier 2.1 can claim Practice asset: none — the practice repo tgifriday7/designedtohold-practice is outside this session's GitHub scope, so the Category 8 cross-reference is owed; the nine dimensions and the three named residual challenges bear on any agent/role-design asset and should be mapped by the scan or by /improve.

Tags: paper-2 practitioner Process Friction Momentum Mirage Capability Momentum COUNTERARGUMENT Source: ANZ Journal of Surgery (Wiley) — read in full at PubMed Central (PMC13724767) URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC13724767/ Date published: 2026-08-13 Date flagged: 2026-11-03

Summary

Wells, Henderson, Varghese, O'Grady, Boyle, Bissett & Campbell report a national population-based cohort of every adult inpatient surgical admission under general or neuraxial anaesthesia in New Zealand public hospitals from 1 January 2022 to 31 December 2024: 202,428 patients, 223,415 admissions, 281,147 surgical ward shifts. 2.8% of patients died within 90 days; complications occurred in 18.7% of admissions; failure to rescue — defined in the paper's Section 2.4 as "death within 90 days of surgery following a complication" — occurred in 15.1% (6,296 of 41,722 admissions with a complication). Table 2 shows all 90-day deaths fall inside the complication group.

The exposure is the novel element. Staffing is not a hospital characteristic taken from a survey; it is "the difference between care hours required and hours available from rostered staff who worked the shift" (Section 2.3), computed shift by shift. The required figure is generated by the ward's own nurses, who record patient acuity each shift inside TrendCare, the instrument used by New Zealand's national Care Capacity Demand Management programme, implemented nationally since 2021 and validated by inter-rater reliability testing. The resulting variable — "clinical variance," in nurses — showed 43.5% of all 281,147 shifts ran understaffed (any negative variance), with a secondary threshold of "below target" set at >8.5% negative clinical variance, "corresponding to a 40 min deficit" per full-time-equivalent on an eight-hour shift. Because ward location was unavailable in the administrative records, each patient's exposure pools all surgical ward shifts in the treating hospital during the admission, weighted by each ward's required hours.

Per one-nurse reduction in staffing per shift, adjusted: 90-day mortality aOR 1.13 (95% CI 1.07–1.18); complications aOR 1.07 (1.05–1.10); failure to rescue aOR 1.07 (1.00–1.14). Correcting the observed staffing deficits was modelled to prevent 182 deaths annually. The authors conclude that "nurse staffing is a modifiable system factor associated with postoperative safety," and state plainly that "direct causal inference is not possible," framing the result as a hospital system-level association rather than a single causal pathway. Stated limitations: residual confounding; exposure misclassification from pooling across wards; no measure of skill mix or experience (healthcare assistants are counted in available hours because skill-mix data were unavailable); possible bias in nurse-entered acuity; and no capture of redeployment as a response to understaffing. On prior work they note only that earlier studies "measured nurse staffing as a static hospital-level characteristic from cross-sectional surveys" and that "population-level evidence using shift-level staffing data is limited."

Five Breakpoints Intersection

Process Friction. The paper measures a structural capacity shortfall against the organization's own standard — 43.5% of 281,147 shifts ran with fewer nursing hours than the ward's acuity-based requirement called for — and associates each one-nurse shortfall with 13% higher odds of death within 90 days; nobody in the system intends the complication to go unmanaged, so what blocks the response is structural capacity rather than purpose or will.

Momentum Mirage. Failure to rescue isolates, by construction, the class of case in which every visible process ran to completion — the operation was performed, the patient went to the ward, the shift was staffed and rostered — and the only thing that did not happen was the corrective response to a complication that had already occurred; the paper's own background names this as mortality following complications "not prevented, recognized, or managed well," which is the appearance of normal progress with the correcting loop silent.

Four Forces Intersection

Capability. The exposure variable is a capability measure in the Four Forces sense — hours of capacity required to do the work against hours actually present — and it is recorded prospectively by the practitioners doing the work rather than reconstructed afterwards by a researcher; the association with mortality (aOR 1.13 per nurse, CI 1.07–1.18) is the cleanest estimate in the paper.

Momentum. The outcome measure prices whether a corrective loop fired in time on cases where it was indicated, and it is the paper's weakest estimate (aOR 1.07, CI 1.00–1.14, interval touching no effect) — which is itself a finding about measuring omissions and is recorded as such below.

Relevance to Brandon's Positioning

This entry's value is methodological, not substantive, and the distinction is load-bearing: it is healthcare, it is not AI, it says nothing about machine speed, and it must never be cited as evidence for any Five Breakpoints claim about transformation. What it supplies is an existence proof against the strongest conclusion the 2026-11-03 research-evidence scan drew from Edmans, Fang & Huang (2022). That scan concluded that organizational failure as Brandon defines it "has no announcement date," and therefore that for omission-shaped failures "better sampling cannot help, because the thing generates no event to sample" — making practitioner observation "the only instrument that can see the case." This paper is a 202,428-patient administrative instrument that prices an omission against an outcome recorded elsewhere, and it needs no announcement date to do it. Edmans et al.'s stated constraint is specifically that stock returns cannot price investment cuts because long-run returns absorb many decisions and quarterly reporting gives no announcement date. Both are limits on market-based identification, not on measurement. Wells et al. show the three moves that substitute for an announcement date: (1) a required level authored by practitioners at the time, not reconstructed later; (2) a shortfall measured against that standard on a dense administrative cadence — the shift calendar dates the omission without anyone announcing anything; and (3) a denominator of opportunities — failure to rescue counts deaths only among patients who had a complication, so the measure distinguishes "did not need to act" from "failed to act."

That template bears directly on three live items. The McAdam standard (knowledge/thesis-challenges.md, OPEN since 2026-10-10) says Five Breakpoints is a practitioner synthesis with no sampling frame, comparison group or outcome variable, and knowledge/master-plan.md records that this challenge "cannot be answered by writing." It still cannot be won by writing — this paper is not evidence for the framework and changes no claim. But the concession paragraph that asset 12 is built around can now name a specific, nationally-implemented instrument design from a practitioner field that had exactly Brandon's problem and solved it, which is materially stronger than naming what would hypothetically test the framework. master-plan.md Tier 0.2, the post-engagement measurement and the only time-gated item on the list, currently specifies one agreed outcome per engagement at twelve months; Wells et al. says that is missing the denominator — without a recorded population of engagements where a design step was indicated, a twelve-month outcome cannot separate the step that was unnecessary from the step that was skipped. Paper 2's Claim 2 need (paper-2-case-file.md, reclassified today to STRUCTURALLY CONSTRAINED) should be narrowed rather than reversed: constrained for event-study identification, open for state-and-shortfall instruments.

The bound on all of the above, stated because it cuts against the argument it supports. The one outcome in this study that is omission-shaped is also the one whose confidence interval touches no effect (failure to rescue, 1.00–1.14), while the two act-shaped or dense outcomes (mortality, complications) are clearly estimated. Prior work is mixed in the same direction — the measure's CBE#0353 endorsement was withdrawn, at least one multisite study found no staffing association with failure to rescue at all, and validity depends on the coding quality of secondary diagnoses in administrative data. So the defensible claim is that omissions can be measured, not that they can be measured as precisely as acts. That is still enough to deny the scan's strong conclusion, and it is not enough to claim the problem is solved.

Sources & Confidence

• Wells CI, Henderson K, Varghese C, O'Grady G, Boyle L, Bissett IP, Campbell D. "Association of Surgical Ward Nurse Staffing With Postoperative Mortality and Failure to Rescue." ANZ J Surg 2026 Oct;96(10):2949–2958; epub 2026-08-13; doi:10.1111/ans.70903. Read in full at PMC13724767 — High for the study's own design, definitions and reported estimates.
• Edmans A, Fang VW, Huang AH. "The Long-Term Consequences of Short-Term Incentives." J Accounting Research 60(3):1007–1046, June 2022 — High; held in the base at research-evidence/knowledge-base/edmans-fang-huang-long-term-consequences-short-term-incentives-2022-06.md. Cited here only for the announcement-date passage this entry contests.
• The wider failure-to-rescue validation record (Silber-lineage definition; Needleman/Buerhaus 2002, 799 hospitals; the withdrawn CBE#0353 endorsement; the English administrative-data validation and its coding-quality caveat; a multisite null) — Moderate, characterized from search summaries and not read at the primary. It is recorded to bound this entry, and no figure from it may be cited.

Filed by / Routing

Filed by: Stafford (judgment layer, daily brief, 2026-11-03) Routing: strategy — the instrument-design template for master-plan.md Tier 0.2 and the asset 12 concession section. Practice asset: none — not a Category 8 prescriptive-design document; this is an empirical instrument.

Tags: paper-2, practitioner, Incentive Fragmentation, Strategic Disconnection, Process Friction, Purpose, Commitment, Capability, COUNTERARGUMENT Source: Journal of Economic Literature 54(1), American Economic Association (Luis Garicano, LSE & CEPR; Luis Rayo, University of Utah & CEPR) URL: https://doi.org/10.1257/jel.54.1.137 (record: https://researchonline.lse.ac.uk/id/eprint/65864/ · full pre-publication text read at https://conference.nber.org/conf_papers/f80906/f80906.pdf) Date published: 2016-03-01 Date flagged: 2026-11-02

Summary

Garicano and Rayo survey the organizational-economics literature on large-scale organizational failure, organizing it under two root causes stated in their own abstract: "Organizations fail due to incentive problems (agents do not want to act in the organization's interests) and bounded rationality problems (agents do not have the necessary information to do so)." They explicitly exclude failures caused by the world changing underneath a sound organization — Schumpeterian obsolescence — and restrict the survey to "failures that result from (ex-ante) poorly designed organizations," at a scale that "would have justified significant organizational change." The method is a three-way combination, stated as such: literature review, a deliberately simple formal model per mechanism, and a named case per mechanism. Case selection is defended on a stated criterion ("notoriety": overwhelming media and academic impact plus economic magnitude sufficient to cause bankruptcy or force redesign), and the paper adds an external-validation pass at a referee's suggestion, classifying the ten largest US bankruptcies 1980–2014 against its own categories.

The survey covers five failure families. Allocation of authority (Section 2): decentralized authority incorporates local information but risks losing coordination and scale, and a flawed allocation produces organization-wide failure — case, the Daimler–Chrysler merger. Short-termism (Section 3): "an agent with wide discretion is rewarded by high-powered, short-term incentives," and therefore "selects apparently-safe projects in order to capture short-term rewards at the cost of placing the organization at risk of catastrophic failure" — cases, AIG and BP Texas City. The authors class both of these as "multitasking problems" after Holmstrom & Milgrom (1991) and Baker (1992), the first static and the second intertemporal. Communication failures from misaligned incentives (Section 4): coarse, biased upward communication in the Crawford–Sobel tradition, plus a model the authors introduce themselves — the "cover-your-ass" or CYA mechanism, in which a subordinate conceals valuable but compromising information from his boss in order to protect the boss from the consequences of knowing it. Resistance to change (Section 5): rent-seeking coalitions blocking adaptation, and rigidities in culture and relational contracts — cases, airlines failing to launch low-cost subsidiaries, the NYSE's inability to introduce electronic trading, HP and the "HP Way." Bounded-rationality failures (Section 6): talent misallocated across hierarchical positions (the Spanish Cajas), and miscommunication from incompatible organizational codes (merger "culture clashes").

Two features of the survey matter more than any single section. First, the authors claim interaction between the failure families, not independence: "our survey suggests some interactions between different sources of failure. For example, by increasing the extent of the conflict between the principal and the agent, decentralized authority (Section 2) aggravates all other incentive and communication problems we have discussed (Sections 3-5)," and the models are "designed to complement one another." Second, the authors concede their own evidence gap in print: "although there is a growing empirical literature on organizations, much of the theoretical work we covered has not been systematically tested. This is particularly the case for resistance to change (Section 5) and shared specific human capital (Section 6). Introspectively, researchers appear to agree that these aspects of an organization matter, but insufficient hard evidence has been produced to date." They also record two phenomena for which the literature they survey has no formal account: that hierarchical position itself disrupts information (co-pilots intimidated by pilots), and that short checklists work (the Pronovost ICU catheterization result, 82% lower bloodstream infection rate).

Five Breakpoints Intersection

Incentive Fragmentation: Section 5.1's Result 5(a) is a formal statement of the mechanism — "firms with strong goodwill do not adopt radical technological changes," because "those agents whose skills are not complementary to a new technology prefer to block change and thus retain their ability to extract rents." That is Paper 1's Incentive Fragmentation case — one executive with veto power stalling a migration — derived rather than narrated, with two named cases (airline low-cost subsidiaries; NYSE electronic trading) and, unlike the framework, two explicit remedies: compensate the losers up front ("kicking them upstairs") or "defuse power in advance across individuals with varying areas of expertise."

Strategic Disconnection: Section 4 is the strategy-execution gap stated as an information problem with incentives attached — misaligned incentives make upward communication "non-credible (and thus coarse and biased)," and "even when information is hard, high-powered incentives may lead bosses to prefer not to receive warnings from their subordinates." The second half is sharper than anything in the framework: it is not that the message failed to arrive, but that the recipient's incentives made him prefer not to receive it.

Process Friction: Section 6.2's code-incompatibility result is structural friction with a stated trade-off the framework does not carry — "a code that is better tailored to the needs of a particular set of agents poses a higher risk of code incompatibility across divisions and across firms following a merger: the strength of codes becomes its weakness." Friction is the price of local efficiency rather than a defect in it.

Momentum Mirage — no analogue found, and the absence is the finding. Nothing in this survey models or cases the appearance of progress without movement. Of the five breakpoints, Momentum Mirage is the one organizational economics does not reach, which bears directly on the relabeling objection (see Positioning).

Technology Illusion — deliberately not tagged. Result 5(a) concerns technology not being adopted because incumbents block it, which is the opposite failure from deploying technology onto broken organizational conditions. Tagging it would be the loose mapping the base's rules exist to prevent.

Four Forces Intersection

Commitment: The vested-interests model is a Commitment result — change requires the organization to act against the interests of agents who hold blocking power, and the authors' two remedies (ex-ante defusion of blocking power; credible compensation of losers) are commitment devices chosen before anyone knows who the losers will be, which is a stronger version of the framework's claim that follow-through fails under pressure.

Purpose: The survey's prescribed response to multitasking failure includes "developing an identity" and "creating a sense of mission" (Akerlof & Kranton 2005; Bénabou & Tirole 2011; Dewatripont et al. 1999) — Purpose entering not as exhortation but as the substitute for incentives when output cannot be measured.

Capability: Section 6.1's talent result is a Capability claim at the level of allocation rather than stock — "what matters is not merely the overall talent available to the organization, but how this talent is leveraged by assigning it to the right hierarchical positions."

Momentum — no implication. Consistent with the missing Momentum Mirage analogue: the survey's models are static, which the authors concede ("much of the work we surveyed has been developed for a static context").

Relevance to Brandon's Positioning

This is the prior work the relabeling objection needed, and it arrives against Brandon's strongest defence. knowledge/thesis-challenges.md sets the objection's winning condition precisely: "a prior work that does the integration — any published framework treating three or more of these five as a single interacting system of failure conditions, with a diagnostic attached." knowledge/asset-suggestions.md #12 records that search as never run and pre-registers both outcomes: empty means the section is a victory lap, a hit Brandon has not cited means "this is not a section, it is a revision." It is a hit. Garicano & Rayo integrate five failure families under two root causes, in the field's top review journal, ten years before Paper 2, and they claim interaction explicitly rather than treating the families as separate literatures. Brandon's papers, the book manuscript and five-breakpoints-source.md contain zero citations to organizational economics (checked 2026-11-02: no Garicano, Holmstrom, Milgrom, Gibbons, or "multitasking"). The integration defence — the first and strongest of the three — does not survive this document.

The other two defences survive, and this document strengthens both. "None of them diagnoses" holds: there is no instrument, no checklist and no self-administered test anywhere in the survey, and the authors go further by naming the checklist as a phenomenon their literature cannot explain ("we found no formal theoretical explanations... simple routinized procedures, such as short check lists, have proven helpful"). "None has an account of speed" holds absolutely — the authors concede the static-context limit themselves. So the correct revision is not a retreat but a re-division: concede integration to organizational economics, keep diagnosis and speed, and name the field whose own survey says a diagnostic is the thing it cannot build.

It also partly discharges the McAdam standard, which is the more valuable of the two effects. That challenge (OPEN, 2026-10-10) says Five Breakpoints is Kotter's method with Kotter's evidence problem — practitioner cases, no sampling frame, no outcome variable. Garicano & Rayo concede the identical gap for the whole organizational-economics literature on resistance to change and shared human capital, in print, in the JEL: researchers "appear to agree that these aspects of an organization matter, but insufficient hard evidence has been produced to date." The honest concession Brandon has to make is therefore a field-level condition rather than a personal deficiency — and he can cite the field's own survey saying so. Their case-selection defence ("notoriety," plus the ten-largest-bankruptcies validation pass) is also the cheapest available template for defending Brandon's own case selection, which currently has no stated criterion at all.

And it answers an OPEN challenge the framework could not answer itself. "The framework has no theory of who holds the remedy" (2026-09-30): Garicano & Rayo's answer is the designer who allocates authority ex ante, before the direction of change is known, because "a priori it may be hard to know the direction of change and therefore who the losers will be." That is a theory of remedy-holding, with a mechanism and cases, and it lands directly on master-plan.md Tier 3.2 (the Commitment test at entry) — which goes from a practitioner intuition out of the CISO case to a practitioner instrument with a formal result behind it.

Not a publication window. A 2016 survey has no clock on it (CLAUDE.md rule 5). It is argument value for master-plan.md Tier 2.1 and Tier 4.1, nothing more.

Sources & Confidence

• Garicano, L. & Rayo, L., Why Organizations Fail: Models and Cases, Journal of Economic Literature 54(1), March 2016, pp. 137–192, DOI 10.1257/jel.54.1.137. Bibliographic record verified at LSE Research Online (eprint 65864) 2026-11-02; LSE holds no full text. — High (AEA peer-reviewed review journal; JEL classification D21, D86, J33, L23, M52).
• Version limit, and it binds every quotation above. The full text read on 2026-11-02 is the pre-publication draft dated 2 February 2015, served at conference.nber.org/conf_papers/f80906/f80906.pdf, not the published 2016 article. Under CLAUDE.md rule 14 the published version must be read before any passage here is quoted in print; section numbering, wording and the referee-prompted bankruptcy footnote may differ. The abstract text in that draft matches the published abstract as recorded by LSE, which is reassurance about the paper's identity and not about its page-level wording. — Moderate on wording, High on existence and substance.
• Luis Garicano's authorship is a positioning fact worth recording: LSE professor of economics and strategy at the time of writing, later an MEP and widely published for general audiences. This is not obscure prior art that reads as scholarship to concede — it is a well-known economist's survey in the field's review journal.
• Secondary summaries (SIOE review; IDEAS/RePEc record) were used only to locate the paper and are cited for nothing. — Low, and load-bearing on no claim here.

Filed by / Routing

Filed by: Stafford (judgment layer, 2026-11-02) — reached through the reference list of the research-evidence scan's 2026-11-02 ingest (arXiv:2604.10290), which cites it in its "How Human Organizations Fail" section. Routing: strategy / market Practice asset: none: not a Category 8 prescriptive design artifact — a theory survey. Its practice bearing is indirect: Section 5.1's two remedies (compensate losers ex ante, defuse blocking power across varied expertise) are the formal result behind master-plan Tier 3.2's Commitment test at engagement entry.

Tags: paper-2, COUNTERARGUMENT Source: NBER Working Paper 35697 (Michael Dalton, Lisa B. Kahn, Andreas I. Mueller) URL: https://www.nber.org/papers/w35697 Date published: 2026-09-01 Date flagged: 2026-10-30

Summary

The authors match near-universal Burning Glass (BG) online job-posting data to the Job Openings and Labor Turnover Survey (JOLTS), the representative employer survey that is the official US measure of job openings. This is a benchmark of a text-inferred labour measure against survey/administrative establishment records, at the establishment level, rather than a comparison of published aggregates.

Three results are stated in the abstract. First, BG converged toward JOLTS over 2007–2022 — the gap between the online-postings universe and the representative measure narrowed across the period. Second, BG's representativeness varies across establishment characteristics, which the authors address by reweighting BG so that its postings align with JOLTS openings along observables. Third, the reweighting has sharply different consequences for different prior findings drawn from BG: the level and cyclicality of skill demand survive the adjustment, which the authors attribute to large establishments accounting for most openings and being well represented in the posting data; but BG overstates labour-market concentration (the Herfindahl–Hirschman Index) by nearly threefold, because small and mid-sized establishments post less online.

The paper is gated (NBER subscription; SSRN lists a $5 charge) and no author copy or free PDF was reachable on 2026-10-30. Everything above is from the abstract as printed on the NBER landing page and reproduced independently on RePEc; the two agree. Per CLAUDE.md rule 14, the landing page was checked for a revision date and none is shown — this is the original September 2026 text, not a revision. No figures, sample counts, reweighting details or coefficients are claimed here, because the body was not read.

Five Breakpoints Intersection

No direct Five Breakpoints mapping, and the tag is withheld deliberately rather than stretched. This is a measurement paper about a data source; it makes no claim about any organization's purpose, incentives, process, technology deployment or momentum. The one organizational fact it reports — that small and mid-sized establishments post fewer jobs online than large ones — is a statement about posting behaviour, not about a breakpoint.

Logged for the proxy problem binding paper-2 Claim 5, and for a genre distinction that must be stated before this source is used anywhere. Burning Glass/Lightcast is employer-side job-postings text. The Résumé-Panel Consensus pattern (ESCALATED, knowledge/emerging-patterns.md) is about worker-side résumé/employment-history panels — Revelio, Ewens & Giroud, Babina. These are different data genres with different selection problems, and today's scan brief wrote down the parallel caution for exposure indices. So this paper is not an administrative benchmark of the résumé panels and must never be cited as one. It is the first administrative benchmark of the adjacent postings genre, and what transfers is a prior about which kinds of quantity survive benchmarking, not a verdict on any résumé result.

Four Forces Intersection

No Forces tags. The source is paywalled and was not read; per CLAUDE.md rule 2 and the scan's anti-pattern on guessed Forces, forces: [] is the correct record. The abstract implicates no Force.

Relevance to Brandon's Positioning

This partially discharges the acquisition target that has been priority 1 on paper-2 Claim 5 for eight consecutive scans — and the half it discharges is not the half Claim 5 needed. The target, carried in paper-2-case-file.md since 2026-10-29 and hunted by keyword four times this month, was: any validation of a text-inferred organizational construct against an observed organizational property. A validation of a text-inferred labour-demand measure against administrative establishment records exists, it was issued in September 2026, and the base did not hold it. What it validates is vacancy counts, skill demand and market concentration — not an organizational property such as hierarchy depth or organization capital. So the target is discharged for the postings genre and stays open for the construct Claim 5 actually rests on.

What transfers, and it is worth more than a clearance would have been: a measured prior on which kinds of text-inferred quantity survive an administrative benchmark. Inside one dataset, one benchmark and one reweighting, quantities describing the composition of labour demand came through intact, while the one quantity describing structure was overstated by roughly a factor of three — in a known direction, from a named mechanism (small and mid-sized establishments post less online). Composition survived; structure did not. That is this base's own 2026-10-07 falsification finding — that the benchmarking which exists is on counts while the claims that matter are about shape, and that "validating that you can count a firm's employees validates nothing about counting its layers" — moved from argued to measured, in an adjacent genre.

Stated as a prior and not as a result, it cuts against Brandon's own evidence as hard as against the challenges. Ewens & Giroud (NBER 34162), the spine of asset 2's delayering argument, infers hierarchy depth — a structural quantity — from text, has never been benchmarked against any administrative record, and has moved its own headline layer count from ten to six across three revisions. Today's paper does not touch it and cannot refute it. But it is now the one data point the base holds on what happens when a structural quantity from text meets administrative records, and the answer was a factor of three. That is a reason to hold the delayering magnitude loosely, not a reason to drop the argument. Symmetrically, the two newest challenges against the thesis — The wave is invisible where the technology is (2026-10-24) and The stock-flow objection (2026-10-29) — are not narrowed by this: Babina's heterogeneity split runs on capitalized SG&A, an audited accounting number, and its AI measure is résumé text rather than postings. Neither is in this paper's genre.

What it does not do: it does not supply Claim 5's missing mediator; HHI is neither hierarchy depth nor organization capital. Its window is 2007–2022 and therefore pre-generative. And reweighting repaired one family of findings while breaking another, which is the finding rather than a caveat on it.

Sources & Confidence

• NBER Working Paper 35697, Do Online Job Postings Capture Job Vacancies? — Dalton, Kahn & Mueller, September 2026, DOI 10.3386/w35697. https://www.nber.org/papers/w35697 — High (NBER working paper) for the abstract's claims; body unread, so no confidence is asserted for anything beyond the abstract.
• RePEc listing for NBER WP 35697, used as an independent route to the same abstract. https://ideas.repec.org/p/nbr/nberwo/35697.html — High as a corroborating route; it reproduces the NBER abstract and shows no revision.
• Hershbein & Kahn (2018, AER) and the OECD review of Burning Glass policy use were surfaced on the same chase as the lineage's earlier validation work. Not read, not cited as evidence here — recorded as the route, per the pre-registration in the scan's query-log.md.

Filed by / Routing

Filed by: Stafford (judgment layer, 2026-10-30) Routing: strategy — a measured prior on which text-inferred quantities survive benchmarking, bearing on asset 2's delayering magnitude; and a genre distinction (postings vs résumés vs exposure indices) that the shared pattern file should record.

Tags: paper-2, Technology Illusion, Capability Source: National Bureau of Economic Research — Working Paper 35684 URL: https://www.nber.org/papers/w35684 Date published: 2026-08-01 Date flagged: 2026-10-29

Summary

Tania Babina, Alex X. He and Renhao Jiang build a firm-level measure of AI investment from AI-skilled employment — the share of a firm's workforce in AI-related roles, inferred from résumé data and matched to Compustat — and state that the measure spans "machine learning through generative and agentic AI." Their headline result, quoted from the abstract, is that "AI investments are associated with productivity growth in recent years, but not over the previous decade."

The paper's contribution is the mechanism it puts between the investment and the gain. The authors construct a novel measure of organization capital from workers' job descriptions — defining organization capital as durable firm-specific knowledge accumulated through learning-by-doing that enables more efficient production — and report that the productivity gains are driven by AI-skilled jobs that build that capital. The abstract's closing claim is that "AI investment generates productivity growth by creating organization capital."

Secondary summaries put the measured window at 2018–2024, against a 2010–2018 period in which the same instrument finds no productivity effect; an ITIF Fact of the Week (Trelysa Long, 2026-09-08) reports roughly one additional percentage point of productivity growth per year for firms expanding AI investment, and search returns attribute 6.2 percentage points of sales-per-worker growth and 7.5 percentage points of TFP growth over six years per standard-deviation increase in the growth of a firm's AI-worker share. The same secondary reports the result as robust to alternative measures and not explained by acquisitions, offshoring or pandemic effects, and records Babina characterizing the finding as consistent with a productivity J-curve.

Read-depth note, and it bounds every number above. The abstract was read verbatim at the NBER landing page and at EconPapers. The primary PDF is access-denied — nber.org/system/files/working_papers/w35684/w35684.pdf returns an NBER "Access denied" HTML page, and CEPR DP21894 (403), SSRN 7345627 (403) and the ITIF page (404 on both URL forms) all refused this environment. No effect size, sample size, firm count, specification or standard error has been read at a primary or at a readable secondary. The effect sizes and the 2018–2024 window are search-snippet provenance only and may not be published. forces: ["Capability"] is read off the abstract's own statement of the mechanism, nothing more.

Five Breakpoints Intersection

Technology Illusion: The paper is the breakpoint's claim rendered as an estimate rather than a case — AI investment on its own is not what moves productivity, because the authors report the gains running through AI-skilled jobs that build organization capital and find no productivity effect at all for the decade of AI investment that preceded it; a firm that bought the capability without accumulating the organizational substrate is the paper's own null arm.

Breakpoints considered and declined, recorded so a later scan does not re-tag them. Momentum Mirage was considered on the strength of a decade of AI investment with no measured productivity gain — declined, because the authors' reading is that the gains arrived late (a J-curve), not that the early activity was illusory, and tagging it would put the framework's interpretation in the source's mouth. Process Friction was considered and declined: organization capital is a stock of knowledge, and the paper measures no handoff, queue or decision right.

Four Forces Intersection

Capability. The framework defines Capability as the skills, infrastructure and organizational capacity to execute, and this paper's proposed mediator is that construct by another name — durable firm-specific knowledge built through learning-by-doing, standing between the technology spend and the realized outcome. This tag is read off the abstract only and is recorded on the same terms as the Huang/Huang/Hong entry: the paper itself was not read, and no second Force is assigned because assigning one from a landing page is the guessed tagging the operating rules forbid.

Relevance to Brandon's Positioning

This is the nearest thing the base has acquired to Claim 5's standing need — a non-technological organizational variable sitting between AI investment and a realized productivity outcome, in a firm panel, with the technology-only arm estimated at zero. It is also a partial answer to the acquisition target opened on 2026-10-28: a post-generative firm panel with complement controls. Partial is the operative word, and the shortfall is the finding. If the secondary-sourced window is right, 2018–2024 straddles the generative break rather than following it, and the AI measure is AI-skilled employment — hiring machine-learning staff, which is the pre-generative genre of AI investment. The exposure recorded on 10-28 — that the strongest evidence for Brandon's central mechanism is pre-generative — is narrowed by this paper and not closed by it.

The sharper positioning point is the interpretive fork, and Brandon is on one side of it by name. Babina's own framing is a productivity J-curve: the gains are late because organization capital takes time to build. That is the maturity-ladder logic in econometric dress — wait, and it arrives — and it is the same shape as the McKinsey reinvention horizon Brandon declined on 2026-10-28. Brandon's framework reads the identical coefficient differently: organization capital is not a thing that accrues with patience, it is a thing that accrues only where purpose, commitment, capability and momentum were designed to hold, and most organizations are not short of time but short of design. The two readings are observationally equivalent on this data and they diverge on a testable point: under the J-curve, the laggards converge; under Five Breakpoints, they do not. That is a paper-2 argument and it is now stateable against a named NBER working paper.

The caution that must travel with any use of this source, and it is the base's own standing rule. Tania Babina and Alex X. He are two of the four authors of Firm Investments in Artificial Intelligence Technologies and Changes in Workforce Composition (NBER 31325), already held. The instrument is the same genre — firm-level AI inferred from résumés — and here the proposed mediator is inferred from the same substrate as the explanatory variable, since organization capital is read out of the text of job descriptions. This paper and NBER 31325 are not two sources, and "organization capital measured from job ads" is a proxy whose validity nobody in this literature has yet checked against an observed organizational property. Under the Résumé-Panel Consensus pattern this is a third member of one data genre, not a third confirmation.

Sources & Confidence

• Babina, Tania; He, Alex X.; Jiang, Renhao. Canaries in the Gold Mine: Early Productivity Gains from Artificial Intelligence Creating Organization Capital. NBER Working Paper 35684, August 2026. DOI 10.3386/w35684. https://www.nber.org/papers/w35684 — abstract read verbatim at the landing page. Moderate, matching the credibility assigned to the sibling NBER 31325 entry; the primary is unread and the publisher tier does not substitute for a read.
• EconPapers listing RePEc:nbr:nberwo:35684 — abstract and JEL codes (D22, D24, D25, E22, G3, G30, G32, J24, O33) read. Moderate.
• ITIF Fact of the Week, Trelysa Long, 2026-09-08 — NOT read. Both URL forms returned 404 to direct fetch. The 2018–2024 window, the ~1pp-per-year figure and the robustness claims are search-snippet provenance only. Low.
• The 6.2pp sales-per-worker and 7.5pp TFP figures — search-snippet provenance, no page read, not confirmed at the primary. Low. Not publishable.
• Babina's data page states the firm-level AI panel will be posted publicly on conditional acceptance; the public replication repository carries synthetic Revelio-style test data only, which is the clearest available confirmation that the underlying substrate is Revelio résumé data. Moderate on the substrate, unread on everything else.
• NOT read: the paper. NBER's PDF endpoint returns an "Access denied" page to this environment; CEPR DP21894 and SSRN 7345627 both returned 403. Registered as a chase target with the obstruction named.
• Non-independence, binding. Shares two of four authors (Babina, He) and the résumé-panel instrument genre with the held babina-fedyk-he-hodson-ai-workforce-composition-2023-06.md (NBER 31325, window 2010–2018, Cognism/Burning Glass). Never cite the two as independent confirmations, on the same terms as the Babina / Ewens & Giroud rule in CLAUDE.md.

Filed by / Routing

Filed by: claude-opus-5 daily scanner (daily scan, 2026-10-29) Routing: strategy Practice asset: none — not a Category 8 prescriptive-design source; this is a firm-panel econometric instrument.

Stafford verification addendum — 2026-10-29, 14:30 run: the primary was obtained and read in full, and three of this file's cautions are now wrong

Why this addendum is appended to the scan's file rather than opened as a second file. One file per source is the rule, and this is the same source. The 12:00 scan filed this entry with the primary unread and recorded six key findings, three of which this read falsifies. Leaving them would merge into the index as standing cautions and suppress usable evidence — the failure mode CLAUDE.md hard rule 14 exists to prevent ("a false positive freezes the base on a superseded number"). Everything above this line is the scan's text, unaltered. Everything below is Stafford's, and the pending staged entry was corrected and held for human verification on the same grounds.

The route, which is the generalizable part

Hard rule 14 names the standing route for any working paper: the author's own copy plus the landing-page revision date. The scan tried nber.org (Access denied), CEPR DP21894 (403), SSRN 7345627 (403) and ITIF (404). It did not try the author route, and the author route fails here too: taniababina.com and alexxihe.com were both read today and neither hosts a PDF — both link only to the SSRN abstract page that 403s.

The paper was found in the authors' own public replication repository. github.com/jrhao7/AI_productivity_replication contains AI_productivity_BHJ.pdf alongside the keyword dictionary and code, fetched as raw.githubusercontent.com/jrhao7/AI_productivity_replication/main/AI_productivity_BHJ.pdf — HTTP 200, 11.5 MB, 267 pages. Standing route addition: when a working paper has a public replication repository, check the repository for the manuscript. Replication repos are built to be readable, carry no paywall, and are frequently linked from the author's own data page — this one is, two clicks from taniababina.com/data.html, which this system had already read once without opening the repo.

Version check, per hard rule 14. The PDF's own footer reads "First draft: February 26, 2026. This draft: August 22, 2026." That matches the NBER landing page's August 2026 and the SSRN late-August posting. One version, no revision history, no figure drift. The repo has a single commit.

What this file said, and what the paper says

The figures, at primary

• Sample: U.S. public firms, non-tech sectors, long differences. Panel 2010–2018: N = 1,222 (sales per worker) / 1,091 (TFP). Panel 2018–2024: N = 1,321 / 1,173.
• Table 2 Panel A (2010–2018): ∆ share of AI jobs on ∆ log sales per worker −0.004 / 0.002; on ∆ TFP 0.014 / 0.012. All insignificant. This is the null decade, measured.
• Table 2 Panel B (2018–2024): 0.048** / 0.062** (sales per worker) and 0.061*** / 0.075*** (TFP). Per one-standard-deviation increase in the growth of a firm's AI-worker share, over six years — i.e. the 6.2 and 7.5 percentage points the secondaries reported, ≈1pp/year.
• Table 4, instrumental-variables estimates (AI investment predicted from firms' pre-2010 hiring ties to universities historically strong in AI research): 0.135–0.160 sales per worker, 0.110–0.144 TFP; first-stage F 10.1–26.2; N 528–633.
• Timing: no differential productivity growth in earlier periods for firms that raised AI investment over 2018–2024; estimated effects rise monotonically after 2018; ex-ante AI investment predicts ex-post productivity and not the reverse.
• Measure: Lightcast near-universe of U.S. job postings to identify AI-relevant skills by co-occurrence with core AI skills, then those 510 validated terms searched in Revelio Labs résumé text covering two-thirds of the U.S. workforce. False-positive rate ≈2% (2.03% at large firms). The AI-relatedness score is data-driven, so it tracks new terminology — agentic systems scores 0.98. By 2024, U.S. public firms employ >300,000 AI-skilled workers, 2.1% of their U.S. workforce, with AI shares rising in every industry; investment outside public firms is roughly half as high.
• Robustness: alternative AI and productivity measures, acquisitions, offshoring, technology firms, retrospective résumé relabeling, pre-existing productivity trends, industry and pandemic shocks, job remote-suitability, and non-AI IT investment.

What the paper contains that the abstract does not, and it matters more than the figures

1. The convergence question is tested, and the sign runs against a Five Breakpoints reading. Table 7 splits firms at the within-industry median of pre-period organization capital (capitalized SG&A normalized by PP&E in 2018 — an accounting measure, not the job-text one) and interacts the split with AI investment.

N = 875 / 836. The firms with the least accumulated organizational knowledge get the largest and the only statistically significant gains; the organizationally rich firms get about half as much and nothing distinguishable from zero. Mean reversion is addressed directly and does not explain it: sorting on OC/PP&E averaged over 2014–2018 leaves the result unchanged (Table C.36), and a pre-period placebo on 2012–2018 growth returns small, insignificant coefficients for both groups (Table C.37) — which rules out both mechanical reversion (opposite sign) and anticipation (same sign).

The authors' mechanism is headroom, not readiness, and this is the hinge for any use of the result: "an established stock of OC simply leaves less to codify: OC-rich incumbents still invest and build, but have less scope for further accumulation." They cite Crouzet et al. (2026) as the contrasting account, in which an old firm's existing organization conflicts with the new technology.

2. Appendix E is a formal model in which the lag requires no organizational friction at all. Extending Atkeson & Kehoe (2005), measured productivity is à = ABψ, with AI modelled as a codification technology that raises the rate χ at which operations convert into organization capital. Three stated conclusions:

• A permanent but gradual level effect — adoption raises productivity permanently by ψ log(χ′/χ); "firms do not grow faster forever, they converge to a higher path."
• Slow adjustment with no frictions required — the post-adoption gap follows a first-order recursion whose convergence rate is the depreciation rate of organization capital plus untreated firms' organization-capital growth; plausible depreciation implies half-lives of five years or more. Verbatim: "Responses that take many years to build are therefore not evidence of adjustment costs or implementation lags—they are what a stock-flow system delivers when stocks are large relative to flows."
• Heterogeneity runs through the organization-capital ratio r = B/A — "Firms with a low level of codified knowledge stock respond most and fastest, which we test empirically." Proposition 3 makes the eventual gain common across adopters and the speed state-dependent.

3. The organization-capital measure is validated, in five steps, one of them out of genre. Job descriptions scored by LLM for concrete evidence of building durable firm capability, threshold 0.6, aggregated to firm-year shares of OC-building and non-OC-building AI jobs, plus an 18-category taxonomy of OC-building activity. Validation: (i) worked contrasting examples — "a machine learning engineer building a model for a single client engagement… is not OC-building"; (ii) a writing-quality placebo — OC-building descriptions are longer but use more repetitive vocabulary, cite fewer specific numbers and tools, and have the same spelling-error rate, and are indistinguishable from non-OC descriptions of similar length on every writing-quality measure; (iii) eight firm case studies; (iv) intuitive variation by function — 79% of AI workers in management roles classify as OC-building versus 22.6% in product support, and only about half of core AI-development roles; (v) external validation against audited accounting data — firms with more OC-building AI jobs subsequently accumulate more organization capital as measured by capitalized SG&A, while non-OC-building AI jobs do not (§4.2, Table C.32). The honest limit, which must travel with the correction. Capitalized SG&A is itself the standard accounting proxy for organization capital (Eisfeldt & Papanikolaou), not an observed organizational property — so this is a text proxy validated against an accounting proxy, and no one has yet compared either to a directly observed organizational fact. It is a narrowing of the unvalidated-proxy objection, not its closure. Note also that capitalized SG&A plays two roles: it validates the text measure and defines Table 7's groups.

4. Structure thickens at the senior end; employment does not grow. Over 2018–2024, AI investment predicts substantial sales growth and no employment growth (in 2010–2018 sales and employment grew at similar rates). Intangibles rise and tangibles do not: SG&A ≈+5%, R&D ≈+2.5%, PP&E −0.7% and insignificant (Table C.24). Workforce composition, computed excluding AI workers to avoid a mechanical result (Table C.25): shares of workers with master's degrees and PhDs and shares in senior roles rise, while shares of non-college workers and those in junior roles fall; average wages rise (Table C.27). The authors link this to evidence that AI adoption disproportionately displaces junior workers (Brynjolfsson et al. 2025a; Hosseini Maasoum & Lichtinger 2025; Kim & Koning 2026). No span of control, no management-layer count, no delayering measure appears anywhere in the paper. Footnote 34 does state the flattening mechanism in theory, citing Garicano (2000): "Codifying that expertise into a firm-wide system reduces the need for the matching that the hierarchy exists to perform." The paper holds that mechanism in its theory and measures senior-share expansion in its data, and does not reconcile the two.

Revised relevance to Brandon's positioning

The section above files this paper as "the nearest thing the base has acquired to Claim 5's standing need — a non-technological organizational variable sitting between AI investment and a realized productivity outcome." After the read, that is not available, and the reason is a fork Brandon has to pick a side of.

• If organization capital is the Capability force by another name (the Four Forces tag above asserts it is), then Table 7 contradicts Breakpoint 4 at n = 875: the firms with the least organizational substrate are the ones AI pays off for, which is the opposite of technology failing where organizational conditions are weak.
• If organization capital is instead codification headroom — the authors' own mechanism, less already codified means more left to codify — then it is not an organizational-readiness variable at all, and the paper is not evidence for Claim 5 in either direction.

Either way the paper is not support. It is the best available statement of the position Brandon's framework has to beat, and Appendix E is why: the model generates the exact observable the framework is built on — technology bought, nothing visible for years — with zero organizational failure anywhere in it. Full challenge: tgifriday7/stafford-research knowledge/thesis-challenges.md, 2026-10-29, The stock-flow objection.

Credibility unchanged at moderate. A working paper read in full is still a working paper; the read upgrades what may be quoted, not the publisher tier.

Filed by: Stafford (judgment layer, 14:30 run, 2026-10-29) — addendum only; the scan's text above is unaltered.

Tags: paper-2, practitioner, Momentum Mirage, Momentum, COUNTERARGUMENT Source: McKinsey Quarterly (Steven Aronowitz, Aaron De Smet, Deirdre McGinty) URL: https://coreinternational.com/wp-content/uploads/2015/06/Getting-organizational-redesign-right-McKinsey-1.pdf (third-party mirror of the McKinsey Quarterly article; mckinsey.com returns HTTP 503 to this environment) Date published: 2015-06-01 Date flagged: 2026-10-28

Summary

McKinsey Quarterly's June 2015 article reports that "less than a quarter of organizational-redesign efforts succeed. Forty-four percent run out of steam after getting under way, while a third fail to meet objectives or improve performance after implementation." It then sets out nine "golden rules" derived by "combining the results of their research with insights from working with multiple companies," and reports that following them raises the odds of success sharply: "73 percent of the executives whose companies followed more than six of them felt that the organizational redesign had succeeded. Executives at these companies were six times more likely to 'declare victory' than those at companies that adopted just one or two. Following all nine rules in a structured approach yielded an even higher success rate: 86 percent."

The article states that the rules are not obvious: "We tested more than 20 common approaches and found that upward of half of them weren't correlated with success," naming competitor benchmarking as a specifically disconfirmed approach. Supporting findings follow the same pattern of self-description: "Sixty percent of the executives in our survey told us they didn't spend sufficient time" on diagnosis, and "almost four out of five survey respondents who owned up to basing decisions on 'gut feel' acknowledged that their chosen blueprint was unsuccessful."

The article carries no methodology section. The PDF text discloses no sample size, no fielding date, no sampling frame, no response rate, and no definition of or scoring rule for "success." The respondents are described only as "a large set of global executives." Sample details attributed to this study by secondary sources — 1,323 companies, a 2014 survey, or the related McKinsey Global Survey figures of 2,063 executives fielded September 2013 and weighted by national GDP contribution — are not stated in the article itself and are recorded here as secondary and unverified. McKinsey published a successor, "The new rules for getting your operating model redesign right," dated 2025-06-25, which revises the nine rules against a reported survey of 2,000 executives; its figures could not be retrieved (see Sources & Confidence).

Five Breakpoints Intersection

Momentum Mirage. The article's own dependent variable is whether executives "felt that the organizational redesign had succeeded" and were willing to "declare victory" — the phrase is McKinsey's, in its own scare quotes — so the field's most-cited quantification of redesign success measures declared progress rather than measured movement toward an outcome, which is the Momentum Mirage operating inside the instrument that is supposed to detect it.

Four Forces Intersection

Momentum. The article's strongest practice finding — that following more of the nine rules raises reported success — is built by crossing a respondent's self-description of their own method against the same respondent's verdict on their own result, so what it demonstrates is the durability of a leader's account of the work rather than the durability of the work; reinforcement based on declaration is the failure of Momentum stewardship rather than evidence of it.

Relevance to Brandon's Positioning

This is the instrument behind the figure Paper 3 uses to replace the retired "70% of transformations fail." Paper 3 (August 2026 update) retires the 70% because it traces to an estimate its own authors called unscientific and a 2011 review found no evidence behind the sources that carried it — then states that "what the evidence does support is harder to wave away: McKinsey's own research on organizational redesign found success rates below 25 percent among organizations that followed few of its nine documented design rules, and 86 percent among those that followed all of them." Read against the standing rule Brandon set on 2026-10-28 — that an instrument in which the respondent self-classifies and is not asked to evidence the classification cannot support a claim about how many organizations are in that category — the 86% is that genre in its most exposed form: self-classified adherence crossed with self-assessed outcome, inside one questionnaire, with no methodology published. The rule as written reaches shares, not conditional relationships, so it does not formally touch Paper 3's use; the reasoning behind the rule reaches it harder, because a correlation between a respondent's account of their method and their verdict on their result is what self-presentation alone predicts. One partial defense is in the article and should travel with the critique: McKinsey reports that more than half of twenty tested approaches showed no correlation with success, which pure halo bias would not produce. Separately, the publisher appears to have restated the same construct downward — a McKinsey email digest reports 59% success for companies using all nine original rules against 97% for the refreshed nine — so the paper's August 2026 update cites a 2015 figure the firm had already superseded fourteen months earlier. Feeds asset 12 ("What the Five Breakpoints Owe") directly: the Paper 3 seventy-percent move, performed on Paper 3's own replacement number.

Sources & Confidence

• McKinsey Quarterly, "Getting organizational redesign right" (Aronowitz, De Smet & McGinty, June 2015), read in full via the coreinternational.com mirror PDF (pdftotext) — High for the article's own text and figures as quoted; Low as an instrument, on the self-report design and absent methodology note documented above.
• Publication date 2015-06-01: the PDF is headed "JUNE 2015"; the day is attested by secondary sources only — Moderate.
• Secondary-attributed sample details (1,323 companies; 2014 survey; 2,063 executives fielded September 2013, GDP-weighted) — Low, not in the primary, not citable.
• McKinsey, "The new rules for getting your operating model redesign right," 2025-06-25, and the 97%/59% comparison carried by a McKinsey email digest — Low, search-summary only; mckinsey.com, mckinsey.com.br and mckinsey.com/~/media/ all returned HTTP 503 to this environment on 2026-10-28. Not citable.
• Aarn Wennekers, "Why I'm Polite, Interested, and Still Skeptical of McKinsey's New Rules for Operating Model Redesign," The Board Room Voice (Substack), 2026-01-29, read in full — Moderate, named critique of the successor article, no methodological audit.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-10-28) Routing: strategy — a citation in a published paper, carrying a correction Brandon should make himself before a reader makes it for him.

Tags: paper-2, practitioner, Technology Illusion, Momentum Mirage, Capability, COUNTERARGUMENT Source: Proceedings of the National Academy of Sciences (PNAS) URL: https://www.pnas.org/doi/10.1073/pnas.2422633122 Date published: 2025-06-25 Date flagged: 2026-10-27

Instrument (recorded before the finding, per CLAUDE.md rule 9)

• Data source: primary field experiment. One large high school in Turkey; roughly 50 classes across grades 9, 10 and 11; nearly 1,000 students. Four 90-minute in-class sessions, collectively about 15% of each grade's class time in the subject.
• Inference method: randomized controlled trial, randomized at class level, analysed intention-to-treat (all students in an arm counted as treated regardless of actual receipt, to preserve randomization). Three arms: GPT Base (standard ChatGPT-style GPT-4 interface), GPT Tutor (GPT-4 with teacher-designed prompts supplying the worked solution and recommended hints rather than answers), and control (textbooks only). Outcome measured twice per session: an assisted practice block with the arm's tool available, then an unassisted exam with the tool removed. Pre-registered; a simpler pre-registered pairwise t-test specification is reported in the appendix and is qualitatively similar. Non-compliance (five sessions lost to logistics) addressed in robustness appendix.
• Observation window: Fall semester, 2023–2024 academic year — generative-era, but early; the authors flag that users were then unfamiliar with the tools and that models have improved since.
• Version note: the SSRN preprint (July 2024, abstract 4895486) is titled "Generative AI Can Harm Learning." The peer-reviewed PNAS version moves the guardrail conditional into the title. Cite the PNAS 2025 version and its title — the preprint title states the finding more strongly than the paper supports.

Summary

The authors ran a randomized controlled trial in a Turkish high school to separate two things that are usually measured together: how well students perform with a generative AI tutor available, and how much they have actually learned once it is taken away. Two tutor designs were deployed against a textbook-only control. GPT Base mimicked a standard ChatGPT interface. GPT Tutor used prompts built by the school's own maths department that supplied the correct solution and teacher-written hints to the model, so the tutor could offer hints and check a student's answer rather than hand over the answer.

On the assisted practice problems, both tools worked, and the better-designed one worked far better: scores rose by 0.137 and 0.361 out of 1 for GPT Base and GPT Tutor respectively, against a control mean of 0.28 — improvements of 48% and 127%. On the subsequent unassisted exam the result reverses for the unguarded arm: GPT Base students scored 0.054 out of 1 below control, meaning unguarded GPT-4 access cut the average control student's unassisted exam performance by 17%, a statistically significant harm. The authors' interpretation is that students used the unguarded model as a "crutch" during practice. The guardrailed arm did not suffer this: GPT Tutor's unassisted-exam point estimate was −0.004, an order of magnitude smaller and statistically indistinguishable from control.

Three further results are reported and are not in the paper's headline. First, self-assessment ran opposite to measured performance in both treated arms: GPT Base students performed significantly worse on the exam but "did not perceive that they performed worse or learned less," while GPT Tutor students did not perform better yet "perceived that they performed significantly better" and rated the practice sessions as more valuable for learning; students in both arms were willing to give up exam time for tool access (about 8.3% and 12.3% of it). Second, both tools compressed the spread of grades during assisted practice — consistent with the existing literature that AI assistance helps the weakest most — but there was no significant effect on dispersion in the unassisted exam, so "the reduction in the skill gap does not persist when access to generative AI is removed." Third, heterogeneity analysis on pre-registered measures of ability, resources and effort found "limited to no statistically significant support for heterogeneous treatment effects," particularly on unassisted performance. The authors state their limitations directly: one school, one subject with objectively gradeable answers, an early-2023 deployment window, short-term outcomes only (a constraint imposed by the partner school), and unknown generalizability to other tutor designs and settings.

Five Breakpoints Intersection

Technology Illusion: This is the breakpoint measured, not merely illustrated. The best-designed deployment in the experiment raised visible output by 127% and moved the underlying capability it was supposed to build by a measured zero (−0.004, indistinguishable from control) — technology laid on top of an unchanged learning process produced a large gain in the number that was being watched and no gain in the capability that mattered.

Momentum Mirage: The appearance of progress without actual movement is measured here in the same subjects at the same time: GPT Tutor students perceived significantly better exam performance while measuring no better, and GPT Base students lost 17% without perceiving any loss. The compression of the grade spread tells the same story — the skill gap visibly narrows while access lasts and does not narrow at all once it is withdrawn.

No Incentive Fragmentation mapping. There is no manager, no scorecard and no evaluation horizon in this design; the incentive layer is Catrinescu's contribution, not this paper's, and tagging it here would be the speculative tagging the base forbids.

Four Forces Intersection

Capability: Direct and primary — unassisted capability is the measured outcome, and unguarded deployment reduced it by 17% relative to never having had the tool. This is the cleanest evidence in the base that organizational capability can be consumed by a deployment rather than merely left unbuilt.

No Purpose, Commitment or Momentum tags. The design holds the deployer's intent and follow-through fixed by construction — all three arms were imposed by the experiment — so the paper cannot speak to them. The perception/reality gap is logged above as a breakpoint (Momentum Mirage), not as evidence on the Momentum force, which concerns whether energy is sustained as novelty fades; this study ran four sessions and explicitly disclaims long-term outcomes.

Relevance to Brandon's Positioning

This is the measurement the paper-2 case file has been asking for and being handed models of instead. The apprenticeship-rung story thread's criterion (c) asked for any measurement that AI-assisted juniors learn more slowly; the file has accumulated two formal models (Catrinescu 2026; Caosun & Aral 2026) and no instrument. Bastani et al. is peer-reviewed, pre-registered, randomized, n≈1,000, and it measures assisted performance and unassisted capability in the same subjects in the same sessions — so it does not merely assert that the two diverge, it shows them diverging in opposite directions at once. It strengthens Claim 3 and it is materially stronger than anything currently supporting it, with the honest limit that the subjects are school students, not employees, and the setting is one school and one subject.

Its most useful service is to dissolve a tension the 10-27 scan left open rather than to confirm the framework. That brief recorded the base as holding "two models on the slower-learning side and one real company on the faster side" — Impala, where less experienced writers reached productivity in one to two months instead of several. Bastani shows those are measurements of different quantities: assisted performance rose 48–127% in the very same experiment where unassisted capability fell or flatlined. Unless Impala measured its writers working without the tool, its result is an assisted-performance number and is not counter-evidence on the learning question at all. The tension should be reframed on that basis rather than resolved in the framework's favour.

It also carries a counterargument against the framework's remedy, and that is why it is tagged COUNTERARGUMENT. The implication Brandon's work invites is that getting the organizational conditions right lets the technology deliver. Here the well-designed arm — teacher-built prompts, hints instead of answers, the deployment this literature would call best practice — bought a null, not a gain. Design averted the damage and built nothing. If that transfers, "design the deployment properly" is harm reduction rather than a capability strategy, and the framework should say so instead of implying the upside. Logged in knowledge/thesis-challenges.md.

Two cautions on citation. First, this paper is upstream of Catrinescu, who uses it as a calibration source; citing the two as independent confirmations of each other is partly circular, and the correct form is Bastani as the measurement, Catrinescu as the organizational extension. This is the Babina / Ewens & Giroud rule applied one link up the citation chain. Second, the skill-gap result cuts against a widely cited optimistic finding — AI helps the weakest most — by showing the compression vanishes when access is removed; any claim in the base resting on durable skill-gap compression should be re-checked against it.

Sources & Confidence

• Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. & Mariman, R. "Generative AI without guardrails can harm learning: Evidence from high school mathematics." PNAS 122(26), online 2025-06-25, issue dated July 2025. DOI 10.1073/pnas.2422633122. Read in full from the authors' own PDF (hamsabastani.github.io/education_llm.pdf), text extracted directly; all figures above quoted from the paper body, not from a summary. — High. Peer-reviewed PNAS; pre-registered field RCT; anonymized data and code released (github.com/obastani/GenAICanHarmLearning).
• Publication date cross-checked: Nature Index and Penn CHIBE both give online publication 2025-06-25; RePEc/IDEAS gives volume 122(26), July 2025. The SSRN preprint (abstract 4895486, July 2024) carries a different and stronger title and is not the version cited. — High.
• Catrinescu, I. arXiv:2610.33102 (2026) — cited here only to record the calibration dependency running from this paper into hers. — Moderate.

Filed by / Routing

Filed by: Stafford (claude-opus-5), judgment layer, exploratory slot 2026-10-27 Routing: strategy Practice asset: none: Category 6/7 measurement entry, not a Category 8 design spec. The guardrail design that produced the null (solution plus teacher-written hints supplied to the model; hints rather than answers to the student) is a design pattern with a measured outcome attached, and should be read against any autonomy-ladder or human-in-the-loop spec that claims a learning benefit — the measured learning benefit here was zero.

Tags: paper-2, practitioner, Technology Illusion, Momentum Mirage, Capability, COUNTERARGUMENT Source: ISS-Corporate URL: https://www.iss-corporate.com/press/roughly-one-third-of-large-u-s-companies-now-disclose-board-oversight-of-ai-iss-corporate-finds/ Date published: 2025-03-19 Date flagged: 2026-10-26

Summary

ISS-Corporate publishes a repeated, proxy-disclosure-based count of board AI oversight in the S&P 500. The 19 March 2025 release reports that "31.6 percent of the S&P 500 disclosed some level of board oversight of AI" as of 2024 — a rise of "more than 84 percent between 2023 and 2024" and "more than 150 percent from 2022 to 2024." Within that, the share of companies disclosing at least one director with AI expertise runs 11% (2022) → just under 14% (2023) → 20% (2024).

The prior release in the same series (21 March 2024) gives the instrument's construction. Board oversight is "defined as specific committee oversight responsibility, director(s) with AI expertise, and/or an AI ethics board," coded from "DEF 14As filed by S&P 500 companies" over filings made "between September 2022 and September 2023." On that window about 15% of the S&P 500 disclosed some board oversight of AI and 13% had at least one director with AI expertise (30% among Information Technology companies). The components of oversight other than director expertise are rare: 1.6% disclosed explicit full-board or committee AI oversight and 0.8% an AI ethics board.

Neither release states how a director was judged to have AI expertise. Both name proxy statements as the source documents; neither publishes coding criteria for the expertise determination itself. A separate, widely cited treatment of ISS director-skills data — Davine, Schoenecker & McCall (Deloitte), Director Skills for Navigating a Complex Business Environment, Harvard Law School Forum on Corporate Governance, 28 July 2025 — states in its footnotes that its skill definitions follow "ISS Corporate Services data collection guidelines," publishes verbatim definitions for Leadership, Finance and Audit, reports Technology and Information Security as tracked skills without defining them, and does not mention AI anywhere.

Five Breakpoints Intersection

Technology Illusion: ISS's own disclosure-based series counts directors with AI expertise at 20% of the S&P 500 in 2024, rising from 11% in 2022 — roughly four and a half percentage points a year — while the ISS board-composition skills dataset compiled in June 2026 codes 68% of S&P 500 companies as having at least one AI-skilled non-executive director (Morgan Stanley Institute for Sustainable Investing, 2026-08-27; held in this base). The designation of board AI competence is expanding several times faster than any count of it built from what companies are willing to assert in a filing, which is direct evidence that the competence being counted is substantially reclassification rather than acquisition.

Momentum Mirage: In the Sept 2022–Sept 2023 window the headline oversight figure (~15%) is carried almost entirely by the director-expertise component (13%), while the two components that name an actual mechanism are near-absent — 1.6% explicit full-board or committee oversight, 0.8% an AI ethics board. "Board oversight of AI" rises as a disclosure category while the artifacts that would operationalize it stay at one to two percent of the index.

Four Forces Intersection

Capability: The release measures governance substrate from filed documents rather than self-assessment, and the substrate is thin in the component that requires standing machinery — explicit committee oversight at 1.6% and an AI ethics board at 0.8% of the S&P 500 — against a director-credential component four to eight times larger.

Relevance to Brandon's Positioning

This discharges the preparation deadline on the 2027 proxy-season window (registered 2026-09-30, prerequisite due 2026-11-15): run the ISS methodology check before any drafting, because the argument rests on one vendor's proprietary skill taxonomy whose coding criteria are unpublished.

The criteria are not published, in either ISS product — so the check fails on its own terms. But it returns something better than the criteria would have: a way to bound the generosity of the taxonomy against the same vendor's own narrower instrument, which removes the need to trust the taxonomy at all. ISS sells a disclosure-based count that reaches 20% of the S&P 500 in 2024 on a trend of about four and a half points a year, and a background-coded count that reaches 68% of the S&P 500 by June 2026. Extrapolating ISS's own disclosure series to mid-2026 lands near 26–29%, not 68%.

The consequence for the window's argument is that its spine changes and strengthens. It can no longer run as "ISS codes 68% of boards as AI-skilled and 62% of those directors were seated pre-2022, therefore relabelling" — that construction asks the reader to trust an unpublished taxonomy. It can run as: the same vendor sells two measures of board AI competence; the one built from what companies disclose about their own directors says roughly a fifth of the index, and the one built from coding director backgrounds says more than two-thirds. Whatever the second number measures, it is not what companies will claim in a proxy statement. That is checkable from two ISS products and requires no methodology disclosure.

Four bounds that must travel with the comparison, and the first is the one a reviewer will reach for. (1) The two figures are not strictly comparable — different years (2024 vs June 2026), different constructs (disclosed expertise vs coded skill), and non-executive directors in the Morgan Stanley half against all directors in the ISS-Corporate half. The argument is about the size of the gap against ISS's own trend, and must be stated that way, never as a like-for-like contradiction. (2) Same vendor group, not demonstrably the same product — ISS-Corporate is the corporate-services arm; the dataset Morgan Stanley used is described only as "a dataset from Institutional Shareholder Services (ISS) compiled in June 2026." Say "the same vendor group's two instruments," not "ISS contradicts itself." (3) The Deloitte/HLS observation that AI is absent from a July 2025 treatment of ISS director-skills data is consistent with the AI tag being a recent taxonomy addition laid over an existing technologist population, which is the relabelling thesis — but a secondary analysis omitting a category is not proof the category did not exist. Dated observation, not a finding. (4) The disclosure series is self-disclosure in proxy statements, so its level is a floor on real expertise and its trend partly tracks disclosure fashion rather than board composition; that cuts against reading the gap as pure taxonomy generosity and must be conceded in print.

One named chase, deliberately not entered as a finding. The D&O Diary (Abrams, 2026-09-28) reports an ISS STOXX study of "more than 3,000 Russell 3000 and S&P 500 companies" for the 2026 proxy season finding 16% "reported having at least one director with AI expertise." If that figure is S&P 500-separable it is the decisive same-year comparison against the 68% and the growth explanation dies. It is secondhand and the primary was not located in one pass — the post identifies the source only by hyperlink, and a search for the named ISS-Corporate post ("Governance Proposals Dominate the 2026 Proxy Season") did not surface it. Under this base's standing rule that no figure enters from a search-result summary, it stays a chase. It is also pooled across Russell 3000 and S&P 500, which drags it below any S&P-500-only figure.

Sources & Confidence

• ISS-Corporate, "Roughly One-Third of Large U.S. Companies Now Disclose Board Oversight of AI", press release, Rockville MD, 19 March 2025 — Moderate. ISS-Corporate is not on CLAUDE.md's high-credibility publisher list. Read first-hand 2026-10-26.
• ISS-Corporate, "Roughly 15 Percent of Large U.S. Companies Disclose Board Oversight of AI", press release, 21 March 2024 — Moderate. Read first-hand 2026-10-26; source of the DEF 14A construction and the component breakdown.
• Davine, Schoenecker & McCall (Deloitte LLP), Director Skills for Navigating a Complex Business Environment, Harvard Law School Forum on Corporate Governance, 28 July 2025 — Moderate. Read first-hand 2026-10-26. Cite only for what it states: that its definitions follow ISS Corporate Services data collection guidelines, the verbatim Leadership/Finance/Audit definitions, and the absence of an AI category. It gives no link or title for the guidelines and does not say whether they are published.
• Morgan Stanley Institute for Sustainable Investing, Governance and AI Adoption, 2026-08-27 — Moderate. Held in this base since 2026-09-30 (knowledge-base/morgan-stanley-board-ai-expertise-governance-allocation-2026-08.md); source of the 68% / 24% / 62% figures and of the ISS June 2026 dataset attribution. Not re-verified today.
• The D&O Diary (Sarah Abrams), 2026-09-28 — Low for the 16% ISS STOXX figure: secondhand, primary unlocated, pooled index. Logged as a chase, not cited as evidence.

Filed by / Routing

Filed by: Stafford (judgment layer, 2026-10-26) Routing: governance / strategy

Tags: paper-2, Technology Illusion, Capability, COUNTERARGUMENT Source: Ramp Economics Lab / Revelio Labs (working paper) URL: https://ramp.com/data/ai-jobs-impact/paper Date published: 2026-06-30 Date flagged: 2026-10-25

Instrument (recorded before the finding, per CLAUDE.md hard rule 9)

• Authors: Ara Kharazian (Lead Economist, Ramp; corresponding), Lisa Simon (Chief Economist, Revelio Labs), Ryan Stevens (Director of Applied Science, Ramp).
• Adoption measure — this is the novel leg: observed AI vendor payments from Ramp card and bill-pay line-item data (the Ramp AI Index, Kharazian 2026). A firm is an adopter once it records at least $100 of AI vendor spend in each of three consecutive months; treatment is absorbing. Intensity split into High/Low by spend per employee.
• Workforce measure: Revelio Labs workforce records, including Revelio's seven-point seniority field — entry-level = seniority 1–2, non-entry = 3–7, manager-plus = the upper band.
• Linkage: exact domain matching between Ramp businesses and Revelio companies. n = 21,559 US firms.
• Observation window: January 2021 – February 2026. Event window e ∈ [−12, +24] months around adoption.
• Estimation: Callaway & Sant'Anna (2021) staggered-adoption difference-in-differences. Preferred design compares adopters to not-yet-adopters within the same eventual intensity group; NAICS sector fixed effects and no other controls. Identifying assumption stated by the authors as conditional parallel trends within eventual intensity group. Pre-period diagnostic "flags" (months −12 to −2 whose 95% CI excludes zero) reported per outcome.
• Genre: telemetry on the adoption side (payment records), resume_panel on the outcome side (Revelio). Partially, not wholly, independent of the postings/résumé genre — see the Résumé-Panel Consensus caution below.

Summary

The paper studies how firm employment changes after a firm starts paying for generative AI, using observed vendor spend rather than occupational exposure indices or job-advert text. The authors motivate the design against the existing literature explicitly: exposure measures "cannot distinguish a firm that has adopted AI from one that has not," and the unit of observation in postings-based work "remains the occupation," which "cannot identify firm-level adoption decisions."

The headline results are in levels. For high-intensity adopters against not-yet-adopters: total headcount +0.097 log points (≈10.2%, p<0.01, 3/11 pre-period flags); total entry-level headcount +0.113 (≈12.0%, p<0.01, 2/11 flags); non-entry +0.074 (≈7.7%, p<0.01); manager-plus headcount +0.065 (≈6.7%, p<0.01, 0/11 flags — the cleanest diagnostics in the table). Low-intensity adopters show no statistically significant change on any headline outcome (total headcount −0.006, n.s.). Growth is broad across functions — sales +10.3%, admin +7.8%, engineering +7.3%, entry-level engineering +6.3%, customer service +6.3% — and the authors report "there does not appear to be a strong correlation between a function's perceived AI exposure and its increase or decrease in headcount." Operations is the only category with no gain. The authors state the results "counter predictions that AI adoption will lead to broad job loss."

The composition results run in the opposite direction to the levels, and the sign flips across intensity. In workforce shares (Table 4), for High-intensity adopters: entry-level share +1.147pp (p<0.05), non-entry share −1.142pp, manager-plus share −1.517pp (p<0.05), all with 0/11 pre-period flags. For Low-intensity adopters the signs reverse: entry-level share −0.520pp, manager-plus share +0.655pp (both p<0.05). The authors are explicit that the share movement is mechanical given the levels — entry-level headcount "expands enough relative to the rest of the firm" that its share rises — not a reduction in managers. Two further bounds are the authors' own: gains "are concentrated in Information and, at present, are statistically significant only in that sector group" (Information high-intensity total headcount +13.4%, no flagged pre-periods), with remaining sectors "close to zero, imprecise, or both"; and adopters "are already larger, more technical, faster-growing firms." A never-adopter specification yields larger estimates (6.5% Low / 12.6% High) which the authors demote — "adopters were already growing at a faster rate relative to never-adopters: total-headcount pre-period flags are 11 of 11 for low intensity and 8 of 11 for high intensity. We therefore include this specification for completeness rather than as the preferred design." Ramp spend is interpreted as "a revealed-adoption measure for a selected, business-spend-active firm population."

Five Breakpoints Intersection

Technology Illusion: The dose-response sign reversal is direct evidence that the technology does not carry the outcome on its own — low-intensity adopters, who have bought AI and show no significant employment change on any headline outcome, are the group whose manager-plus share rises (+0.655pp) and whose entry-level share falls (−0.520pp), while the firms investing materially show the opposite composition and 10.2% growth. Buying the tool and getting the organizational result are separated here by a measured threshold, which is the Technology Illusion claim stated as a dose-response rather than as a case study.

No Momentum Mirage mapping is claimed: the paper measures realized headcount from payroll-adjacent records, not self-reported progress, so there is no gap between appearance and movement for it to evidence. No Incentive Fragmentation, Process Friction or Strategic Disconnection mapping — the paper contains no incentive, process or strategy variable of any kind.

Four Forces Intersection

Capability: The sector result is a Capability finding in Brandon's sense rather than a technology finding — the gains are "statistically significant only in" Information, and the authors attribute this to where "the most commercially mature AI gains appear clearest," i.e. to the firms that already possessed the engineering capacity to put coding agents into their core production. The same spend in sectors without that substrate produces estimates "close to zero, imprecise, or both."

No Purpose, Commitment or Momentum tags: the instrument contains no measure of stated intent, follow-through, or reinforcement. Tagging them would be a guess and hard rule 2 forbids it.

Relevance to Brandon's Positioning

This is the strongest counterargument instrument the base holds against the junior-displacement reading, and it arrives on better identification than the literature it contradicts. The apprenticeship-rung thread and Xu et al.'s automation/augmentation model both predict the entry rung thinning at adopters; at firms whose adoption is observed by payment rather than inferred from advert text, entry-level headcount is the fastest-growing headline outcome (+12.0%, p<0.01, 2/11 flags). It is also the nearest thing the base has to the Regime Fork thread's long-unmet criterion — management-layer share measured against AI adoption intensity — and it must be filed with that bound stated: Revelio's seniority field is a job-title classifier, so this is manager share by title band, not spans of control, reporting lines, decision rights or the manager wage premium. Those three remain unmet.

For paper-2 Claim 5 it does not close gap item 1: the right-hand side is AI adoption intensity, which is technological, not an organizational variable. What it does supply is the other half of the chain the base has been missing — a credibly identified adoption→outcome link — against which an organizational moderator could be placed by someone with the data.

For asset 2 (The Flattening Bet) it is load-bearing and it cuts two ways, and this is the entry's most important use. It strengthens the draft's concession paragraph: manager-plus headcount +6.7% here converges almost exactly on Chandar & Klein Teeselink's senior +6.7% from a different adoption instrument, so "no thinning of management at adopters" now has two instruments agreeing on the number. But it removes the junior half of the concession: C&KT's junior −2.5% was the statistically weak half of their result, and this better-identified design finds entry-level rising at p<0.01. The draft must not cite the Ramp share result as measured flattening — manager headcount rose; only its share fell, because juniors grew faster. A reader who checks will find that, and the essay's own standard is that it does not make the error it charges others with.

Sources & Confidence

• Kharazian, Simon & Stevens, A New Look at AI's Impact on Jobs: Firm-Level AI Spending and Workforce Adjustment, Ramp Economics Lab with Revelio Labs, 2026-06-30 — Moderate. Read in full from the primary PDF (pdftotext); all figures above are from Tables 3–6 and §§3–6 of the paper itself, not from coverage. Rated Moderate rather than High on publisher: this is a vendor economics lab publishing on its own transaction data, not a peer-reviewed or NBER/CESifo working paper, and the firm population is Ramp's own customers. Methodologically it is stronger than its publisher class — Callaway–Sant'Anna estimator, not-yet-treated comparisons, pre-period diagnostics reported per outcome, and the weaker never-adopter specification explicitly demoted by the authors.
• Acknowledgements name Nick Bloom, Scott Cunningham, Martha Gimbel, Jon Hartley, Torsten Sløk, Noah Smith and Ernie Tedeschi as having given feedback — context only, not endorsement, and not a basis for raising credibility.
• Non-independence caution: Lisa Simon is Revelio Labs' Chief Economist and the workforce/seniority data is Revelio's. This entry may not be cited as independent corroboration of Chandar & Klein Teeselink or Hosseini Maasoum & Lichtinger on the seniority-composition question — the outcome measure is the same vendor's classifier. The adoption measure is genuinely independent of the postings genre. See the Résumé-Panel Consensus pattern.

Filed by / Routing

Filed by: Stafford (judgment layer, 2026-10-25) Routing: strategy — bears directly on asset 2's dated wager; the composition result must not be read as delayering.

Tags: paper-2, Technology Illusion, Commitment, COUNTERARGUMENT Source: Revelio Labs (research note, author Nate Lawrence) URL: https://www.reveliolabs.com/news/ai-and-work/is-ai-just-a-scapegoat-for-layoffs Date published: 2026-08-25 Date flagged: 2026-10-25

Instrument (recorded before the finding, per CLAUDE.md hard rule 9)

• Cohort definition: companies that publicly attributed layoffs to AI in news coverage, identified using System2. This is a media-attribution cohort, not a sample of firms that cut layers.
• Adoption measure: Revelio Labs' "AI adoption rate," defined in the note as "the percentage of job postings for roles that can either build AI systems or leverage AI tools." Postings-text inference — the same genre and the same vendor as Chandar & Klein Teeselink and Hosseini Maasoum & Lichtinger.
• Comparison: size-matched industry peers.
• Observation window: the 24 months before each announcement, plus a persistence check six months after.
• Sample size: not stated anywhere in the note. No confidence intervals, no estimator, no identification strategy — this is a descriptive vendor research note, not a working paper.
• Last updated 2026-10-22 (published 2026-08-25), so the figures are current as of three days before flagging.
• Genre: postings.

Summary

The note asks whether companies blaming layoffs on AI are making a genuine workforce shift or using AI as cover. Its answer is mixed and it reports both sides. On the "genuine shift" side, it finds that firms attributing layoffs to AI had a median AI share of headcount nearly double that of size-matched peers, both before the announcement and six months after it — which the author reads as evidence the elevated AI orientation is real and persistent rather than rhetorical.

The decomposition underneath that headline runs the other way, and is the more informative part. Over the two years before announcement, the layoff cohort's AI headcount grew a median of just over 11%, against just over 13% for size-matched peers — that is, slightly slower AI build-out than their peers — while their non-AI headcount fell more than 3%, against roughly flat for peers. So the cohort is distinguished from its peers not by accelerating AI hiring but by contracting everything else.

The note's own summary of how much of the cohort fits the AI story is internally inconsistent and unreconciled: the body states that "nearly half of layoff companies do demonstrate shifts in their workforces toward AI," the conclusion states that "more than half of firms attributing their layoffs to AI did show strategic shifts in their workforces," and a summary bullet states that "nearly half the cohort does not fit the story." Named examples are given on several sides: the Washington Post, Dropbox and UPS as firms shifting toward AI; Atlassian, Meta and Coinbase as firms that had grown headcount sharply beforehand (the note raises over-hiring as the alternative explanation); Block and McKinsey as firms already two years into headcount reduction before the AI attribution; and Cloudflare as an apparent shift away from AI. The note does not mention management layers, spans of control, or delayering anywhere.

Five Breakpoints Intersection

Technology Illusion: The cohort's AI headcount grew slower than its peers' (11% vs 13%) while its non-AI headcount fell more than 3% against peers' flat — so the technology is carrying the public justification for a structural contraction that the firm's own comparative AI build-out does not support. That is Technology Illusion at the level of the stated rationale: the announcement asserts a technological cause whose measurable footprint is, relative to peers, slightly below average.

Momentum Mirage is not tagged, and the reason is worth recording because it would be the tempting tag: the note's persistence check — AI share still nearly double peers' six months after the announcement — argues against reading this as announced progress that then fails to materialize. The data cuts the wrong way for that mapping and tagging it would be the speculative-tag failure hard rule 4 forbids.

Four Forces Intersection

Commitment: Commitment is follow-through — decision-makers acting consistently with stated direction under pressure. This cohort states AI as the direction and then builds AI capacity more slowly than its size-matched peers while cutting the rest, which is a measured divergence between stated direction and resource allocation. Read with the note's own caveat: it is a median comparison with no sample size and no interval, so this is a directional reading, not an estimate.

No Purpose, Capability or Momentum tags — the note contains no measure of intent clarity, organizational capacity, or reinforcement.

Relevance to Brandon's Positioning

This is the source that bears directly on the wager Brandon dated on 2026-10-25 in drafts/the-flattening-bet.md, and it stresses it. The draft's reframed bet is that "the delayering cohort and the AI-adopting cohort barely overlap," and it tells the reader this "is checkable today, by anyone, with no new data." The check has been published since 2026-08-25 and was updated on 2026-10-22, by the same vendor whose corpus supplies the draft's concession — and its answer is not "barely." Firms attributing layoffs to AI carry a median AI share of headcount nearly double their size-matched peers'. On this instrument the two populations overlap substantially, and the stronger phrasing loses to a reader with one search.

The repair is in the same note and it is a better argument than the one it replaces. The cohort's AI headcount grew slower than peers (11% vs 13%) while its non-AI headcount fell. So the defensible claim is not these are not AI firms but these are firms whose AI build-out is unremarkable against their peers and whose only distinguishing move is the cut — the technology does argumentative work for a contraction that its own relative investment does not explain. That survives the overlap finding, and it converts the overlap from a refutation into the premise. It also composes with the Ramp result filed the same day: measured high-intensity adoption is associated with growth (+10.2%) concentrated in Information, so within the sector that cuts most on AI belief, the adoption evidence predicts hiring.

Bounds that must travel with any use. The cohort is defined by media attribution, not by layer reduction, so it is evidence about firms that said AI, not about delayering — which is the object asset 2 actually bets on, and no instrument yet measures that. The note gives no sample size, no estimator, and contradicts itself on the headline share ("nearly half" vs "more than half", unreconciled). And it is the same vendor and the same postings genre as Chandar & Klein Teeselink: Brandon cannot cite C&KT at full strength as the strongest objection to his essay and then discount this note on vendor grounds. Both or neither.

Sources & Confidence

• Nate Lawrence, AI Layoffs: Real Workforce Strategy or Scapegoat?, Revelio Labs, published 2026-08-25, last updated 2026-10-22 — Low. Vendor research note with no stated sample size, no estimator, no intervals, and a self-contradictory headline figure. Read from the primary. Rated Low on methodology despite being the only instrument that addresses the question, and it must be cited as the only available check rather than as a reliable magnitude.
• Cohort identification via System2; workforce and postings data via Revelio Labs — Moderate as data providers, not independent of the base's existing Revelio-based entries.

Filed by / Routing

Filed by: Stafford (judgment layer, 2026-10-25) Routing: strategy — a dated, publication-bound wager rests on the claim this note contradicts; the repair is in the same note.

Tags: paper-2, COUNTERARGUMENT Source: Autonomous Policy Evaluation Project (APEP), "APEP Autonomous Research" URL: https://ape.socialcatalystlab.org/papers/apep_0537/v1/paper.pdf Date published: 2026-03-06 Date flagged: 2026-10-24

⛔ DO NOT CITE. This file exists so that nobody in this system cites it.

This paper has no human author. Its byline is "APEP Autonomous Research"; its only named contributor is a GitHub handle, @olafdrw; and its acknowledgements, on page 29, read verbatim: "This paper was autonomously generated using Claude Code as part of the Autonomous Policy Evaluation Project (APEP)." The title page carries "This paper was generated autonomously. Total execution time: auto (cumulative: auto)" and correspondence to scl@econ.uzh.ch. Project repository: github.com/SocialCatalystLab/ape-papers.

It surfaces in web search as "the first independent replication" of the seniority-bias finding — which is how this session found it, while running the owed O*NET/occupational-rank venue pass. A pipeline that triaged it on its abstract would have ingested a synthetic artifact as a peer-grade replication contradicting the base's own holdings.

The fabrication signature, which is the reusable part

The identifier is real and the human-readable metadata around it is confabulated. Its reference list reads, verbatim:

"Maasoum, Negar Hosseini and Jeremy Lichtinger, "Generative AI as Seniority-Biased Technological Change: Evidence from 14 Million Resumes," Working Paper 5425555, SSRN 2025."

Four errors in one entry, against a correctly cited SSRN working-paper number:

The invented first names are the tell, and they are the cheapest check available: verify an author's given name, not just the identifier. The body repeats the error — "Hosseini Maasoum and Lichtinger (2025) analyze 14 million résumés" — understating the source it claims to replicate by a factor of roughly four and a half.

Summary (what it asserts, reported so the claims are recognizable if they resurface)

The paper reports that between 2015 and 2024 the entry-level share of U.S. employment fell from 50.2% to 45.7% while the senior share rose from 29.0% to 33.3%; that industries above median AI exposure saw a 1.8 percentage point larger decline in entry-level share after 2022 and a 2.2 percentage point larger rise in senior share (both p < 0.05); and — the finding that would have mattered — that an event study shows a statistically significant pre-trend from 2015 onward, predating ChatGPT's release, so the shift "cannot cleanly attribute" itself to generative AI. Its own conclusion states the differential pattern is "entirely driven by pre-existing secular trends rather than a discrete post-2022 break."

Its three named datasets are all real and all public: BLS Occupational Employment and Wage Statistics, O*NET Job Zone classifications for seniority, and Felten, Raj & Seamans AI Occupational Exposure scores. That is exactly why it reads as credible.

Five Breakpoints Intersection

No direct Five Breakpoints mapping. Logged as a provenance and method record, not as evidence about organizations. Its one durable contribution is negative: a documented instance of a synthetic primary — an AI-generated working paper, on this base's exact topic, with real public datasets, a real working-paper identifier and fabricated author names, circulating in search results as a replication.

Four Forces Intersection

No direct Force implication.

Relevance to Brandon's Positioning

The claim it makes is the right question and it cannot be sourced here, and that gap is worth more to Brandon than the paper would have been. If the entry-level employment decline has a pre-trend running from 2015, then a large part of the seniority-bias literature is measuring a secular shift and attributing it to generative AI. That is a serious objection to the fastest-moving genre in this base. It is cheap to test, on entirely public data (BLS OEWS + O*NET Job Zones + Felten AIOE), and this session found nothing credible that has run it — only this. Registered as an acquisition target rather than a finding.

Second, it is the sharpest example available of the argument The Flattening Bet already makes — that nobody reads the instrument. A paper that fabricated the names of the authors it claimed to replicate sat in search results as an independent replication. Offered as optional illustrative material for that essay in knowledge/asset-suggestions.md; the essay's research punch list is finished and this does not reopen it.

Third, and the uncomfortable one worth stating plainly: the agent writing this file and the agent that wrote that paper are the same tool. The defence is not instinct, it is the standing rule — go to the primary, and check the author's name against it.

Sources & Confidence

• APEP, "Is Generative AI Seniority-Biased? Evidence from U.S. Occupational Employment Data," 2026-03-06 — Low, and excluded from the evidence base by provenance. Full text retrieved and read via pdftotext (web fetch could not parse the PDF); all quotations above are verbatim from the extracted text.
• Hosseini Maasoum & Lichtinger, "Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data," SSRN 5425555 — Moderate, and unread at primary (SSRN returned HTTP 403). Author names, affiliations, title, Revelio/WRDS data source and sample figures verified through search-result quotation of the SSRN abstract and the October 2025 version, not by reading the paper. Previously logged as "the highest-value chase on this list" in research-evidence knowledge/source-candidates.md, known only through Friebel et al.'s footnote 1.

Filed by / Routing

Filed by: Stafford (stafford-research, exploratory slot 2026-10-24) Routing: governance — a new standing caution for research-evidence knowledge/sources.md; strategy — asset 2 optional material Practice asset: none.

Tags: paper-2, COUNTERARGUMENT Source: SSRN working paper (Bharat Chandar, Bouke Klein Teeselink; linked from the Stanford Digital Economy Lab) URL: https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7498743 Date published: 2026-09-20 Date flagged: 2026-10-24

Instrument (recorded before the findings, per CLAUDE.md hard rule 9)

• Data source: Revelio Labs — public profile and job-board data scraped from LinkedIn and job boards, the same vendor and the same genre as Hosseini Maasoum & Lichtinger (2025). Over a billion job postings searched; 41 countries. The employment-record count is not stated on the authors' own page (secondary coverage gives 154 million; unverified at primary).
• Inference method: a firm is scored an adopter if its job postings indicate generative AI is used in some part of its work — postings searched for terms including ChatGPT, Claude, Gemini, "large language model," prompt engineering, copilot, retrieval augmented generation, then screened by an LLM to drop false matches and postings that mention AI only to restrict it. Seniority is Revelio's own standardized seniority label. Unit of analysis is the foreign affiliate of a multinational (a company's operations in one country), with parent-headquarters metro-area adoption rates used as an instrument; matched adopter/non-adopter comparison.
• Observation window: pre-period anchored on the November 2022 ChatGPT launch; five-year horizon in secondary coverage. The authors state pre-ChatGPT trends are "not perfect."
• Genre dependency: this is a résumé/postings-inference instrument and belongs under the ESCALATED Résumé-Panel Consensus pattern. It is not independent of Hosseini Maasoum & Lichtinger — same vendor (Revelio), same underlying LinkedIn corpus. Never cite the two as mutually confirming.

Summary

Chandar and Klein Teeselink ask what happens to the shape of a firm's workforce after that firm starts using generative AI, across 41 countries. They identify adopters from job-posting text, then compare adopting affiliates of multinationals against matched non-adopting affiliates.

Their headline composition result is that the junior share of the workforce falls by about two percentage points at adopters, and it falls in 23 of the 31 countries where estimates are statistically precise. But the decomposition is the part that matters: the authors state the junior-share decline comes primarily from growth in senior employment rather than from a fall in junior employment. Senior employment rises 6.7%; junior employment falls about 2.5%; total employment at adopters ends up roughly 3.3% higher than at non-adopters, which the authors read as modest productivity gains.

The significance pattern runs the same way. The 3.3% total-employment increase is described as marginally significant. The junior-side estimates are the weak ones: the junior shift away from AI-exposed jobs (0.6 percentage points) and the junior labour-saving effect (about 2%) are both not statistically significant, and the junior-share decline is significant in only seven countries. On the senior side the labour-expanding effect is about 5% and the shift toward exposed jobs is 0.9 percentage points.

Five Breakpoints Intersection

No direct Five Breakpoints mapping. This measures labour-demand composition, not an organizational condition — there is no purpose, incentive, process or momentum variable anywhere in it, and no concrete sentence can tag a breakpoint from a specific claim in the source without importing an inference the authors do not make. Logged as COUNTERARGUMENT evidence bearing on paper-2 Claim 1 (hierarchy as an information-routing solution) and on the Regime Fork story thread, and as a genre member of the Résumé-Panel Consensus pattern. Not staged — the staging gate requires at least one breakpoint rationale, and per hard rule 4 a weak rationale means the tag is wrong.

Four Forces Intersection

No direct Force implication — the instrument contains no Force variable. No Forces tags, per the rule against guessing them.

Relevance to Brandon's Positioning

This is the largest firm-level test yet of the prediction that AI flattens organizations, and it finds the opposite sign — which cuts for Brandon's thesis and against his Tier 1 essay's framing at the same time. Coordination-compression models (Farach; Klein & Wieczorek) predict that cheaper coordination widens spans and thins hierarchy. At firms whose AI adoption is actually detectable, headcount grows about 3.3% and the senior end thickens by 6.7%. Nothing here looks like delayering.

The constructive reading, and it is a better line than the one The Flattening Bet currently carries: the delayering wave being attributed to AI is not happening at the firms that measurably adopt AI. That makes the "AI redundancy washing" confound the base logged on 2026-10-19 (layoffs that would have happened anyway, re-attributed to automation) a measurable asymmetry rather than a quip, and it is a stronger support for the essay's central prediction than any source the essay now holds — the structural change the narrative promises is absent precisely where the technology is present.

The hostile reading has to be carried in the same breath, because a reader will find it: the wave your essay predicts will fail is invisible in the best firm-level data, so what exactly is failing? Logged to knowledge/thesis-challenges.md 2026-10-24.

Two bounds that must travel with every use. (1) "Senior" is a Revelio seniority label, not a management layer — senior individual contributors are senior. This says nothing about spans of control, layer counts or the manager wage premium, so the Regime Fork thread's criterion stays unmet. (2) Adoption is a lower bound measured from job-posting text, the population is digitally connected multinational affiliates, and the authors call their matching and instrument "not a foolproof approach" with country patterns "mostly suggestive."

Sources & Confidence

• Chandar & Klein Teeselink, "How Does AI Change Labor Demand? Evidence from 41 Countries," SSRN 7498743, 2026-09-20 — Moderate. SSRN working paper, not peer reviewed; SSRN landing page returned HTTP 403 to this session, so the paper itself is unread at primary.
• Bharat Chandar, "AI's impacts on jobs around the world," bharatchandar.substack.com, 2026-09-21 — Moderate, and this is the author's own copy, which is the route hard rule 14 prescribes. All figures above are verbatim from it.
• Secondary coverage (Business Standard, phys.org, businessmirror.com.ph) reporting a 3% junior decline against the author's own 2.5%, and 154 million employment records — Low. Do not quote the secondary figures. The discrepancy is itself the reason the author's own copy is the standing route.

Filed by / Routing

Filed by: Stafford (stafford-research, exploratory slot 2026-10-24) Routing: strategy — asset 2 (The Flattening Bet), paper-2 Claim 1, Regime Fork thread, Résumé-Panel Consensus pattern Practice asset: none — not a Category 8 prescriptive-design source.

Tags: paper-2, Momentum Mirage, COUNTERARGUMENT Source: Journal of Substance Use and Addiction Treatment (Elsevier), 165:209437 URL: https://doi.org/10.1016/j.josat.2024.209437 Date published: 2024-10-01 Date flagged: 2026-10-21

Instrument (recorded BEFORE the finding, per CLAUDE.md hard rule 9)

Read grade — state it first, because it bounds everything below. This entry is abstract-and-registration grade. The article is subscription-gated (Europe PMC: isOpenAccess: N, inPMC: N, no PMC id, no author copy located in this run). What was read in full and verbatim: the authors' structured abstract via the Europe PMC REST API, and the ClinicalTrials.gov registration record NCT03042832 via the v2 API. The full text was not read. No claim below rests on inference about unreported content, and the places where the full text could overturn a reading are named as such.

Data source genre. Repeated web-based provider self-report surveys (leadership, implementation climate, implementation citizenship behaviour), plus one administrative/process count — the number of motivational-interviewing sessions a provider recorded and submitted for fidelity coding.

Inference method. Multiple-cohort cluster randomized trial. 60 clinics across nine behavioral health organizations in California and Arizona; clinics randomized within organization to LOCI or to a leadership-training-webinar control, in three consecutive cohorts. n_LOCI = 201, n_Control = 179 direct service providers. Multilevel autoregressive modeling, providers (level 1) nested within clinics (level 2). Motivational-interviewing reach tested by independent-sample t-test, reported as chi-square (1, n = 370) = 5.59, p = .018.

Observation window. Trial start 2015-09-30; primary completion 2020-04-20 (ClinicalTrials.gov, actual dates). Registered 2017-02-02, retrospectively — roughly sixteen months after the trial began. Follow-up assessments predicted at 4, 8 and 12 months. Pre-generative with respect to AI, and the intervention is not technological.

Shared-programme flag — this is a network-level non-independence, not an author-level one. Same intervention (LOCI), same research group (Aarons, Ehrhart) as the two WISDOM-trial entries already in this base: williams-aarons-ehrhart-loci-leadership-randomized-digital-uptake-2026-03.md and williams-ehrhart-aarons-loci-mechanism-implementation-leadership-climate-fidelity-2024-03.md. The registration is separate (NCT03042832 vs NCT04096274), the sector is separate (substance use treatment vs youth mental health), and the randomization is separate — so this is a second trial, not a re-analysis. It is not an independent confirmation of the LOCI programme, and must never be cited as one.

Summary

LOCI — Leadership and Organizational Change for Implementation — is a manualized, twelve-month implementation strategy that works through "iterative cycles of leadership and climate assessment and feedback, leadership training and coaching, and strategic planning with upper-level leaders." The authors state the study "tested the effects of LOCI on transformational and implementation leadership, implementation climate, implementation citizenship behavior, and EBP reach."

The reported results are: "between condition differences at 4 months for implementation leadership, implementation climate, and implementation citizenship behavior such that greater improvements were evidenced in the LOCI condition compared to the control condition." On reach: "Reach of MI was significantly greater in the LOCI vs control condition such that LOCI providers were significantly more likely to engage in the fidelity monitoring process (chi-square (1, n = 370) = 5.59, p = .018)."

The conclusion states that "The LOCI implementation strategy resulted in more positive hypothesized outcomes compared to the control condition," and that organizational change strategies "have utility for implementing health innovations in complex, multilevel contexts and for greater sustainment of facilitative leader behaviors, strategic implementation climate, and improved implementation outcomes."

Five Breakpoints Intersection

Momentum Mirage. The conclusion claims the strategy produces "greater sustainment of facilitative leader behaviors," while the between-condition differences the abstract reports are at 4 months, against a registration (NCT03042832) whose primary endpoints — Implementation Leadership and Full Range Leadership — are both specified at 12 months, with results never posted to ClinicalTrials.gov despite primary completion on 2020-04-20; a sustainment claim resting on the earliest of three pre-specified timepoints is the appearance of durable change standing in for demonstrated durability, which is the breakpoint exactly.

No second breakpoint is tagged. The measurement-substitution reading — a registered secondary outcome of "Fidelity" at 12 months reported instead as providers' willingness to submit sessions for fidelity coding — is the same Momentum Mirage mechanism (a participation proxy occupying the place of the practice outcome), not a separate breakpoint, and tagging it twice would inflate the count.

Four Forces Intersection

No Forces tags — the source is subscription-gated and was not read in full (CLAUDE.md hard rule 2). forces: [] is the correct value, not missing data. Momentum is the obvious candidate given the sustainment language, and it is deliberately withheld: the Forces layer is a mechanism claim, and the abstract does not carry enough of the mechanism to earn one. A full-text read is the named condition for adding it.

Relevance to Brandon's Positioning

This cuts both ways on the LOCI prior-art challenge (knowledge/thesis-challenges.md, logged 2026-10-08, OPEN), and the half that hurts should be stated first. The objection gets stronger. The competing manualized remedy now has two separately registered cluster randomized trials, in two sectors, across roughly 80 clinics, applying one undifferentiated twelve-month protocol and reporting positive results in both. Brandon's differentiating claim — that five failure modes are distinguishable and that knowing which is live changes the remedy — remains untested, and today's sweep did not test it; it watched the undifferentiated alternative acquire a second sector. A hostile reader of Asset 12 can now say: the competing cure has eighty randomized clinics and you have a taxonomy.

The half that helps is the premise, and it is the more interesting half. The sharp form of the objection is why buy a five-way diagnostic when one protocol has trials behind it? — which presupposes that general organizational quality is the right level of description. This trial registered Full Range Leadership (the transformational/MLQ construct, and the thing essentially the whole leadership-development market sells) as a co-primary outcome at 12 months, and it appears nowhere in the reported results. In the WISDOM mechanism paper already in this base, the same construct was measured and found inert — mediating 0.29 (p = .136, null) against implementation leadership's 0.82 (p = .004), and barely moved by the intervention. So across the LOCI programme the generic construct is a registered primary outcome in one trial and an explicitly null mediator in another, while the construct specific to the thing being implemented carries the effect in both. The principle that you must name the specific thing rather than improve the organization in general is a finding of the programme being used against Brandon — now in two sectors.

The bounded conclusion for Asset 12 (master-plan Tier 2), which is the gate this find serves: concede the prior art at full strength and cite both trials by registration number; then narrow to the two things LOCI does not do — it moves uptake and participation, not converted value, and it applies one protocol rather than discriminating among failure modes — and note that its own programme twice found the general construct inert. The defensible sentence is unchanged in shape and stronger in support: that organizational conditions gate whether technology gets used is established by randomized trial; which condition is gating, and whether matching the remedy to it matters, is not.

What must not be claimed. That the 8- and 12-month effects were null. The abstract reports the 4-month differences; the paper analysed all three timepoints and the full text may report the later ones. What is established is the reporting pattern against the registration, not the sign of the unreported results.

Sources & Confidence

• Aarons GA, Sklar M, Ehrhart MG, Roesch S, Moullin JC, Carandang K. "Randomized trial of the leadership and organizational change for implementation (LOCI) strategy in substance use treatment clinics." J Subst Use Addict Treat 2024 Oct;165:209437. DOI 10.1016/j.josat.2024.209437, PMID 38866139. Abstract read verbatim via Europe PMC REST API — High (peer-reviewed trial report; abstract-grade read only).
• ClinicalTrials.gov NCT03042832, registration record via v2 API: registered 2017-02-02 retrospectively, start 2015-09-30, primary completion 2020-04-20, primary outcomes Implementation Leadership and Full Range Leadership at 12 months, secondary outcomes Implementation Climate / Attitudes Toward EBP / Implementation Citizenship Behavior / Fidelity at 12 months, hasResults: false — High (primary registry record).
• Cross-reference, already in base: Williams, Ehrhart, Aarons et al., Implementation Science 2024;19:29 — High. Same research group; used here only for the transformational-leadership comparison, never as corroboration.

Filed by / Routing

Filed by: Stafford (tgifriday7/stafford-research), Wednesday falsification sweep, 2026-10-21 Routing: strategy — feeds master-plan Tier 2 (Asset 12, "What the Five Breakpoints Owe") and updates the LOCI challenge in knowledge/thesis-challenges.md Practice asset: none — this is implementation-science prior art for the differentiation argument, not a Category 8 AI-native design spec.

Tags: paper-2, Capability, COUNTERARGUMENT Source: Cruces, Fernández Meijide, Galiani, Gálvez & Lombardi — arXiv / NBER WP 34851 / CEPR DP 21299 / Universidad Torcuato Di Tella School of Government WP 2026-03 URL: https://arxiv.org/abs/2608.04198 Date published: 2026-08-04 Date flagged: 2026-10-18

Summary

A randomized online experiment with 1,174 adults aged 25–45 drawn from heterogeneous educational backgrounds. Each participant completed an incentivized, workplace-style business problem-solving task either with or without access to a generative-AI assistant, and then completed a second task unassisted.

The headline result is strong equalization on education. AI raised performance for everyone, and the gains were substantially larger for participants with less education. Without AI, participants with more education outperformed those with less by 0.548 standard deviations; with AI, that gap fell to 0.139 — roughly three-quarters of the initial difference closed. In plain terms: the measured performance advantage that more schooling bought on this task mostly disappeared once both groups had the assistant.

Two qualifications the authors supply themselves. First, the chat logs show that the two groups used the tool differently — lower-education participants obtained substantial assistance, while higher-education participants used the assistant more effectively. Second, the equalization is partly tool-dependent rather than learned: on the unassisted follow-up task, treated participants did not perform worse than controls and lower-education participants retained part of their improvement, but a sizable education gap reappeared. The authors' own conclusion is that generative AI narrows effective productivity differences in task execution while human-capital differences continue to shape both unassisted performance and the quality of tool use. They also note that most existing evidence studies AI inside firms and occupations, where organizational selection has already compressed educational variation.

Five Breakpoints Intersection

No direct Five Breakpoints mapping — this is an individual-level task experiment with no organization in it. Logged because it is one of the five instruments that fix the individual-level baseline against which the framework's organizational claim has to be stated, and because the base has been carrying a false version of its finding since 2026-10-18 (see below).

Four Forces Intersection

Capability — the experiment separates two things the Capability force usually runs together: capability in execution and capability in tool use. The first equalizes almost completely (0.548 → 0.139 SD); the second does not, and the authors say so directly — higher-education participants "used AI more effectively" even as the output gap closed. Any capability-building prescription that treats "give people the tool" and "make people good with the tool" as one intervention is measuring the first and selling the second.

Relevance to Brandon's Positioning

First, a correction. Today's research-evidence scan recorded Cruces et al. as finding "productivity gains but no equalization across educational, occupational, or demographic groups" — in the Brosch et al. KB file, in knowledge/story-threads.md, and in the scan brief's PATTERN BUILDING section, where it is one of the two primaries said to contradict the equalization result. That sentence in Brosch et al. describes Haslberger et al. (2025), the next paper in the paragraph, not Cruces et al. Brosch et al.'s actual words on Cruces are: "find that AI closes about 75% of the education-based productivity gap." Verified against the authors' own copy (arXiv 2608.04198) and their institutional working-paper listing, both of which report the 0.548 → 0.139 figures. Cruces et al. belongs in the equalization column, and it is the strongest result in it.

Second, what it is actually worth. This is the cleanest statement available of the proposition Brandon's framework has to survive: give people a good tool on a well-specified task and most of the human-capital advantage in output evaporates. It belongs in the corpus as a conceded premise, not a contested one — the same move Paper 3 made with the 70% figure. The framework's claim is not that individuals don't gain; it is that individual gains do not aggregate without organizational conditions, and conceding the individual result at full strength is what makes the aggregation argument credible rather than defensive.

Third, the part that cuts Brandon's way, and it is in the authors' own caveat. The equalization is in task execution, on a task somebody else specified, with the tool in hand. It does not survive the tool's removal, and it does not extend to tool use. Both of those are organizational variables: who specifies the task, and who is accountable for using the tool well. Read with Kanazawa et al. (2025, filed today), Brosch et al. (2026-10), Haslberger et al. (2025) and Otis et al. (2026), the five instruments sort on exactly those two conditions — see today's brief and knowledge/thesis-challenges.md (person-level moderator, update 2026-10-18). The authors also hand over a structural point worth keeping: existing within-firm evidence understates educational heterogeneity because hiring already selected on it.

Sources & Confidence

• Cruces, G., D. Fernández Meijide, S. Galiani, R. H. Gálvez and M. Lombardi, "Does Generative AI Narrow Education-Based Productivity Gaps? Evidence from a Randomized Experiment" — High (NBER working-paper series; authors at Nottingham, San Andrés, Tulane, UTDT).
• Version record (hard rule 14). Four versions located: NBER WP 34851 (2026-02), CEPR DP 21299 (2026-03), UTDT School of Government WP 2026-03, arXiv 2608.04198v1 (submitted 2026-08-04). The version read here is the arXiv v1 author copy, cross-checked against the UTDT listing; the 0.548 → 0.139 figures agree across both. The NBER landing page was not read, so a later revision cannot be ruled out — re-check before print use.
• Brosch et al. (2026), "Task Expansion with Generative AI," Stanford Digital Economy Lab working paper, 2026-10-07 copy, related-work paragraph §3 — High, read directly as the source of the misattribution.
• Haslberger, M., J. Gingrich and J. Bhatia, "No Great Equalizer: Experimental Evidence on Productivity Effects of Generative AI Use in the UK Labor Market," SSRN 4594466 — Moderate; the copy located is the 2023-10-06 draft via Birkbeck eprints, while Brosch et al. date it 2025. Direction of its finding verified, magnitudes not pinned.

Filed by / Routing

Filed by: Stafford (judgment layer, 2026-10-18) Routing: strategy — the individual-level premise to concede in print, stated at full strength; also the correction record for the 2026-10-18 misattribution. Practice asset: none — outcome instrument, not a prescriptive-design source.

Tags: paper-2, Technology Illusion, Capability, COUNTERARGUMENT Source: Management Science (INFORMS) 72(2), 1376–1388 URL: https://doi.org/10.1287/mnsc.2023.01631 Date published: 2025-06-09 Date flagged: 2026-10-18

Summary

Kanazawa, Kawaguchi, Shigeoka and Watanabe study a demand-prediction app deployed to drivers at a Japanese taxi firm. The app forecasts where ride demand will be and suggests where to cruise. The outcome measured is the time a driver spends searching for a passenger — the productivity margin that actually matters in the job.

The AI raises productivity by shortening cruising time, and the gain is not distributed evenly: it accrues only to low-skilled drivers, narrowing the productivity gap between high- and low-skilled drivers by 14%. High-skilled drivers gain little or nothing — their own accumulated knowledge of where demand will be already matches what the model produces. The identification turns on the fact that where a ride ends is close to random, which randomly determines where a driver begins searching and therefore whether the app's suggestion is available and used.

Two further results the authors report. Productivity gains for low-skilled drivers run around 7% in the published version. And adoption is incomplete in the direction that matters: many less-skilled drivers did not use the app even though it could have improved their performance substantially. Watanabe has described the tool publicly as potentially "deskilling" — the model substitutes for the spatial knowledge that previously distinguished a good driver from an average one.

Five Breakpoints Intersection

Technology Illusion — the tool worked and the deployment still under-delivered, for an organizational reason rather than a technical one: the authors report that many low-skilled drivers, the only group the app helped, declined to use it. The capability was real, available and free to the firm; the realized return was bounded by adoption inside the group with the most to gain, and nothing in the deployment addressed that. This is the cleanest example in the base of the gap between a technology's measured effect and a firm's realized effect being produced entirely by the organizational layer around it.

Four Forces Intersection

Capability — the app substitutes for driver skill rather than extending it. The gain exists only where unaided skill is lowest, and the authors raise deskilling explicitly: the knowledge that distinguished high- from low-skilled drivers is the knowledge the model supplies. This is a Capability result in both directions — capability acquired by the bottom of the distribution, capability potentially eroded in the accumulation channel that used to produce it.

Relevance to Brandon's Positioning

This is the instrument today's scan missed, and it is the strongest member of the equalization literature because of where it ran. Every other study in Brosch et al.'s related-work paragraph is an online or school-based experiment on assigned tasks; this one is a real firm, a real production tool, a real revenue outcome, and an identification strategy that does not rely on self-report. It also sits directly against the thesis in its blunt form: a 7% productivity gain for the bottom of the skill distribution with no redesign, no incentive change and no leadership involvement. Filed as COUNTERARGUMENT for that reason.

Its more useful role is as the second point on an axis. Taken with Cruces et al. (2026, filed today), Brosch et al. (2026-10), Haslberger et al. (2025) and Otis et al. (2026), the five instruments sort cleanly on two conditions that are organizational rather than individual: whether somebody else specifies the task, and whether the output can be checked objectively. The taxi app is the extreme case on both — the tool states the task ("cruise there") and the market verifies the answer within minutes — and it produces the cleanest equalization. Otis et al.'s entrepreneurs hold both conditions themselves and produce the only reversal. That sort is the named mechanism the amplification story thread has been asking for since 2026-09-17, and it is the same two variables Leyden & Scur model at the worker level (direction and verification). See today's brief and knowledge/thesis-challenges.md (person-level moderator, update 2026-10-18).

Against what is already in the base, this is stronger than the résumé-panel structural work on every dimension that matters for a causal claim and much narrower in scope — one firm, one occupation, one tool, one country. It cannot carry a claim about organizations in general. It can carry the claim that skill-substituting AI equalizes where the task is specified and checkable, which is the claim the axis needs.

Sources & Confidence

• Kanazawa, K., D. Kawaguchi, H. Shigeoka and Y. Watanabe (2026), "AI, Skill, and Productivity: The Case of Taxi Drivers," Management Science 72(2), 1376–1388; published online 2025-06-09, doi:10.1287/mnsc.2023.01631 — High (peer-reviewed, top-tier journal).
• Free prior version: IZA Discussion Paper 15677, https://docs.iza.org/dp15677.pdf; earlier as CIRJE-F-1202 (2022-10) — High.
• University of Tokyo press release (7% low-skilled gain; Watanabe's deskilling remark), https://www.u-tokyo.ac.jp/focus/en/press/z0508_00381.html — High for the figures it attributes to the published paper, Moderate for the framing.
• Version caution (hard rule 14). The headline numbers moved across versions: a 2021 preliminary draft reported a 5% average search-time reduction with roughly 10% for low-skilled and near zero for high-skilled; the published version carries the 14% gap-narrowing and 7% low-skilled figures. Cite the published Management Science version only. The 14% figure is also the one Brosch et al. cite, which is a second check on it.

Filed by / Routing

Filed by: Stafford (judgment layer, 2026-10-18) Routing: strategy — the equalization case stated at its strongest, which is what any client or reviewer will raise against the organizational-prerequisite argument. Practice asset: none — not a Category 8 prescriptive-design source; this is an outcome instrument.

Tags: paper-2, Technology Illusion, Capability, COUNTERARGUMENT Source: Harvard Business School Working Paper 26-090 (Hyunjin Kim, INSEAD; Rembrand Koning, HBS) URL: https://www.hbs.edu/ris/Publication%20Files/26-090_96f92aa0-37d9-4789-beaa-5c0cb87a4032.pdf Date published: 2026-06-09 Date flagged: 2026-10-16

Summary

Kim and Koning ask how firms "built around AI capabilities" are organized. They construct two samples: all Y Combinator startups in batches Winter 2020 through Fall 2024, collected via an open-source public API, and all U.S.-based PitchBook startups whose first financing deal closed between 2020 and 2024. AI-native status in the YC sample is taken from the startup's own public "AI" tag on its YC profile; a text model trained on YC descriptions extends the classification to the broader PitchBook universe. Both samples are linked to Revelio Labs workforce records — job titles, functions, seniority levels 1–7 — covering records through a January 2025 snapshot. Firms Revelio cannot link any employees to are imputed as zero-employee.

Relative to non-AI startups in the same industry-cohort cell, AI-tagged startups have approximately 25% fewer employees (roughly 13 Revelio-matched employees against 42). Their engineer share is 5 percentage points higher — a 13% greater share — offset by lower shares in sales, finance, operations and administration. They skew senior: the entry-level share is just under 15% lower and the senior share about 20% higher. Their hierarchies are half a seniority level flatter, with 15% fewer managers, controlling for firm size. Funding raised and valuations are comparable to non-AI peers, so the smaller teams raise roughly 20% more capital per employee. The patterns hold across both the YC and PitchBook samples.

The paper's second half decomposes the result into two channels. A product channel — AI embedded in what the firm sells — is measured by LLM classification of founders' own business descriptions. A process channel — AI tools used by the firm's own workers — is measured by whether any job posting names a specific tool from a 33-entry dictionary (ChatGPT, GitHub Copilot, Cursor, V0, Midjourney and others). In the 1,666-firm regression sample (467 AI-tagged, 1,199 non-AI), the product channel is present at 69.8% of AI-tagged firms against 16.1% of non-AI firms; the process channel at 13.5% against 5.2%, the roughly 2.6× relative gap. Regressing employment and hierarchy outcomes on both channels with batch × industry fixed effects, the product channel drives the organizational pattern and the process channel does not: product-channel firms have about 10 fewer employees, are 13% smaller in logs and 19% in Poisson, and operate with about 0.2 fewer hierarchy layers (all significant); the process channel's levels estimate is positive and its log, Poisson, hierarchy and manager-share estimates are not statistically significant. The authors state the implication directly: "equipping workers with ChatGPT, Copilot, or Cursor does not, on its own, predict smaller firms." Differences are largest in services industries where pre-AI delivery required staffing people — AI services startups are roughly 70% smaller with hierarchies nearly a full level flatter — illustrated by near-neighbour product-space pairs such as Educato AI (AI exam prep, 7 employees) against Careerist (third-party tutors, 912 employees).

Five Breakpoints Intersection

Technology Illusion: This is the first firm-level evidence in this base that separates deploying AI tools onto an organization from organizing around AI, and it finds the first does nothing structural — firms whose job postings name ChatGPT, Copilot or Cursor are not smaller, not flatter, and do not have fewer managers once the product channel is controlled for, while firms that embed AI in the product are all three. The authors' own sentence is the breakpoint stated as a measured result rather than an argument: "equipping workers with ChatGPT, Copilot, or Cursor does not, on its own, predict smaller firms."

No other breakpoint is mapped. The paper measures headcount, function, seniority and hierarchy depth; it does not measure purpose clarity, incentives, process friction or post-launch persistence, and tagging those from this instrument would be speculative.

Four Forces Intersection

Capability: The paper's central theoretical claim is that AI relocates productive capability out of the organization and into the product — "AI enables firms to shift some of this processing from the organization into the product itself, allowing the firm to deliver knowledge-work capabilities without building them internally through people and hierarchy." That is a Force-level claim about where capability resides, measured against headcount and hierarchy depth, and it is the first entry in this base where capability is located somewhere other than in people or practices.

No Purpose, Commitment or Momentum implication. None is measured. The authors explicitly flag Momentum's question as beyond their design — their data covers firms for at most a few years and they state that "whether these organizational forms persist as firms scale, or whether the demands of coordination reassert themselves, is something we cannot yet address." That is a stated absence, not evidence, and is recorded as such.

Relevance to Brandon's Positioning

This is the Claim 3 evidence the paper-2 case file has lacked — a direct comparison of AI-native organizations to conventional ones — and it arrives split in two directions. For the thesis: the process-channel null is the strongest firm-level support Technology Illusion has ever had, because it is the one analysis in the literature that holds the product constant and asks whether internal AI tool adoption alone changes structure. It does not. Any claim that an incumbent can reach the AI-native organizational form by giving its workforce AI tools is contradicted by the only dataset that decomposes the two channels. Against the thesis: the organizational change the market attributes to AI is, in this data, happening through entry rather than repair — AI-native firms are 25% smaller and half a level flatter without having run a transformation, fixed incentives, or passed any readiness assessment, because they were built that way. A framework addressed to incumbent repair is addressed to the population in which this result did not occur. Logged as a new OPEN challenge (the entry objection, knowledge/thesis-challenges.md 2026-10-16) and routed to asset 2, The Flattening Bet, whose delayering argument this both arms and constrains.

Binding usage note (CLAUDE.md hard rule 9). The workforce measures come from Revelio Labs — résumé/workforce inference, the same data genre as Babina et al. and Ewens & Giroud — on a U.S. VC-backed startup population with an overlapping window, read from a single January 2025 cross-section. This must never be cited as independent confirmation of Ewens & Giroud's delayering result; the two share a data genre and plausibly share firms. The authors' own limitations are binding too: AI status is self-identification, and "our cross-sectional design cannot pin down causal mechanisms." Composition is moving underneath the measure — the YC AI-tagged share runs from about 15% in early 2020 batches to over 50% in 2024 batches, and PitchBook AI first-deal counts in 2024 are nearly 8 times the 2020 average.

Sources & Confidence

• Kim H & Koning R, AI-Native Firms, Harvard Business School Working Paper 26-090, 9 June 2026. Read in full from the openly served HBS PDF. High (HBS working paper; unrefereed, and cross-sectional by the authors' own statement).
• Revelio Labs workforce data, January 2025 snapshot, as used by the authors. Moderate — résumé/workforce inference genre; see the usage note above.
• Kim H, Kim D & Koning R, Mapping AI into Production: A Field Experiment on Firm Performance (SSRN 6513481 / INSEAD WP 2026/20/STR), cited by this paper as the process-channel experiment. Moderate — abstract-grade in this base; see the pre-registration verification in briefs/2026-10-16.md.

Filed by / Routing

Filed by: Stafford (scheduled run, 2026-10-16) — found by citation-chasing the scan's own headline find, not by the discovery sweep. Routing: strategy (asset 2, The Flattening Bet), personal (the entry objection).

Tags: paper-2, Incentive Fragmentation, Technology Illusion, Capability, COUNTERARGUMENT Source: Daniel Rogger (World Bank Research Department / IFS) & Ravi Somani (UCL Department of Economics) — working paper URL: https://www.danrogger.com/files/papers/RoggerSomani_2019_Hierarchy%20and%20Information.pdf Date published: null Date flagged: 2026-10-15

Summary

The paper builds a measurement framework for the accuracy of what public officials believe about the jurisdictions they administer, and tests whether supplying them the data fixes it. The authors surveyed 1,831 public officials of the Government of Ethiopia between mid-June and mid-September 2016, eliciting beliefs about the variables their work depends on — the population they serve, primary-age children in school, antenatal care rates, hectares under agriculture — and scored each belief as absolute error against objective census and administrative benchmark data. The errors are large: only 21% of officials' population estimates fall within 20% of the census figure, while 47% claim their district's population is 50% bigger or smaller than it is. Mean absolute error is 37–38% of the benchmark across education variables, 35% on the antenatal care rate, and agriculture officials overestimate agricultural hectares "by almost a factor of 2." Among central officials, 47% cannot accurately state which of the 25th and 75th percentile districts in their region has the larger population.

The errors are consequential and the paper prices them. Controlling for the variables the budget formula distributes on plus district remoteness, ethnic fractionalization, population, rural share and poverty rate, officials' population estimates remain significant predictors of actual budget allocations (5% level clustered at organization, 10% clustered at district): "districts whose officials overestimate their population by 1-percent receive a 2-percent larger budget over and above the budget specified by the formula."

The authors' central descriptive result is that the determinant of accuracy is institutional, not positional. Demographic variables and proxies for officials' remoteness from their constituents explain little of the variation; "the institutional environment in which officials work explains the dominant portion of the variation," and the coefficient on their decentralization measure is 41 times larger in absolute terms than the coefficient on the local poverty rate. Management practices that reward information acquisition mediate decentralization's effect: "In those organizations in which management practices are least likely to reward information acquisition, there is no difference between the accuracy of claims at different tiers of government." The authors draw the conclusion directly — "unconditional decentralization, without effective corresponding management practices or more broadly incentives for information acquisition, does not improve the information environment." They also find no difference in accuracy between managers and employees, consistent with control rights being vested in management and thereby removing subordinates' incentive to acquire information independently. Their summary: superior information at decentralized levels is "a result of greater control rights and complementary incentives rather than proximity."

The experimental half is a randomized reduction in the cost of acquiring the information, and its headline estimate is covariate-adjusted rather than unconditional. A random sub-sample of organizations and district jurisdictions was sent the official administrative data for their region as a government circular, in the Ethiopian Public Service's own standard format, typically two weeks before the survey (registered as AEARCTR-0003946; treatment and control balanced on all but one characteristic). Column 1 of Table 4 — the unrestricted treatment effect — is "positive and statistically insignificantly different from zero at the usual levels" under both clustering approaches. With controls the sign reverses and the estimate strengthens: −, p = 0.08/0.06 with the paper's standard control set; −0.28 (p = 0.004/0.003) adding the de facto authority controls; −0.23 (p = 0.04/0.03) adding the management indices; −0.074 (p = 0.03) using proportional rank of error. The abstract reports the adjusted version: briefings "reduce these errors by a quarter of a standard deviation, but in line with theoretical predictions, organizational incentives mediate their effectiveness." The mediation is stark and is the authors' own emphasis: "the treatment effect arises totally from organizations with below-average management practices." Their reading — "While external interventions can reduce the scale of errors made by public officials, suitable organizational incentives ensure self-acquisition of that information" — and the policy corollary: "when the public sector is designed appropriately, the rationale for compensating interventions — which are often ad hoc and focused on short term goals - is weakened."

Five Breakpoints Intersection

Incentive Fragmentation: This is the entry's load-bearing mapping and it is measured against objective benchmarks rather than self-report — where management practices least reward information acquisition, the accuracy advantage of decentralized tiers disappears entirely, which is the authors' stated finding that "unconditional decentralization, without effective corresponding management practices or more broadly incentives for information acquisition, does not improve the information environment." A structural change intended to put information closer to the work delivers no information benefit when nobody is rewarded for acquiring it.

Technology Illusion: The evidence briefing is an information technology deployed on top of an incentive structure, and its measured effect "arises totally from organizations with below-average management practices" — with the authors concluding that appropriate organizational design weakens the rationale for such compensating interventions. Caveat binding on this tag: the unconditional intent-to-treat estimate is positive and insignificant, so this mapping rests on the covariate-adjusted specifications, not on a clean experimental contrast. Stated at that strength and no higher.

No Strategic Disconnection or Momentum Mirage mapping is claimed. The population-belief errors and the budget consequence are tempting as a dispersion measure against Breakpoint 1's self-check, and they are not one: officials are wrong about a fact with an external benchmark, which is not the same object as leaders disagreeing about purpose. Recorded so it is not later overstated.

Four Forces Intersection

Capability: Information accuracy is measured here as a property of the organization rather than of the individual or their location — demographic variables and remoteness proxies explain little, while "the institutional environment in which officials work explains the dominant portion of the variation," with the decentralization coefficient 41 times the local poverty rate coefficient in absolute terms.

No Purpose, Commitment or Momentum tags. None is measured in this instrument; management practices rewarding information acquisition are an incentive structure, not a measure of decision-makers' follow-through under pressure, and reading Commitment into them would be the loose tagging the base's rules prohibit.

Relevance to Brandon's Positioning

This bears on paper-2 Claim 2 — "AI removes the information routing problem but leaves the underlying misalignment intact" — and specifically on the clause the case file has no contemporary evidence for. Claim 2 is a conjunction, and the base's evidence is lopsided: the IBM/Censuswide instrument ingested by the scan on 2026-10-15 measures the second clause (a signal arrives and the organization does not respond), while the first clause — that a structural or technological change actually removes the routing problem — is evidenced only by analogy through the Previous-Shock Archive (Ash et al. 2007, Koppel et al. 2005). Rogger & Somani is the first instrument in the base that measures a routing improvement itself, against an external benchmark, and finds it conditional on incentives: decentralizing authority produces no information gain where management practices do not reward acquiring information. That is a stronger and more useful form of Claim 2's first clause than the claim as currently written, because it replaces "AI removes the routing problem" with a measured conditional — a structural change improves the information environment only where the incentive to acquire information exists — which is Brandon's own Breakpoint 2 doing the work, and which no AI-era instrument in this base has measured. It also supplies a quantified consequence the base has never had for an information failure: a 1% population overestimate buys a 2% budget premium over formula.

It carries a genuine counterargument and it must travel with the find. If a single administrative circular measurably improves officials' factual knowledge — and does so precisely in the badly-managed organizations — then at the information margin a cheap intervention substitutes for organizational repair, and the prescription "fix the organizational conditions first" is over-broad for that margin. The objection is real and is logged in knowledge/thesis-challenges.md (2026-10-15, OPEN). Two things limit it, and both are inside the paper: the unconditional ITT is positive and insignificant, so the substitution result is covariate-adjusted rather than cleanly randomized; and the experiment's outcome is belief accuracy only — no decision, allocation or service outcome is measured after treatment. The circular moved what officials knew. Whether anything downstream changed is unmeasured, which is Claim 2's proposition rather than a refutation of it.

Two disciplines for any use of this entry. It is Ethiopian public administration in 2016, not an AI result and not a firm: it enters on the Previous-Shock Archive's stated footing ("entries from the previous shock enter on the Bloom/Garicano/Sadun/Van Reenen footing, say so in the file, and are never cited as AI results"), and the analogy is load-bearing. And it is a third literature for that pattern — development economics and public administration, after economics of organization (Bloom et al.) and clinical informatics (Ash, Koppel) — which is the pattern's own pre-registered escalation condition; whether it escalates is a proposal for research-evidence, not a change made here.

Sources & Confidence

• Rogger, D. & Somani, R., "Hierarchy and Information," September 2019 — author's own copy at danrogger.com, read in full via text extraction. High on publisher tier (World Bank Research Department and UCL, peer-track research-institution work on the NBER/IZA footing; ethics approval from Oxford's Blavatnik School; preregistered at the AEA RCT Registry as AEARCTR-0003946). Version caveat, binding: this is a working paper dated September 2019 and no published journal version was located on a search run 2026-10-15. Per hard rule 14 the author's own copy was used rather than any repository mirror, and no landing-page revision date was available; a published version may carry different numbers and must be re-checked before this entry becomes load-bearing in a draft. Date published is recorded as null because the document carries month precision only (September 2019) and the base's hard rule 1 forbids manufacturing a day.
• Benchmark data: Ethiopian census and administrative/survey records, as collected and matched by the authors. High as used here — it is the instrument's independently verified substrate, which is what distinguishes this entry from the base's self-report genre.
• Treatment-effect figures (Table 4 columns 1–6) read directly from the paper's results section. High for fidelity of transcription; the unconditional column 1 null is reported here because the abstract does not report it.

Filed by / Routing

Filed by: Stafford (judgment layer, exploratory slot, 2026-10-15) Routing: strategy — the measured conditional is a better first clause for Claim 2 than the claim as written; governance — the budget-premium figure prices an information failure in public money; personal — the substitution counterargument needs Brandon's call, logged OPEN in thesis-challenges.md.

Tags: paper-2, Incentive Fragmentation, Momentum Mirage, Strategic Disconnection, Purpose, Commitment, ESCALATED Source: The Information (exclusive, internal memo obtained) — restated by BigGo Finance, The Decoder URL: https://www.theinformation.com/briefings/exclusive-meta-tells-engineers-ai-token-usage-part-performance-reviews Date published: 2026-09-02 Date flagged: 2026-10-12

Summary

On 2026-09-02 Meta sent all engineers updated performance-expectation guidance, jointly issued by Maher Saba and Santosh Janardhan, stating that the company will not use AI adoption dashboards or token usage to evaluate employee impact. Managers are directed instead to assess output quality, work speed, and the complexity and scope of the problems engineers take on. The guidance also removed references to specific AI tools and to AI-usage directives tied to particular tasks. In the same memo Meta states that 93% of its code changes are now completed with AI agent assistance, and that AI adoption "is not the goal in itself." The memo was obtained by The Information; the account here comes from that reporting as restated by BigGo Finance (2026-09-03) and The Decoder.

This reverses a policy Meta had put in place over the preceding ten months. In November 2025 Meta's Chief People Officer Janelle Gale told employees that "AI-driven impact" would become a core expectation from 2026; earlier in 2026 Meta mandated that engineers use AI tools and added token usage and AI-assisted programming to the hundreds of metrics in its performance management system. Reporting describes the result as "tokenmaxxing" — employees consuming tokens to earn better ratings. The Decoder (2026-06-13, also citing The Information) reports an internal "Claudeonomics" leaderboard that logged 73.7 trillion tokens in just over thirty days, a memo to roughly 6,000 employees flagging an "exponential increase" in AI usage with individual employees and teams having no visibility into or control over their own consumption, and internal AI costs that could reach billions of dollars in 2026. CTO Andrew Bosworth wrote in an internal memo (April 2026, reported June) that "No one should use AI tools just for the sake of using them," and that token usage alone does not measure impact. From 2027 Meta plans budgets and allocations governed by a central dashboard built internally and called "AI Gateway."

Two further items are recorded as unverified: reporting states that Amazon encountered a similar tokenmaxxing problem, and that a July complaint by former and current employees alleges token usage and other activity measures contributed to layoff selection. The latter describes allegations, not judicial findings.

Five Breakpoints Intersection

Incentive Fragmentation: Meta added token usage to its performance-management system as a measure of AI-driven impact, and engineers responded by consuming tokens — a leaderboard logging 73.7 trillion tokens in roughly a month — so the measure rewarded the consumption it counted rather than the work it stood for, which is the breakpoint's definition and is conceded in the company's own words when the retracting memo says adoption "is not the goal in itself."

Momentum Mirage: The retraction memo reports 93% of code changes AI-assisted in the same document that declares the adoption number unusable for judging impact — near-total measured adoption, simultaneously disavowed as evidence of progress by the organization reporting it; and the activity the metric sustained was large enough to be visible as a cost line in the billions before it was visible as a performance problem.

Strategic Disconnection: The operational definition of the AI objective that reached engineers was token consumption and dashboard adoption attached to evaluation, and the corrective memo replaces it with output quality, speed, and problem complexity and scope — the firm restating its own intended outcome after discovering that the version that reached the front line was a volume proxy.

Four Forces Intersection

Purpose: Bosworth's "No one should use AI tools just for the sake of using them" and the memo's "not the goal in itself" are both the organization stating, after the fact, that the objective it had operationalized was not the objective it held.

Commitment: Meta reversed a policy issued by its Chief People Officer within roughly ten months and in writing, which is follow-through in the sense of changing a stated direction when it proved wrong — recorded with the caveat that the trigger was visible gaming and cost, not an evaluation of the metric's validity.

Only the two Forces above are tagged. The remaining two are not implicated by anything this source measures, and are left untagged rather than inferred.

Relevance to Brandon's Positioning

This is the third independent member of the "AI Mandate Backfire" pattern in knowledge/emerging-patterns.md, which has sat at one source (Duolingo) since 2026-04-27, and it takes it to ESCALATED — three named organizations that attached a quantified AI-usage measure to individual evaluation, generated documented gaming, and then withdrew or reduced the measure while keeping the technology. It also resolves a standing unverified lead in the Paid for Deployment pattern, which recorded "HR Grapevine, 2025-11-17, reporting Meta tying employee reviews to 'AI-driven impact' from 2026; no primary located, treat as unsupported": the policy is now corroborated and attributed to a named officer (Janelle Gale, November 2025) and documented as retracted. Paid for Deployment explicitly asked whether "the same activity-not-outcome shape appears below the executive layer" — it did, at Meta, in the performance-management system, and it was withdrawn.

Against the other two members this is the strongest on gaming and the weakest on outcome: Meta supplies a quantified gaming figure and a named, dated document on both the imposition and the withdrawal, but no measured consequence — nothing here shows what the retraction did to output, where Gong et al. measure a 7% sales rise at MedTech. The three members are complementary rather than redundant: Duolingo is testimony, MedTech is measurement, Meta is documentation at scale.

For paper-2 this supports Claim 4 (AI produces the appearance of execution) with the pattern's only non-pseudonymous large-firm instance, and it is the material the Tier 2 asset (asset-suggestions.md #12) needs, because the sequence it shows — substituted objective → gamed measure → sustained activity read as progress → technology working throughout — is a breakpoint interaction, which is the contribution claim asset 12 now rests on after CFIR took the integration claim away.

Sources & Confidence

• The Information, "Exclusive: Meta Tells Engineers AI Token Usage Won't Be Part Of Performance Reviews," briefing, 2026-09-02 — Moderate as used here. The Information is a strong primary-adjacent outlet and obtained the memo, but the briefing is paywalled and was not read in this session; the memo itself has not been seen. Every figure below comes from secondary restatement.
• BigGo Finance (editorial team), "Meta Ends AI Usage Metrics in Engineer Performance Reviews, Curbing 'Tokenmaxxing' Trend," 2026-09-03, https://finance.biggo.com/news/441b5966-1eda-479d-bbaf-eed2816be6c6 — Moderate. Credits The Information; supplies the memo date, both signatories, the 93% figure and the "not the goal in itself" line.
• The Decoder, "Meta shifts from 'tokenmaxxing' to token managing as internal AI costs reportedly hit billions," 2026-06-13, https://the-decoder.com/meta-shifts-from-tokenmaxxing-to-token-managing-as-internal-ai-costs-reportedly-hit-billions/ — Moderate. Credits The Information; supplies the Claudeonomics leaderboard and 73.7-trillion-token figure, the ~6,000-employee memo, the Bosworth quotation, the 2027 AI Gateway plan and the billions-of-dollars cost estimate.
• Fortune, 2026-04-13, https://fortune.com/2026/04/13/duolingo-ceo-luis-von-ahn-ai-usage-requirement-employee-performance-evaluations/ — High (cited here only for the Duolingo comparison that makes this the pattern's third member).
• Unverified and labelled as such: the Amazon tokenmaxxing parallel (single secondary mention, no primary located) and the July employee complaint alleging token usage fed layoff selection (allegations only, no case document located).

Standing caution on this entry. It is a reported-document chain: a named, dated internal memo, two named signatories, consistent figures across two independent restatements published three months apart, both crediting the same outlet — which is materially stronger than the unattributed-headline layer this base routinely skips, and materially weaker than a read primary. Do not quote the 93% or the 73.7 trillion as measurements in print without reaching The Information's text or a second outlet that saw the memo. The pattern claim does not depend on either figure.

Filed by / Routing

Filed by: Stafford (judgment layer, exploratory slot 2026-10-12) Routing: strategy / governance

Tags: paper-2, Technology Illusion, Capability Source: International Journal of Environmental Research and Public Health (MDPI) — Chomutare, Tejedor, Svenning, Marco-Ruiz, Tayefi, Lind, Godtliebsen, Moen, Ismail, Makhlysheva & Ngo URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC9738234 Date published: 2022-12-06 Date flagged: 2026-10-11

Summary

Chomutare et al. screen 2,784 studies and include nineteen reports of AI systems actually implemented in healthcare settings (or regulator-approved, e.g. FDA/EMA), published 2015–2021, and code the reported barriers and facilitators onto the Consolidated Framework for Implementation Research (CFIR). Eligibility required human subjects, machine learning, and real implementation — not pilots or models.

The coded result is lopsided. "Most of the cited facilitators were related to engagement with and management of the implementation process" — management and engagement accounted for 27 of 57 facilitator codes. The leading barriers were the intervention's generalisability and interoperability with existing systems (7 of 36 barrier codes), data quality and management (6 of 36), and trust and transparency (5 of 36). The authors also record that no reviewed article reported a barrier related to management and engagement — the category that dominates the facilitator side is absent from the barrier side entirely.

The conclusion is a capability claim about the implementing organisations: "implementation science expertise should be a part of every AI implementation project in healthcare." Critically for how this source may be used, the review applied CFIR to the literature retrospectively — it notes that none of the nineteen included studies was reported as using CFIR itself. The framework is the reviewers' coding scheme, not the primary studies' instrument.

Five Breakpoints Intersection

• Technology Illusion: across nineteen real AI deployments in healthcare, the barriers that recur are interoperability with existing systems, data quality and clinician trust — properties of the receiving organisation and its existing infrastructure, not of the algorithm — which is direct evidence for the claim that technology deployed on top of unrepaired organisational conditions does not deliver.
• Process Friction: interoperability with existing systems is the single most-cited barrier category (7 of 36 codes), which is structural friction between the new system and established workflow blocking execution rather than a failure of the technology itself.
• Not evidence for Strategic Disconnection, Incentive Fragmentation or Momentum Mirage: the coding scheme has no construct for leader disagreement or sustainment, so an absence here is an absence in the instrument and must not be read as an absence in the world.

Four Forces Intersection

• Capability: the review's own conclusion — that implementation-science expertise should be part of every healthcare AI project — is a statement that the binding constraint is organisational execution capacity rather than model quality, argued from nineteen implemented cases.
• No Purpose, Commitment or Momentum tags. The coding scheme does not reach them, and the asymmetry noted below means the facilitator codes cannot be read as evidence about what organisations did.

Relevance to Brandon's Positioning

Two uses, and the second is a correction to a pattern opened this morning.

First, it is the best available evidence that the Technology Illusion mechanism is documented at implementation level in a field that studies implementation professionally — nineteen deployments, the failures located in interoperability, data and trust rather than in the model.

Second, and more valuable: it bounds the Borrowed Apparatus pattern that the 12:00 scan opened today at two members (research-evidence/knowledge/emerging-patterns.md). That pattern claims the AI-and-organisations literature applies none of the validated instruments sitting in adjacent pre-AI literatures. The naive reading of CFIR-and-AI work would kill it — CFIR has been pointed at clinical AI repeatedly (this review; imaging AI in BMJ Open 2024; ICU nursing in JMIR 2026; a 2026 JMIR Medical Informatics systematic review of AI clinical decision support). This source shows the kill is wrong and the pattern survives in sharper form: of the nineteen primary AI-implementation studies from 2015–2021, none used CFIR. The reviewers coded it on afterwards. Neglect in the primary literature, confirmed in the field most favourable to the instrument.

The honest boundary that follows: the pattern must be restated as a claim about the management and organisational AI literature, not about all AI research, because prospective CFIR use on clinical AI does appear from roughly 2024. That restatement is better for Brandon than the original — it makes management's non-use a disciplinary choice rather than a measurement impossibility, since another field demonstrably does it on AI.

One caution on the facilitator asymmetry, which should stop anyone citing the 27-of-57 figure as a finding about organisations. That management and engagement is the top facilitator and appears as a barrier in zero studies is far more likely an artifact of who writes implementation reports and what they choose to narrate than a fact about AI projects. Authors describing their own deployment credit their management of it; they do not list their own management as the obstacle. Cite the barrier distribution; treat the facilitator distribution as evidence about the genre.

Sources & Confidence

• Chomutare T, Tejedor M, Svenning TO, Marco-Ruiz L, Tayefi M, Lind K, Godtliebsen F, Moen A, Ismail L, Makhlysheva A, Ngo PD. Int J Environ Res Public Health 19(23):16359 (2022-12-06), doi:10.3390/ijerph192316359, PMCID PMC9738234, PMID 36498432 — High (peer-reviewed; read from open-access PMC full text this session).
• Prospective CFIR-on-clinical-AI studies named above (BMJ Open 2024 imaging AI; JMIR 2026 ICU nursing; JMIR Medical Informatics 2026 CDSS review) — Low as used here. Reached via search result summaries only, not fetched and not read this session. They are cited solely to bound the Borrowed Apparatus pattern's scope, and the bound must be re-verified against the primaries before it becomes load-bearing in any draft.

Filed by / Routing

Filed by: Stafford (judgment layer, exploratory slot follow-on, 2026-10-11) Routing: strategy — scope bound for the Borrowed Apparatus pattern (PROPOSED-UPDATE in the 2026-10-11 brief); Technology Illusion evidence for the paper-2 case file.

Tags: paper-2, Technology Illusion, Incentive Fragmentation, COUNTERARGUMENT Source: Implementation Science (BioMed Central) — Damschroder, Aron, Keith, Kirsh, Alexander & Lowery URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC2736161 Date published: 2009-08-07 Date flagged: 2026-10-11

Summary

Damschroder et al. build the Consolidated Framework for Implementation Research (CFIR) by synthesising nineteen published theories, models and frameworks of implementation (Greenhalgh et al.'s diffusion model plus eighteen others) into a single structure of five domains and 37 constructs: intervention characteristics (8), outer setting (4), inner setting (12), characteristics of the individuals involved (5), and process (8). The stated purpose is to replace a fragmented set of overlapping theories with one shared vocabulary of implementation determinants.

The framework is explicitly intended for diagnostic use. The authors write that researchers can select the constructs most relevant to their setting and "use these to guide diagnostic assessments of implementation context," and that "prior to implementation, capacity and needs assessments are done to identify potential barriers and facilitators." It is described as a practical framework for systematically assessing barriers and facilitators rather than as a theory.

The authors are equally explicit about what the framework does not do. It presents constructs "without depicting interrelationships, specific ecological levels, or specific hypotheses," and "does not specify the interactions between those constructs." A 2022 updated CFIR and a CFIR Outcomes Addendum (separating implementation outcomes from innovation outcomes) have since been published. Named constructs include "organizational incentives and rewards" and "external policies and incentives", "goals and feedback", "leadership engagement", "relative advantage", "complexity", "planning", and "reflecting and evaluating". There is no construct for structural process friction (the nearest is "planning") and no construct for sustainment or loss of momentum ("maintenance" appears only in a reference to RE-AIM, not as a CFIR construct).

Five Breakpoints Intersection

• Technology Illusion: CFIR's organising premise is the Five Breakpoints claim about technology in generalised form — the same intervention succeeds in one setting and fails in another because of inner setting conditions rather than the intervention's own properties, which is why the framework devotes 12 of its 37 constructs to the receiving organisation and only 8 to the intervention itself.
• Incentive Fragmentation: the inner-setting construct "organizational incentives and rewards" names the mechanism directly, and the outer-setting construct "external policies and incentives" extends it beyond the firm boundary — so this breakpoint has a named, dated construct inside a validated multi-domain diagnostic framework, not merely a 1975 antecedent.
• Logged as COUNTERARGUMENT because of what it does to the originality claim: CFIR is an integration of nineteen theories with a diagnostic use, which is precisely the winning condition the relabeling objection (knowledge/thesis-challenges.md, CONTESTED, 2026-09-17) set for itself. Two of that objection's three limiting responses — "nobody integrated them" and "none of them diagnoses" — are false against this source.

Four Forces Intersection

• Capability: the readiness for implementation sub-domain (including "available resources," "access to knowledge and information" and "leadership engagement") is an operationalisation of organisational capacity to execute a change, assessed before deployment rather than inferred after failure.
• Purpose: the inner-setting construct "goals and feedback" sits under implementation climate and concerns whether goals are communicated and progress fed back — adjacent to, but narrower than, Strategic Disconnection's question of whether leaders agree on the outcome. Tagged on the construct's existence, not on a claim about leader agreement, which CFIR does not measure.
• Momentum: not tagged. CFIR has no sustainment construct; "maintenance" appears only as a reference to RE-AIM. The absence is the finding and must not be filled with a guessed tag.

Relevance to Brandon's Positioning

This is the most consequential prior-art find in the base and it was named as the risk twenty-four days before it was opened. knowledge/asset-suggestions.md asset 12 ("What the Five Breakpoints Owe") listed CFIR as "the most dangerous of the five" candidates for the unrun integration search, and said plainly: "If it comes back with a hit Brandon hasn't cited, this is not a section, it is a revision." It came back with a hit. Asset 12's three stated novelty claims were: nobody integrated these, nobody diagnosed, and none has an account of speed. The first two do not survive this source.

What survives is narrower, better sourced, and in the prior art's own voice — which makes it the strongest available spine for the Tier 2 differentiation essay. CFIR's authors explicitly decline to specify how their constructs interact or to advance hypotheses. Five Breakpoints is a claim about the interaction and the sequence — that treating one condition while ignoring the others is why transformations fail. Brandon no longer has to assert that his contribution is the interaction claim; the most-cited determinant framework in implementation science declines, in writing, in 2009, to make that claim about itself. Add to that two conditions CFIR has no construct for at all — structural process friction and loss of momentum, i.e. Process Friction and Momentum Mirage — and the machine-speed extension CFIR could not have made. That is a defensible contribution claim stated precisely, replacing two claims a reader with implementation-science training can refute from the abstract.

The uncomfortable part, which belongs in the record. CFIR was not an undiscovered source. It is cited twice inside this base already — knowledge-base/bmc-hsr-miake-lye-readiness-assessments-review-2020-02.md maps 1,370 readiness-assessment items onto it, and knowledge-base/implementation-science-helfrich-orca-predictive-validity-2011-08.md records Helfrich's 2021 mapping of ORCA to it. The base has used CFIR as a coordinate system since 2026-08-30 without once opening it as prior art for the framework it coordinates.

Sources & Confidence

• Damschroder LJ, Aron DC, Keith RE, Kirsh SR, Alexander JA, Lowery JC. Implementation Science 4:50 (2009-08-07), doi:10.1186/1748-5908-4-50, PMCID PMC2736161 — High (peer-reviewed; read from the open-access PMC full text this session; all quoted phrases verbatim from it).
• Read bound, stated honestly: the PMC full text is ~134,000 characters and the first ~100,000 were read. The five domains, construct counts, the nineteen synthesised theories, the interrelationships disclaimer, the diagnostic-use statements and the named constructs above are all from the read portion. No claim here rests on the unread final 34,000 characters, and the 2022 updated CFIR's construct list has not been checked — construct-level claims are about the 2009 original.

Filed by / Routing

Filed by: Stafford (judgment layer, exploratory slot, 2026-10-11) Routing: strategy — rewrites asset 12 and amends knowledge/thesis-challenges.md The relabeling objection; the Tier 2 gate in knowledge/master-plan.md cannot be drafted on its current argument.

Tags: paper-2, practitioner, Strategic Disconnection, Momentum Mirage, Capability, Momentum, COUNTERARGUMENT Source: Fast Company (Leadership) URL: https://www.fastcompany.com/91616068/change-management-counterintuitive-strategies-that-actually-work Date published: 2026-10-07 Date flagged: 2026-10-10

Summary

Greg Satell — co-founder of the change advisory ChangeOS, a Wharton lecturer, and author of Cascades and Mapping Innovation — argues that the standing practice of change management is built on an unevidenced foundation. He locates the foundation precisely: Kotter's 1995 Harvard Business Review article, "based on his observations of more than 100 companies," from which the eight steps and their descendants (he names ADKAR explicitly) derive. His charge is methodological rather than substantive: "the only evidence he cites is his own observations of 100 companies. From his subsequent writing, these appear to be case studies based on executive interviews. There doesn't appear to be any significant effort to control for his own biases, those of his subjects, or any other significant sources of error." He contrasts this with Doug McAdam's work on Freedom Summer recruitment, where the researcher analysed applications from 720 volunteers plus 239 who withdrew and 55 who were rejected, conducted 80 in-depth interviews, and — because applicants listed their social contacts — could document the social ties among them. His generalisation: "That type of documentation simply doesn't exist in Kotter's work or the work of his imitators. To be fair, it is rare in business research as a whole."

Against that, he offers three prescriptions drawn from social-movement and diffusion research, crediting Everett Rogers' five decades of documentation. (1) Start with a majority rather than Kotter's "sense of urgency": he cites the finding that people presented with disconfirming facts "tend to question the evidence rather than our own preconceived notions," the KAP (knowledge-attitude-practice) gap in which knowledge and attitude shifts do not produce practice change, and the social rather than mass-media propagation of change (1950s air-conditioner diffusion, Freedom Summer recruitment). The operative instruction is to shape networks rather than opinions and to secure a local majority first — "even if that majority is just three people in a room of five" — because "once you're in the minority, you will feel immediate pushback." (2) Wait for a trigger, on the grounds that change "isn't always possible" and that pushing when it is unlikely to succeed "is more likely to put you deeper in the hole"; the discipline is to build the movement before the trigger, because "one reason why so many change efforts have failed in recent years is that the movements were built in response to a moment. Once the moment comes, it's too late." His examples are the Eastern European colour revolutions targeting elections likely to be falsified, King's move from Albany to Birmingham because Bull Connor was "likely to employ the type of brutal tactics that would trigger a moment," and COVID-19 triggering digital transformation. (3) Design a dilemma rather than force a confrontation, illustrated by a single consulting engagement of his own: a sales director who "never overtly objected, but would nod her head and then quietly sabotage progress" was presented with a client-handover meeting she had already agreed to in principle, had to either permit it or derail it, chose to derail it, "and was fired for cause." He traces the tactic to Alice Paul's Silent Sentinels picketing Wilson's White House with his own quotes in 1917, notes "increasing academic interest," and states the design principle: the action must be "seen as a constructive act rooted in a shared value."

His closing claim is institutional as well as evidentiary: "the traditional change models simply don't work. They aren't based on any serious research and have shown themselves, over a period of decades, to fail consistently. Often, organizational change management units are used by consulting firms and vendors to cheerlead larger engagements." He explicitly declines to call this a con — practitioners "are well-meaning and under the impression that they make a positive impact" — and identifies the structural reason the impression survives: "They usually aren't around to see what happens after the engagement ends."

Five Breakpoints Intersection

Momentum Mirage: Satell supplies a named mechanism for the appearance of progress without movement, and it is one the base did not hold — "the movements were built in response to a moment. Once the moment comes, it's too late." The claim is that mobilising at a trigger generates visible activity on a network that was never built, so the energy is real and the capacity to convert it is absent; the prescription (build before the trigger) is the mechanism stated as a design rule.

Strategic Disconnection: the KAP gap as he states it — "shifts in knowledge and attitudes do not necessarily lead to changes in practice" — is a named reason why a direction that has been communicated, understood and even agreed to still fails to reach operational reality. That is the illusion of alignment with a measurement tradition behind it rather than an assertion, and the base has no entry on the KAP literature.

Not tagged, deliberately — the incentive-fragmentation, process-friction and technology-illusion mappings were all considered and rejected. The sales-director anecdote reads like the first of those (a senior employee protecting her book of business against a restructuring that would redistribute it), but the source supplies no reward structure, no measurement and no counterfactual, so the tag would rest on my inference rather than on his claim. AI is not mentioned anywhere in the piece. (Names hyphenated above on purpose: the --from-kb converter's name scan reads loosely across the whole file, so writing a declined tag in its display form registers it as an untagged breakpoint and fails the entry. See the pipeline note in today's reply.)

Four Forces Intersection

Momentum: the trigger argument is a direct, falsifiable claim about the timing of sustained energy — that movements built in response to a moment fail, and movements built in advance of one succeed. It is the Force stated as a sequencing rule, and it is testable in principle even though he tests it nowhere here.

Capability: the KAP gap locates change failure in the capacity to convert intent into practice rather than in intent itself, which is the Capability claim stated from the other side. Note this cuts against the piece's own first prescription: if knowledge and attitude shifts do not produce practice change, a secured majority of enthusiasts is a shift in attitude and inherits the same gap. He does not address that.

Not tagged: Purpose and Commitment. The source is fully readable, so these are honest judgements rather than access nulls — he makes no claim about clarity of intent or about decision-makers holding direction under pressure.

Relevance to Brandon's Positioning

This is the hostile reader from the 2026-09-17 sparring note, published in a national outlet three weeks later, and the single most useful thing in it is that he attacks Kotter on a different axis than the one Brandon has been defending. The relabeling objection (thesis-challenges.md, CONTESTED, logged 2026-09-17) says the breakpoints are renamed canon — Momentum Mirage is Kotter's seventh error, Strategic Disconnection his third and fourth. Satell says the canon was never evidenced: eight steps from one author's uncontrolled observation of 100 companies, via executive interviews, with no bias controls. Those two objections are close to opposites, and the second partly dissolves the first: a priority claim over Kotter is worth much less if Kotter's own claim was never established, which makes "you are relabeling Kotter" a weaker charge than it looked on 09-17. That is a real gain for Paper 2's contribution claim and Brandon should take it.

The cost is larger than the gain and it needs stating first. Satell's standard indicts Five Breakpoints exactly as it indicts Kotter. The framework is five named failure modes drawn from a practitioner's own case experience, and the base's own ledger concedes the missing piece repeatedly — asset 3 is titled "The Missing Instrument," growth edge #2 has asked since 2026-08-29 for an ex-ante Five Breakpoints measure with an outcome criterion and has never been answered, and the recurring complaint across the structural literature in this base is that nobody has an outcome variable. Brandon has no McAdam-grade instrument and neither does anyone else in this field; the honest position is to say so before a reviewer does, which is the same unforced-concession move Paper 3 already made once on the 70% figure. See the new challenge logged in thesis-challenges.md today.

Two defects in the piece are worth holding, because they are the answer to it if it is ever cited at Brandon. First, its premise is the claim Brandon has already publicly retired: "Study after study finds that the vast majority of organizational transformations fail" is asserted with no study, no figure and no citation, in an article whose entire argument is that change management rests on unsourced assertion. Paper 3's August 2026 update retires exactly that claim on Hughes (2011) and traces its provenance to Hammer & Champy (1993) — and Hammer & Champy is also the process-friction antecedent in the relabeling challenge, so Satell's unevidenced premise and Brandon's unattributed lineage run back to the same book. Second, and more damaging, the piece indicts Kotter for uncontrolled case studies from executive interviews and then evidences its third prescription with a single self-narrated consulting engagement in which the author is the protagonist, the subject was fired, and the outcome ("sales shot up") is self-reported with no counterfactual. That is n=1 with maximal researcher interest, offered three paragraphs before the methodological indictment. Third, less a defect than a boundary: all three prescriptions are imported from voluntary-affiliation settings — diffusion, civil rights recruitment, colour revolutions, suffrage picketing — into involuntary-affiliation settings with hierarchy and payroll. "You are under no obligation to invite the bomb throwers in" is precisely what an org chart forbids, and the transfer is the same analogy error this base catches in other people's work (the Previous-Shock Archive pattern; the ERP-as-AI caution on Bloom et al.).

Positioning, and explicitly not a window. This opens no publication window — there is no dated peg, and per hard rule 5 an argument piece is argumentative value, not urgency. What it is: the first evidence in this base that a credentialed competitor (Wharton lecturer, trade book, 50+ HBR articles, his own change advisory) is making the "change management's evidence base is broken" argument publicly. That is Brandon's own register, and the asset-12 Antecedents piece now has a live interlocutor rather than a hypothetical reviewer.

Sources & Confidence

• Satell, G., "The change management playbook is broken. Here are 3 counterintuitive strategies that actually work," Fast Company, 2026-10-07 — Moderate. Trade press, not on the high-credibility list; the author carries real standing (Wharton, HBR volume) but the piece contains no data of its own and its one organisational case is self-reported n=1. Full text read, supplied by Brandon 2026-10-10; the URL returned HTTP 403 (DataDome) from this container on every route attempted, so the text is Brandon's paste rather than a first-party fetch.
• Secondary claims the piece makes that this base has not verified and must not inherit: Kotter's 1995 article as described; McAdam's Freedom Summer counts (720 / 239 / 55 / 80 interviews); Rogers' five decades of documentation; 1950s air-conditioner diffusion; the KAP gap literature; Alice Paul's Silent Sentinels, 1917; the "increasing academic interest" in dilemma actions. Each is a checkable primary source and none is in the base — mcadam, freedom summer, everett rogers, KAP gap, silent sentinel and dilemma action all return zero files across both repositories. Low confidence until chased, which is the entry's main forward value.
• Brandon's own corpus, used above and verified in-repo: Paper 3 (August 2026 update) retiring the 70% figure on Hughes (2011) with the Hammer & Champy provenance endnote; knowledge/thesis-challenges.md relabeling objection, CONTESTED, logged 2026-09-17 — High (primary, in-repo).

Filed by / Routing

Filed by: Stafford (targeted ingest at Brandon's request, 2026-10-10; full text supplied by Brandon after the URL proved unfetchable) Routing: market (a credentialed competitor is now making Brandon's argument in public), personal (the strongest available material for the unanswered 2026-09-17 sparring note on growth edge #1)

Tags: paper-2, Technology Illusion, Commitment, Capability, COUNTERARGUMENT Source: Implementation Science (BioMed Central / Springer Nature), 19:29 URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC10976812/ Date published: 2024-03-28 Date flagged: 2026-10-09

Instrument (recorded BEFORE the finding, per CLAUDE.md hard rule 9)

Data source — and it is split, which is the whole value of this entry. The outcome (measurement-based-care fidelity) comes from system-generated electronic metadata from the OQ-A platform combined with monthly caregiver reports — the authors' words: "The index was generated using electronic metadata from the OQ-A system combined with monthly caregiver reports." Neither source is the clinician whose behaviour is being scored. The mediators (implementation leadership, transformational leadership, implementation climate) are all clinician-reported questionnaires: the 12-item Implementation Leadership Scale and the 20-item Multifactor Leadership Questionnaire, both rated by clinicians about their first-level leaders, and the 18-item Implementation Climate Scale, rated by clinicians about their own clinic. Clinician ratings were aggregated to clinic level after r_wg(j) agreement checks.

So the causal chain has one contaminated link and one clean one. Leadership → climate is same-respondent: the same clinicians supplied both the predictor and the outcome of that step. Climate → fidelity is not: clinician-reported climate against a platform log plus caregiver reports. This matters for how the two headline numbers may be used and is developed below.

Inference method. Secondary mechanistic (causal mediation) analysis within the WISDOM cluster randomized trial, NCT04096274 — the same trial as the uptake paper filed here on 2026-10-08. 21 outpatient youth mental health clinics randomized (11 LOCI, 10 control); both arms received identical measurement-based-care training and technical assistance, the LOCI arm additionally received 12 months of the leadership-and-climate intervention. Multilevel causal mediation (Imai et al.), R mediation package. Mediation models ran on 20 clinics for the climate step and 17 clinics for the fidelity step, the reduction attributed to missing data.

Observation window. Leadership, transformational leadership and climate measured at baseline, 4, 8, 12 and 18 months. Fidelity measured over six months for youth entering treatment in the 12 months after clinician training. 252 clinicians completed assessments (analytic Ns 247–251); 234 youth enrolled, 231 analyzed for fidelity. Clinics enrolled 2019; measurement runs 2020–2022. Pre-generative with respect to AI.

Setting. Small outpatient youth mental health clinics in Idaho, Oregon and Nevada. Not enterprises, not AI.

Summary

This is the mechanism paper from the trial whose uptake results this base filed on 2026-10-08, and it answers the question that paper deferred: through what did the leadership intervention work? The authors specified a chain in advance — LOCI improves leadership, improved leadership produces a better implementation climate, and a better climate produces higher fidelity in the clinical practice being implemented — and tested each link by causal mediation.

The chain held, and it discriminated between two leadership constructs. Improvement in implementation leadership (leader behaviours specific to the practice being implemented — being proactive in removing obstacles before they land, knowledgeable about it, supportive of staff doing it, and perseverant through its difficulties — the four Implementation Leadership Scale subscales) from baseline to 4 months mediated LOCI's effect on clinic implementation climate at 12 months: indirect effect 0.51, proportion mediated 0.82 (95% CI 0.25–1.00), p = .004. Improvement in climate then mediated LOCI's effect on fidelity over the same period: indirect effect 0.14 (95% CI 0.01–0.37), proportion mediated 0.71 (95% CI 0.02–1.00), p = .045. Transformational leadership — the generic charismatic-and-visionary construct that most leadership literature measures — did not mediate anything. LOCI moved it slightly at 4 months (d = 0.31, p = .019) and not at all thereafter (all ps > .05), and its mediation of climate was null (indirect effect 0.16, p = .135; proportion mediated 0.29, p = .136). LOCI's effects on implementation leadership and climate themselves were large and durable at every follow-up (ds 0.76–1.34, all ps < .01).

The authors' own limits. The fidelity index is the product of a completion rate and a viewing rate and, in their words, "does not indicate whether clinicians used the feedback to guide treatment" — the same value-wall caveat the uptake paper carries. Training and technical assistance were in both arms, so LOCI's components cannot be separated from each other. The stated primary limitation is generalizability, since all clinics and caregivers volunteered. The paper reports a priori power of 0.74–0.90 and does not list small sample size among its limitations, which is a judgment this base need not share given the fidelity model ran on 17 clinics. The paper does not discuss common-source bias at all; it frames clinician raters as a strength, as "third-party informants whose behavior is most salient."

Five Breakpoints Intersection

Technology Illusion: The intervention held the technology and the training identical across both arms and moved only leadership and climate, and the thing that improved was how faithfully the deployed platform's core practice was actually carried out — completion of the measure and clinician viewing of the feedback, read from the platform's own metadata. That the same software produces materially different practice depending on an organizational condition that was assigned rather than observed is the breakpoint's causal claim tested at the strongest available identification.

Momentum Mirage — considered and deliberately not tagged. The uptake paper's post-intervention decline is the Momentum evidence from this trial; this paper's analysis window ends at 18 months and it reports no sustainment result, so tagging it here would double-count one trial's single finding.

Correction 2026-10-21 (Stafford) — the Implementation Leadership Scale has FOUR subscales, not three

This entry originally described implementation leadership with three behaviours (knowledgeable, supportive, perseverant). The instrument has four, verified 2026-10-21 against the ILS-X validation paper (PMC12676754), which states the original dimensions explicitly: proactive, knowledgeable, supportive, perseverant. Proactive — the leader who removes obstacles before they land — was missing, and it is the dimension closest to this framework's own remedy language, so the omission mattered. Corrected in both occurrences above.

Recorded at the same time, because it was the other half of the same check: the Implementation Climate Scale contains incentives, and this entry never said so. The 18-item ICS has six three-item subscales — Focus on EBP, Educational support for EBP, Recognition for EBP, Rewards for EBP, Selection for EBP, Selection for openness — verified against the ICS German validation (PMC10066389, Table 3). One Rewards item reads "This team/agency provides financial incentives for the use of evidence-based practices." Implementation climate is defined as implementation being "expected, rewarded, and supported by the organization."

So LOCI does not ignore Incentive Fragmentation — it measures it as a perception. What it never establishes is that LOCI changed any actual reward, or who holds the authority to set one: the leadership construct is rated by subordinates about their first-level supervisor, a layer that does not set financial incentives. See the altitude argument in knowledge/thesis-challenges.md.

Four Forces Intersection

Commitment: The mediator that carried the effect was implementation leadership — measured as whether a leader is proactive in clearing obstacles for the specific practice, knowledgeable about it, supportive of staff doing it, and perseverant through its difficulties — and it carried roughly four-fifths of the intervention's effect on climate while the generic transformational construct carried none that reached significance. Commitment as this framework defines it (acting consistently with a stated direction when it costs something) is the specific construct, not the inspirational one, and this is the first result in the base to put the two in the same model and separate them.

Capability: Evidenced as a null again, and for the same structural reason as the uptake paper: identical training and identical technical assistance in both arms means skills-and-tooling capability was held constant by design and cannot account for the fidelity difference.

Relevance to Brandon's Positioning

This is the first randomized, mediation-tested evidence that specificity in the organizational remedy is what does the work — and it arrives inside the prior-art objection logged against the framework yesterday. The LOCI exposure logged on 2026-10-08 says: the cure already exists, it is manualized, and it has a registered trial, so why buy a five-way diagnostic? This paper sharpens that objection (the cure now has a mechanism analysis and the mechanism is understood) and simultaneously undercuts its most important joint: LOCI's own analysis finds the general leadership construct inert and the implementation-specific one load-bearing. The principle that generic organizational quality is the wrong level of description, and that you must name the specific thing being implemented, is now LOCI's finding as much as Brandon's premise.

The bound has to travel with it, because it is not the five-way test. LOCI's specificity is specificity to the practice being implemented — one axis, two constructs compared. Brandon's claim is that five failure modes are distinguishable and that knowing which is live changes the remedy. That remains untested, and this paper does not test it. What it establishes is the weaker, prior claim the framework needs: that the level of description matters at all, demonstrated where the organizational variable was assigned rather than observed.

For the paper-2 case file, it moves Claim 5's request (b) one rung and does not close it. The 2026-10-08 split recorded request (a) — does a non-technological organizational variable move technology adoption, identified against selection — as met, and request (b) — does it move converted value — as open. Fidelity is not converted value; it is a composite of whether the measure was completed and whether the feedback was opened, and the authors say plainly it does not show the feedback changed any clinical decision. But it is the first thing in the base on the quality-of-delivery rung that this morning's scan named as exactly what Claim 5 needs and the AI literature has never operationalized — found independently, from the other direction, on the same day.

One correction to this base's own record, in Brandon's favour. The 2026-10-08 entry on the uptake paper flagged its lay summary's claim that LOCI clinics "followed best practices more closely" as not supported by the outcomes analyzed there and said it must not be repeated. That flag was right about that paper and wrong about the trial: the support exists, in this paper, as the fidelity result. The claim is quotable — cited to Implementation Science 19:29, not to the uptake paper.

And the base's own standard has to be applied against the find, not for it. The 2026-10-05 challenge accuses this base of applying a one-questionnaire standard to vendors and waiving it for peer review. Applied here, it demotes the paper's best-looking number: the tight, strongly significant estimate (leadership → climate, proportion mediated 0.82, p = .004) is the same-respondent link, and the link that reaches a non-self-reported outcome (climate → fidelity) is the weak one — proportion mediated 0.71 with a 95% confidence interval running from 0.02 to 1.00, p = .045, on 17 clinics. In plain terms: the part of the story measured by asking the same people twice is crisp, and the part that reaches an independently recorded outcome is barely distinguishable from noise. Both must be reported together or neither.

Sources & Confidence

• Williams NJ, Ehrhart MG, Aarons GA, et al. "Improving measurement-based care implementation in youth mental health through organizational leadership and climate: a mechanistic analysis within a randomized trial." Implementation Science 2024;19:29. DOI 10.1186/s13012-024-01356-w. Published 2024-03-28 (received 2023-11-16, accepted 2024-02-27). Trial registration NCT04096274. Open access, full text read. — Moderate. The design is strong where it is strong (randomized parent trial, a non-self-reported outcome, pre-specified mediation chain, multilevel causal mediation) and the publisher is peer-reviewed but outside the named high-credibility list. Held at moderate rather than high on three grounds: the fidelity mediation model runs on 17 clinics with a proportion-mediated confidence interval spanning nearly the whole range; the mediator-to-mediator link is same-respondent and the paper does not acknowledge it; and all clinics volunteered.
• Companion uptake paper already in the base: Williams NJ, Aarons GA, Ehrhart MG, Vega N, Marcus SC, Translational Behavioral Medicine 16(1):ibag007, 2026-03-13 — filed at knowledge-base/williams-aarons-ehrhart-loci-leadership-randomized-digital-uptake-2026-03.md. Same trial, same clinics, same research group: these two papers are one instrument and must never be cited as two independent confirmations.
• Noted, unread, logged as chase targets rather than evidence: a second LOCI cluster randomized trial in substance use treatment clinics (Journal of Substance Use & Addiction Treatment, 2024, S2949-8759(24)00149-8) — a different sector and different clinics but the same intervention and research group; and a normalization-process-theory paper from the WISDOM trial on sustainment over 16 months (Implementation Science, 10.1186/s13012-026-01485-4). — Low as held here, because neither has been read.

Filed by / Routing

Filed by: Stafford (judgment layer, exploratory slot, 2026-10-09) Routing: strategy — the specificity result belongs in the paper-2 counterarguments section as the answer to the LOCI prior-art exposure; personal — the fidelity index is the worked template for asset 3b's outcome definition.

Tags: paper-2, Commitment, Momentum, COUNTERARGUMENT Source: Substance Use & Misuse (Taylor & Francis) — Stein, Bassett, Welsh, Clair-Michaud, Abdel-Salam, Monico, Gallagher, Murgo, Yang, Friedmann & Clarke; CJ-DATS-2 Assessment Work Group URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC9969992/ Date published: 2023-01-11 Date flagged: 2026-10-08

Instrument — recorded before the finding (mandatory for structural instruments)

Data source. Mixed. The outcome is rated from documents: site-level success in meeting process-improvement goals and objectives, scored by trained researchers from each site's End of Implementation Report on a 0–6 scale (a goal counted as met at 4–6). The fidelity constructs blend administrative records (fidelity checklists completed, weeks to complete phases and meetings, facilitator logs, meeting notes, Champion and team-member stability, number of team meetings held) with self-report from the local change teams. Leadership was measured with a 9-item scale (Broome et al. 2009) completed by N = 1,052 staff plus supervisors.

Inference method. Mixed-method cluster randomized design: 21 sites (Local Change Teams) nested within 9 research centers, randomized to early-start (10 sites) or delayed-start (11 sites), with each center contributing at least two sites. Condition effects by ANCOVA with leadership controlled; fidelity-to-outcome associations by correlation; and moderated regression testing leadership as a moderator of each fidelity construct. Four fidelity constructs after Dane & Schneider (1998): responsiveness (5 indicators, α = .68), adherence (6 indicators, α = .76), quality (4 indicators, α = .69) and dose (1 indicator).

Observation window. Data collected 2010–2013. Phases: pre-phase 1–2 months, needs assessment 3–4 months, process-improvement planning 3–4 months, implementation 6 months, follow-up/sustainability 3 months. Early and delayed sites started on average 14 months apart.

Design limit that bounds every number below, stated first. The authors describe the work as exploratory, applied no alpha correction, and note that with 21 sites the nesting of staff within teams and within research centers could not be modeled and factor analysis was not possible. Sites were not randomly selected from any broader population.

Summary

The authors asked whether how faithfully an organization executes an implementation effort predicts whether that organization meets its own process-improvement goals, and whether leadership changes the answer. The setting is CJ-DATS-2, a federal trial of an Organizational Process Improvement Intervention across correctional and community-treatment sites.

The headline result is mostly nothing. Randomized condition did not affect any fidelity construct (responsiveness F(1,18) = .026, p = .873; dose F = .82, p = .378; quality F = 3.73, p = .069; adherence F = 3.08, p = .098 — the last two large in effect-size terms, d = .85 and d = .93, but not significant at 21 sites). Associations among the fidelity constructs ran at a median correlation of .19, and associations between fidelity and the goal-attainment outcomes at a median of .23, with most p-values non-significant. Fidelity scores also declined significantly over time.

One interaction survived, and it points the opposite way to intuition. Leadership did not moderate dose, quality or adherence. It did moderate responsiveness — staff engagement and dedication to the effort — for percent of goals met (b = .15, p = .045) and percent of objectives met (b = .15, p = .03), though not for the mean goal or mean objective ratings (p = .09 and .06). Decomposed, the effect lives entirely at the bottom of the leadership distribution: at low leadership, higher staff responsiveness predicted worse goal attainment (percent goals b = −.19, p = .03; percent objectives b = −.20, p = .01), with no significant relationship at moderate or high leadership. The authors state it directly in the abstract — "At low levels of leadership, high staff responsiveness (i.e., engagement in the OPII) related to poorer outcome" — and in the conclusion: "Even when staff members are more responsive in attempting to accomplish organizational goals, it appears that this does not compensate for impact of relatively weak leadership."

Five Breakpoints Intersection

No direct Five Breakpoints mapping, and the restraint is deliberate. The tempting tag is the breakpoint covering effort that produces no movement, on the strength of the negative coefficient at low leadership. It is withheld: two of sixteen interaction tests reached significance in a self-described exploratory analysis with no alpha correction at 21 sites, the two that did were the percentage outcomes while the two rating outcomes did not, and the setting is correctional-and-treatment process improvement in 2010–2013 rather than a technology deployment. The finding is strong enough to argue from and not strong enough to tag. Logged as (a) the counterpart instrument to the randomized leadership trial filed the same day (williams-aarons-ehrhart-loci-leadership-randomized-digital-uptake-2026-03.md) and (b) direct evidence on the sample arithmetic blocking asset #3b.

Four Forces Intersection

Commitment: The one interaction that survived says the leadership layer is the binding constraint and that effort below it does not substitute — the authors' own sentence is that staff responsiveness "does not compensate for impact of relatively weak leadership." Suggestive rather than established, for the reasons above.

Momentum: Fidelity scores declined significantly over the trial, independent of condition — measured decay of execution discipline across an eighteen-month organizational change effort, which is the Momentum force's claim observed rather than asserted.

Relevance to Brandon's Positioning

This entry's value is almost entirely methodological, and it is bad news that Brandon needs. The paper-2 gap ranking has asked, for weeks, for an organizational variable measured against an implementation outcome with selection handled. Implementation science runs that design as routine. This is one of the executed versions, and what it returns is a median fidelity-to-outcome correlation of .23 with most results non-significant — at a sample size (21 sites) very close to the ~30 that asset #3b has been priced at since 2026-08-30.

So the price of 3b was wrong in a way that matters. The 2026-08-30 update to asset #3 took Helfrich's power calculation (~30 sites for 90% power at R² ≥ .21) as the target. Stein et al. is what a ~20-site measurement study of organizational execution against organizational outcomes actually yields: mostly nulls, and effects that are large in standardized terms (d = .85–.93 on quality and adherence) yet non-significant because the n is sites, not people. Thirty sites is a floor, not a target — and the base should stop treating 3b's blocker as "find a partner" and start treating it as "this design is underpowered at any realistic organizational sample."

The escape is in the same day's other find, and the contrast is the finding. Stein measured the organizational variable at 21 sites and got noise. Williams et al. randomized the organizational variable at 20 clinics and got an order-of-magnitude effect with administratively logged outcomes. Same field, nearly the same number of organizations, opposite informativeness. The difference is manipulation versus measurement, and it is the single most useful design lesson this base has acquired for the instrument Brandon wants to build.

One guardrail on the asset-#3 essay. That essay's spine is that the discipline built to test whether readiness matters never finished the test (Helfrich's protocol with no results paper; Weiner's "the question of whether readiness for implementation matters remains open"). Stein does not break that spine — fidelity-to-implementation-strategy is a different construct from ex-ante readiness — but it does narrow the claim the essay may make. The essay must say readiness measured before a named change was never properly tested, and must not imply that the field never tests organizational variables against implementation outcomes. It does, it reports the results, and the results are weak.

Sources & Confidence

• Stein LAR, Bassett SS, Welsh WN, Clair-Michaud M, Abdel-Salam S, Monico L, Gallagher C, Murgo C, Yang Y, Friedmann PD, Clarke JG; CJ-DATS-2 Assessment Work Group. Substance Use & Misuse 2023 Jan 11;58(3):320–330. DOI 10.1080/10826084.2022.2161311. PMID 36629127; PMCID PMC9969992. — Moderate. Peer-reviewed and cluster-randomized, but exploratory by the authors' own description, 21 sites, no alpha correction, nesting unmodellable, fidelity partly self-reported, sites not randomly selected, and only a single leadership moderator tested. The authors call for "replication and a larger, fully powered trial."
• Supporting instruments cited by the paper, not independently verified here: Dane & Schneider (1998) fidelity framework; Broome et al. (2009) leadership scale; Working Alliance Inventory. — Moderate.

Filed by / Routing

Filed by: Stafford (scheduled run, 2026-10-08) — own exploratory slot, same query as the Williams et al. entry. Routing: strategy (asset #3b pricing), personal (design lesson for the instrument).

Tags: paper-2, Technology Illusion, Commitment, Capability, Momentum Source: Translational Behavioral Medicine (Oxford University Press / Society of Behavioral Medicine) — Williams, Aarons, Ehrhart, Vega & Marcus URL: https://academic.oup.com/tbm/article/16/1/ibag007/8519520 Date published: 2026-03-13 Date flagged: 2026-10-08

Instrument — recorded before the finding (mandatory for structural instruments)

Data source. Administrative system telemetry. Both outcomes are monthly clinic-level counts "derived from the OQ-A system which automatically logged every occurrence" — the measurement-based-care platform's own event log, not a survey, not a self-report, and not a researcher's observation. Clinician- and leader-report data were used only to build the leadership and climate development plans that the intervention arm acted on, never as an outcome. Panel: 632 clinic-months (33 monthly observations per clinic).

Inference method. Two-arm cluster randomized hybrid type III effectiveness–implementation trial (the WISDOM trial, NCT04096274). Randomization was at the clinic level, 1:1, covariate-constrained by clinic size and state. Both arms received identical training and technical assistance (TTA) in measurement-based care; the intervention arm additionally received LOCI for 12 months. 21 clinics enrolled, 20 analyzed (11 LOCI, 9 control — one clinic closed for unrelated reasons, one excluded as an extreme outlier). Outcome differences estimated at 3, 13 and 35 months with phase-specific slopes.

What was randomized, stated plainly because it is the whole value of the entry. The organizational variable itself. Leadership and implementation climate were manipulated, not measured — so the usual objection that capable organizations do more of everything cannot apply. Selection into "good leadership" was broken by design.

Observation window. Clinics enrolled June–October 2019; first LOCI training November 2019; measurement-based-care training December 2019; logging began one month after that training and ran 33 months, i.e. roughly January 2020 through late 2022. Entirely pre-generative with respect to AI, and spanning the COVID-19 disruption.

Setting. 20 outpatient mental health clinics serving youth in Idaho, Oregon and Nevada. Small service organizations, not large enterprises.

Summary

The authors tested whether changing an organization's leadership and implementation climate changes how much a deployed digital tool actually gets used. Every clinic in both arms received the same technology and the same training in measurement-based care — the practice of administering a standardized outcome questionnaire to a client and reviewing the resulting feedback report. Clinics randomized to the intervention arm additionally received LOCI (Leadership and Organizational Change for Implementation) for twelve months: first-level leaders completed quarterly trainings totalling 30 hours plus weekly ~20-minute coaching calls and quarterly one-hour organizational strategy meetings; executives attended those strategy meetings and monthly planning meetings; and each clinic leader worked an individualized leadership-and-climate development plan informed by survey data from their own clinicians. In the final quarter the activities began to address sustainment.

The two outcomes were monthly counts of youths who had the questionnaire administered, and youths whose feedback report was actually viewed by their clinician. The intervention arm was higher on both, at every measured timepoint, and all three timepoint differences were statistically significant. Adjusted differences were 5.09 [95% CI 1.63–8.55] more youths with the measure administered and 3.81 [1.26–6.37] more with feedback viewed at 3 months; 9.03 [1.64–16.41] and 8.31 [3.07–13.56] at 13 months; and 3.82 [1.28–6.36] and 1.41 [0.22–2.60] at 35 months. Expressed as ratios in the Discussion, LOCI clinics had roughly 10.6 times greater measure administration and 10.8 times greater feedback viewing at 3 months, falling to 3.4 and 6.5 times at 13 months.

The within-phase slopes tell a second story. During the implementation phase the arms' slopes did not differ (measure administered b = −0.11, P = .145; feedback viewed b = −0.05, P = .548) — the gap opened early and then held rather than widening. During the sustainment phase, LOCI clinics' measure administration declined less steeply than control (b = 0.38, SE = 0.11, P < .001), while feedback-viewing slopes did not differ (b = 0.01, P = .947). The authors report a post-LOCI decline in uptake and explicitly propose, but do not test, explanations for it. They did not analyze fidelity in this paper, and they report no mediation or moderation analysis here, citing a separate mechanistic paper from the same trial.

Five Breakpoints Intersection

Technology Illusion: This is the breakpoint's causal claim run as an experiment with the technology held constant. Both arms received the identical digital platform and the identical training in how to use it; the only difference was a twelve-month intervention on leadership and implementation climate, and that difference produced roughly tenfold more actual use of the same tool at three months and a still-significant advantage at thirty-five. The framework's assertion that deploying technology onto unchanged organizational conditions yields the appearance of deployment rather than its substance is normally inferred from failures after the fact; here the organizational condition is the randomized treatment and the tool's own event log is the outcome.

Four Forces Intersection

Commitment: The intervention is an operationalization of decision-makers acting consistently with a stated direction under pressure — weekly coaching of first-level leaders, quarterly executive strategy meetings, monthly executive planning — and randomizing it moved the outcome. This is the closest thing in this base to Commitment manipulated rather than asserted.

Capability: Evidenced as a null, which is why it matters. Skills and infrastructure were held identical across arms by design: every clinic got the same platform and the same training. Capability in the tooling-and-training sense was therefore sufficient for neither arm's uptake and cannot explain the gap — the base's standing claim that training is the input organizations buy instead of the one they need has, here, an experiment behind it.

Momentum: The advantage attenuated from roughly tenfold at three months to roughly threefold at thirteen, and absolute uptake declined in the sustainment phase in both arms, more slowly in the intervention arm on one of the two outcomes. The twelve-month support ended and the gap narrowed; the authors name the decline and decline to explain it. Sustained energy is the variable the trial's own tail is about.

Relevance to Brandon's Positioning

This is the answer to the request that has sat at the top of the paper-2 gap ranking for weeks, and it arrives in the wrong field and with the wrong outcome variable — both of which have to be said out loud before it is used. The standing request (gap ranking 2026-10-08, item 1) is for a non-technological organizational variable on the right-hand side of a technology-outcome specification, identified against selection. Williams et al. do not merely control for selection; they destroy it by randomizing the organizational variable, and they read the outcome off the platform's own log rather than asking anyone. On the identification question this is the strongest instrument in the base, and nothing else is close.

The two limits are severe and non-negotiable. First, the outcome is uptake, not value — nobody measured whether any youth got better. Leadership raised how much the tool was used; the trial is silent on whether use produced results. That is the exact distinction this base enforces against everyone else's instruments (the merged-PR proxy thread, the Unconverted Gain family), and it must be enforced here. Second, it is not AI — it is a questionnaire-and-feedback platform in youth mental health clinics, measured 2020–2022, in twenty small clinics in three states, and the authors themselves say the small number of clinics means inferences "should be treated with caution."

What it changes, and it is a design decision rather than a citation. Asset #3b has been blocked since 2026-08-30 on sample arithmetic: the Helfrich template says score organizations at baseline, wait 6–12 months, and see whether the score predicts, which needs ~30 sites for adequate power and a partner with access to deployments. Williams et al. got an order-of-magnitude effect at k = 20 — because manipulating an organizational variable buys far more statistical power than measuring one. The implication for 3b is that the readiness-instrument design is the expensive, weak version of the study, and the breakpoints-remediation design is the cheap, strong one. A randomized remediation that fails is also a cleaner answer to the standing unfalsifiability challenge than a predictive instrument that comes back null, because the latter is always attributable to the instrument.

And it creates a prior-art exposure that has to be logged rather than admired. LOCI is a named, manualized, trialled, twelve-month leadership-and-climate intervention with Gregory Aarons' research programme behind it. If Brandon's remedy is "fix the organizational conditions before you deploy," that remedy has an established competitor with a registered RCT — the same class of exposure as the Toyama/Kerr/Kotter lineage work of 2026-09-17. Logged to knowledge/thesis-challenges.md.

Deliberately not tagged, and the reason is the discipline: the breakpoint that reads "appearance of progress without movement" is declined for this entry. The uptake decline in the sustainment phase is real attenuation of a real behaviour change, not appearance without substance, and tagging decay as that breakpoint would stretch it. The durability finding is recorded under the Momentum force, where it belongs.

Sources & Confidence

• Williams NJ, Aarons GA, Ehrhart MG, Vega N, Marcus SC. "Effects of a leadership-focused implementation strategy on uptake of digital measurement-based care in mental health clinics: a cluster randomized trial." Translational Behavioral Medicine 16(1):ibag007. DOI 10.1093/tbm/ibag007. Received 18 Sep 2025; accepted 5 Feb 2026; published 13 Mar 2026. Trial registration NCT04096274. — Moderate. The design is the strongest in this base on the identification question (cluster randomized, covariate-constrained, administratively logged outcomes, 632 clinic-months, registered trial). The rating is held at moderate rather than high on three grounds the base applies to everyone: the publisher sits outside the named high-credibility list; k = 20 clinics, with the authors' own caution about inference; and the analyses were not formally preregistered even though the trial was registered. Code and data are not publicly archived.
• Authors' stated limitations carried verbatim where load-bearing: outcomes "do not capture clinicians' use of feedback to guide treatment"; the missing-at-random assumption "cannot be empirically verified"; the post-LOCI decline is proposed but untested. The paper's lay summary claims LOCI clinics "followed best practices more closely" — that claim is not supported by the outcomes analyzed here, which are uptake counts only, and must not be repeated.

Filed by / Routing

Filed by: Stafford (scheduled run, 2026-10-08) — own exploratory slot, executing the method instruction the 2026-10-08 research-evidence scan handed forward. Routing: strategy (asset #3b design decision), personal (prior-art exposure on the remedy).

Tags: paper-2, Strategic Disconnection, Momentum Mirage, Capability, Commitment, Momentum, COUNTERARGUMENT Source: University of Antwerp, Faculty of Social Sciences (institutional repository) URL: https://repository.uantwerpen.be/docman/irua/975b53motoM68 Date published: 2026-06-29 Date flagged: 2026-10-04

Summary

A doctoral dissertation defended at the University of Antwerp on 29 June 2026, supervised by Koen Verhoest, Jan Wynen and Bjorn Kleizen — the research group that owns the change-complexity construct the 2026-10-04 scan logged as its highest-value unread chase target (Verlinden et al. 2023a, cited here as the measurement precedent). The organising construct is change intensity: "employees' perceived exposure to cumulative and ongoing change environments in which multiple change initiatives unfold simultaneously or in rapid succession." The dissertation builds an integrative framework from three perspectives — information processing (intensity raises informational demands and uncertainty), social exchange (intensity erodes psychological contract fulfilment), and stress/strain ("continuous adaptation depletes resources") — and tests it in three empirical articles.

All three articles draw on the same site: a large Belgian public organisation in the welfare domain (~2,900–3,100 employees), surveyed in two waves one year apart (wave 1: 9 March – 10 May 2021, 1,393 respondents, 40.9%; wave 2: 20 April – 25 May 2022, 1,139 respondents, 39.7%). The restructuring under study "primarily involved incremental changes to organizational processes aimed at improving service delivery" and "did not result in job losses among employees." Article 1 (661 respondents, quasi-natural experiment with entropy-balancing matching) finds that past change intensity is negatively related to appraisal of a new change in terms of personal benefit, while leaving appraisal of organisational benefit untouched: change complexity on personal-benefit appraisal −0.452 (SE 0.17, p = 0.009; model F(1,660) = 6.84, p = 0.0091); on organisational-benefit appraisal the model is not significant (F(1,660) = 0.25, p = 0.6162). The perceived-pace measure reproduces the same split: personal benefit −0.638 (SE 0.237, p = 0.007), organisational benefit 0.03 (SE 0.269, p = 0.912). The author notes the effect "holds even when accounting for levels of trust in change management." Article 2 (699 respondents linked across waves) finds intense change environments associated with lower psychological contract fulfilment, and tests change-management quality as a moderator. Article 3 (survey plus eight focus groups) finds intensity associated with higher qualitative job insecurity — concern about the future quality and sustainability of one's role rather than fear of job loss — operating through four mechanisms, the second of which is "erosion of role clarity during implementation and ongoing fine-tuning, as roles and responsibilities failed to stabilize."

The dissertation's cross-cutting result is about the remedy rather than the disease, and the author states it twice. From Article 2: "Change management shows a direct, rather than moderating, effect — suggesting that while it is positively associated with outcomes, it does not fundamentally alter the relationship between change intensity and psychological contract fulfillment." In the synthesis: "Across the studies, change management emerges as a distinct and relevant factor that is positively associated with employee outcomes. However, it does not moderate the relationship between change intensity and these outcomes. This suggests that, while effective change management can improve employees' experiences of change, it does not fundamentally alter how intense change environments translate into psychosocial strain." The closing recommendation follows from it: supporting employees under sustained change "requires not only investing in communication, participation, implementation support, and visible leadership, but also actively managing the pace, sequencing, and overlap of change trajectories."

Five Breakpoints Intersection

Strategic Disconnection: Article 3's second mechanism is role clarity eroding "during implementation and ongoing fine-tuning, as roles and responsibilities failed to stabilize" — a case where the change was announced, communicated and managed, and employees still could not say what their job was, which is the illusion of alignment measured at the individual level rather than asserted of the organisation.

Momentum Mirage: Article 1's central split — employees with a high-intensity change history appraise the next change less positively for themselves while continuing to grant it organisational benefit (personal benefit −0.452, p = 0.009; organisational benefit not significant) — is the appearance of progress and the absence of movement, measured on the same respondents at the same moment, and the divergence grows with the number of prior changes.

Four Forces Intersection

Capability: The stress/strain arm treats absorptive capacity as a depleting stock — "continuous adaptation depletes resources" — and the no-moderation result means the depletion is not offset by running the change well, so capacity to absorb change behaves as a quantity spent rather than a skill exercised.

Commitment: Psychological contract fulfilment, measured as agreement with "I feel my employer has fulfilled the promises made during the change processes of recent years," falls as change intensity rises, and the qualitative arm attributes it to overlapping trajectories that "undermine continuity, follow-through, and alignment with earlier expectations" — follow-through under pressure, in the Four Forces sense, measured directly.

Momentum: The appraisal of a future change degrades as a function of the history of prior changes, which is momentum decay measured prospectively rather than inferred from a stalled programme.

Relevance to Brandon's Positioning

This is the strongest measured objection in the base to the structure of a remedy framework, and it needs stating before its support value. Paper 1's architecture implies an interaction: do the Four Forces work and the breakpoints do less damage. This dissertation tests the nearest available proxy for "did you do the work" — change-management quality — as a moderator of change intensity's effect, and finds a direct effect only. Good change management makes people better off on average; it does not change the rate at which cumulative change damages them. That is one interaction test in one organisation and must not be inflated beyond that, but it is a test, it is prospective, and it points the wrong way. It also converges with an instrument already in the base from an unrelated field: Faros AI (research-evidence/knowledge-base/faros-ai-engineering-report-acceleration-whiplash-2026-04.md, 22,000 developers, within-organisation telemetry) found strong DORA-style engineering foundations give "no protection against this downstream deterioration regardless of baseline maturity." Two instruments, two substrates, two literatures, same shape: the prescribed condition has a level effect and not a protective one. The base previously read the Faros result as support for the thesis, by substituting an unmeasured moderator (redesign) for the measured one (maturity); that reading should be revisited rather than repeated here. Logged as a new challenge in knowledge/thesis-challenges.md (2026-10-04).

Where it supports, it supports in paper-2's emptiest cell, and in the vocabulary the 2026-10-04 scan said that cell needs. The scan opened The Unpriced Delayering that morning on the observation that no organisation prices the capacity a structural change spends, and posed as the day's open question what unit such a price would even be in. This dissertation is a field that has been measuring exactly that for fifteen years and has a published answer: change intensity as a count of concurrent and successive change types per employee over three years (Article 1 reports 71% of respondents experiencing 4–6 distinct change types in three years), with the explicit prescription to manage "pace, sequencing, and overlap." The construct and the instrument exist; what nobody has done is put an AI variable on the other side of them. That is a cheaper and more defensible route into the capacity question than inventing a span-and-queue measure, and it is a direct contribution to growth edge #2 in knowledge/brandon-development.md.

Hard limits, and the first one bounds every use. No AI variable anywhere, and the window is pre-generative — the surveys ran March 2021 and April 2022, with change history covering roughly 2018–2021, so this enters on the same previous-shock footing as the CPOE and information-routing literature and may not be cited as evidence about AI deployment. Single site, single country, single sector (Belgian public welfare), and a restructuring that produced no job losses — so it is the low-stakes end of the change-intensity range, not the delayering end. Psychological contract fulfilment in Article 2 is measured with a single survey item. The sample is representative overall but the author reports imbalance on Level, Tenure and some team-cluster dummies, handled with control variables rather than weights. Article 1's organisational-benefit point estimate could not be read reliably from the repository PDF's table extraction — only the model-level non-significance (F(1,660) = 0.25, p = 0.6162) is cited here, and the coefficient is deliberately omitted rather than guessed.

Sources & Confidence

• Pieters, D. (2026), Navigating Turbulence in the Public Sector: The Psychosocial Impact of Intense Organizational Change on Civil Servants, doctoral dissertation, University of Antwerp, defended 2026-06-29 (https://repository.uantwerpen.be/docman/irua/975b53motoM68) — Moderate. Full text read directly from the institutional repository; source of every figure and quotation above.
• Filed moderate by the same rule the 2026-10-04 scan applied to GAO: CLAUDE.md's high-credibility publisher list is enumerated, and the University of Antwerp is not on it. The instrument is stronger than that grade suggests — a jury-examined dissertation whose constituent articles publish in Review of Public Personnel Administration — and that discrepancy is the second instance today of the enumerated list under-grading a disinterested instrument. Routed with the scan's own proposed CLAUDE.md amendment; not acted on unilaterally.
• Verlinden, S., Wynen, J., Kleizen, B., & Verhoest, K. (2023a), "Blurred Lines: Exploring the Impact of Change Complexity on Role Clarity in the Public Sector," Review of Public Personnel Administration 43(3), DOI 10.1177/0734371X221093573 — Moderate; NOT READ. Publisher returned HTTP 403. Cited here only as this dissertation reports it (Australian Public Service Employee Census; rising change complexity correlates with greater reductions in role clarity). It remains an open chase target in source-candidates.md and no claim in this base may rest on it until the text is read.

Filed by / Routing

Filed by: Stafford (judgment layer, exploratory slot — 2026-10-04) Routing: strategy (the no-moderation result changes how the framework's remedy claim should be worded); personal (growth edge #2 — the change-intensity instrument is the readiness measure's missing dose variable)

Instrument note (recorded before the finding, per CLAUDE.md)

Data source genre: employee survey, two-wave panel, single organisation, plus eight focus groups (Article 3). Not résumé or postings inference — this instrument does not share a right-hand side with the Babina / Ewens & Giroud lineage and is independent of it. Inference method: Article 1 is a quasi-natural experiment comparing high- and low-change-intensity employees with entropy-balancing matching on confounders including trust in change management; Articles 2–3 are two-wave regressions with demographic controls plus focus-group analysis. No causal identification beyond matching on observables; no administrative or payroll records; no outcome variable external to the respondent — every dependent variable is self-reported perception (change appraisal, psychological contract fulfilment, qualitative job insecurity). Observation window: March 2021 – May 2022, with change history covering approximately 2018–2021. Pre-generative; no AI variable. Sample: 661 (Article 1), 699 linked (Article 2), of a ~2,900–3,100-employee population at ~40% response.

Verification addendum 2026-10-06 (Stafford) — what this dissertation establishes about Verlinden et al. 2023a, the paper it cites as its measurement precedent

Why this addendum exists. Verlinden et al. 2023a carries paper-2 Claim 3's Strategic Disconnection cell and cannot be retrieved: four catalogues (Crossref, OpenAlex, Semantic Scholar, Unpaywall) independently report exactly one copy in existence, publisher-hosted at SAGE, which returns HTTP 403 to automated clients; Unpaywall gives oa_locations count 1 / host_type: publisher, OpenAlex gives any_repository_has_fulltext: false. It is openly licensed (CC BY-NC 4.0) and un-fetchable, which are not the same thing and were being recorded as the same thing. This dissertation was re-fetched and full-text searched on 2026-10-06 because it cites that paper as its measurement precedent, and a methods section describing a borrowed measure is a partial read of the source it borrowed from.

Verified from this document (direct quotations):

1. The change-complexity measure is a count of discrete events, not a rating scale. Article 1: respondents report whether they "experienced this particular change in the past 3 years (the 3 answer categories are: not at all, once, more than once)" across seven types — "change in physical workplace, change in job function, change in job content and procedures, change in team composition, change in departmental structure, IT-related changes as well as a category 'other'" — and "the measure of change complexity consists of the summative score of the number of change types experienced … and therefore aligns with the measurement of change complexity as used by Verlinden et al. (2023a)."
2. The dichotomous recoding is cited to Verlinden 2023a directly. Article 2: "As we are interested in the types of change experienced, these items were recoded into dichotomous versions indicating whether respondents have experienced a given change type or not (Verlinden et al, 2023a)."
3. The research group names common-source bias and describes the measure as mitigating it: "It should also be noted that the change complexity variable has the added benefit of not being measured through a Likert scale, reducing common source and/or method bias issues."
4. Verlinden et al. 2023a used probability weights. "Two ways to deal with remaining bias are to employ sample weights and to incorporate known sources of bias as control variables. This study opts for the latter … This approach avoids potentials biases associated with using both control variables and probability weights simultaneously (Verlinden et al., 2023a)."
5. Setting and result direction, stated twice. "the study by Verlinden et al. (2023a) in the Australian public service found that rising levels of change complexity correlate with greater reductions in role clarity, indicating that the pace of change can sometimes surpass employees' ability to adapt"; and "Verlinden et al. (2023) found that high levels of change complexity lead to increases in perception of role ambiguity."

Grade, stated precisely. Passages 1–3 describe this dissertation's variable and attribute the operationalization to Verlinden 2023a by explicit citation; Verlinden is a member of this research group and is thanked in the acknowledgments. This is a near-authoritative second-hand attribution by co-authors — substantially verified, short of a direct read of Verlinden's methods section. It is not a substitute for reading the paper.

Still NOT verified, and still quarantined. Census wave(s), all sample sizes, matching covariates, every coefficient. The figures circulating in web summaries (APS 2014 census, 99,392 employees / 89 agencies, 76,275 after item non-response, a random 10% subsample of 7,627) remain secondhand from a single route and are NOT entered in the base. Finding 4 above is consistent with a complex census sample requiring weights; it confirms none of those numbers. Hard rule 14 applies.

What this changes elsewhere. It partly refutes the 2026-10-05 challenge logged against this base's own method ("You apply the one-questionnaire standard to vendors and waive it for peer review") — the field named the bias and designed against it on the independent-variable side — and replaces the base's common-source binary with a three-rung ladder. See the Update 2026-10-06 entry in knowledge/thesis-challenges.md for the re-scoring and the revised case-file grade proposed for Claim 3 (suggestive — asymmetric common-source (rung 2), not the flat suggestive, common-source proposed on 10-05).

Tags: paper-2, Technology Illusion, Process Friction, Momentum Mirage, Capability, Momentum, CONVERGENT SIGNAL Source: DORA (Google Cloud) URL: https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report Date published: 2025-09-23 Date flagged: 2026-10-03

Summary

DORA's 2025 edition — retitled from Accelerate State of DevOps to State of AI-assisted Software Development — reports survey responses from "nearly 5,000 technology professionals" worldwide plus "over 100 hours of qualitative data," published 23 September 2025. Adoption is near-saturation in the sample: 90% of respondents report using AI at work and more than 80% believe it has increased their productivity.

The headline result is a split, stated by DORA in two consecutive sentences. On throughput: "Unlike last year, we observe a positive relationship between AI adoption on both software delivery throughput and product performance." On stability: "However, AI adoption does continue to have a negative relationship with software delivery stability." DORA reads the pair as confirmation of a prior expectation: "This confirms our central theory — AI accelerates software development, but that acceleration can expose weaknesses downstream. Without robust control systems, like strong automated testing, mature version control practices, and fast feedback loops, an increase in change volume leads to instability." The 2024 edition had found AI adoption negatively related to both throughput and stability; the throughput sign flipped in 2025 and the stability sign did not.

The report's organizing claim is the amplifier: "AI doesn't fix a team; it amplifies what's already there. Strong teams use AI to become even better and more efficient. Struggling teams will find that AI only highlights and intensifies their existing problems. The greatest return comes not from the AI tools themselves, but from a strategic focus on the quality of internal platforms, the clarity of workflows, and the alignment of teams." The moderating condition is architectural and procedural rather than cultural: "Teams working in loosely coupled architectures with fast feedback loops see gains, while those constrained by tightly coupled systems and slow processes see little or no benefit." The report also introduces seven team archetypes, of which two are named in the announcement — "Foundational challenges" ("trapped in survival mode… low performance, high system stability, and high levels of burnout and friction") and "Harmonious high achievers."

Five Breakpoints Intersection

• Technology Illusion: This is the breakpoint stated almost verbatim by a research programme that does not use the term — "AI doesn't fix a team; it amplifies what's already there… The greatest return comes not from the AI tools themselves, but from a strategic focus on the quality of internal platforms, the clarity of workflows, and the alignment of teams" — which is the proposition that technology deployed on top of unchanged organizational conditions returns little, asserted on roughly 5,000 respondents rather than on cases.
• Process Friction: The moderator is a handoff property, not a talent property — "teams working in loosely coupled architectures with fast feedback loops see gains, while those constrained by tightly coupled systems and slow processes see little or no benefit" — so the same tool yields opposite outcomes depending on machinery nobody redesigned, which is the breakpoint's definition.
• Momentum Mirage: Throughput and product performance rose while delivery stability fell on the same population in the same wave, meaning the metrics a rollout reports moved favourably while the metric the production system experiences moved adversely — progress visible in the reporting layer and not in the system.

Incentive Fragmentation and Strategic Disconnection are not tagged. The announcement contains no finding on how teams or individuals are measured, on decision rights, or on whether the organization agreed what the AI deployment was for. Tagging either would be the loose mapping the rules forbid.

Four Forces Intersection

• Capability: Capability in the framework's flow sense is the report's explicit subject — automated testing, version-control maturity, feedback-loop speed and loose coupling are named as the conditions determining whether added change volume becomes delivery or becomes instability, which is the operating model's capacity to absorb work rather than the talent available to do it.
• Momentum: The split result describes momentum failing to convert rather than fading — change volume up, stability down — so the organization generates more movement than its control systems can turn into reliable outcomes.

No Purpose or Commitment tag. "Alignment of teams" and "clarity of workflows" are adjacent to Purpose but are reported as platform and workflow properties, not as clarity of intent, and the announcement carries no item on either Force. Honest omission rather than a stretched tag.

Relevance to Brandon's Positioning

This entry is simultaneously the strongest external corroboration of Brandon's Technology Illusion claim in the base and the largest positioning exposure the base has recorded — and the second matters more than the first. DORA states Brandon's central proposition, in plainer language than he has published it, backed by ~5,000 respondents, under Google Cloud's distribution, to precisely the engineering-leadership audience his papers address, and it did so in September 2025. A reader who knows DORA will not find Brandon contradicted; they will find him arriving second without acknowledgement. Verified today: DORA, Forsgren, Accelerate, "State of DevOps" and "error budget" appear nowhere in any of Papers 2–6 or in books/built-to-be-replaced.md.

It also discharges half of a reading assignment this ledger has carried for thirteen days. knowledge/thesis-challenges.md ("Google is a special case," CONTESTED, 2026-09-18, updated 09-20 and 09-21) records the conversion path as "Accelerate plus the current DORA report, then the two-variable diagnostic stated in print with at least one of them behind it." The current DORA report is now read. The result cuts for Brandon on the narrow question — DORA measures capabilities against outcomes and still does not measure the installation of a shared interface measure between two parties with separate leadership and separate metrics, so the two-variable diagnostic remains unclaimed territory — and against him on the broad one, by owning the general claim.

Its third and most useful role is as a contrast instrument. This base's merged-PR cluster (DX, LinearB, Murphy-Hill et al., all filed 2026-10-03) shares one outcome variable produced at the review gate. DORA measures delivery stability — change failures, rework, recovery — which is detected after release by the production system rather than by a reviewer, and therefore cannot be held flat by review capacity running out. On that outcome, at n≈5,000, the sign is negative. That makes DORA an independent discharge of the escalation criterion the Merged-PR Proxy pattern pre-registered on the day it opened, and it agrees with Faros AI's telemetry (faros-ai-engineering-report-acceleration-whiplash-2026-04.md: bugs per developer +54%, incidents-to-PR ratio +242.7%) against He et al.'s gated metrics (he-denisov-blanch-2x-mandate-review-capacity-2026-07.md: "merge and revert rates held steady").

One direct conflict must be carried, not smoothed. Faros reports that strong pre-existing DORA-style engineering foundations provide no protection against downstream deterioration regardless of baseline maturity. DORA reports that those same foundations are exactly what separates gain from instability. Same function, same period, opposite answers to the most practically consequential question either asks. Faros is within-organization telemetry on 22,000 developers; DORA is a large opt-in survey with self-reported practice maturity. Neither should be cited on this point without the other, and the disagreement is more useful to Brandon than either verdict.

Sources & Confidence

Instrument note (recorded before the finding, per CLAUDE.md hard rule 9). Data-source genre: opt-in practitioner survey, ~5,000 self-selected technology professionals worldwide, supplemented by 100+ hours of qualitative interviews. Inference method: correlational — relationships between a self-reported AI-adoption measure and self-reported delivery outcomes, with moderators (testing maturity, version-control practice, feedback-loop speed, architectural coupling) likewise self-reported; no causal identification, no panel, no randomization, and both sides of every reported relationship come from the same respondent, which is the instrument's binding limit and the one Faros's telemetry does not share. Observation window: 2025 survey wave; report published 2025-09-23. Shared-respondent caution: DORA's AI-adoption measure and its outcome measures are reported by the same person, so this is not an independent-measurement design and must never be cited as telemetry.

• DORA / Google Cloud, 2025 DORA Report: State of AI-assisted Software Development, published 2025-09-23 — https://cloud.google.com/blog/products/ai-machine-learning/announcing-the-2025-dora-report — High (publisher: Google Cloud; multi-year established research programme). Source of every verbatim quotation above, the ~5,000 sample, the 100+ qualitative hours, the throughput/stability split, the amplifier statement, the coupling moderator and the two named archetypes. Fetched from the first-party announcement page.
• Report landing page and full-report access: https://dora.dev/research/2025/dora-report/ and https://cloud.google.com/resources/content/2025-dora-ai-assisted-software-development-report — High as locators. The full report PDF was not read here; this entry rests on the first-party announcement, and no figure from the report body is cited. A 2025 errata page exists at dora.dev/research/2025/ and has not been checked — flagged, because this base has twice frozen on a superseded number.
• Version check performed today, per CLAUDE.md hard rule 14. dora.dev/research/publications/ lists no 2026 report or 2026 publication of any kind; the newest entries are the 2025 State of AI-assisted Software Development and the 2025 AI Capabilities Model. DORA's own 2025 year-in-review names three 2025 outputs — Impact of Generative AI in Software Development (March), State of AI-assisted Software Development (September), AI Capabilities Model (December) — and announces no 2026 edition. So the September 2025 report is the current edition as of 2026-10-03, and the several third-party pages titled "DORA 2026" are aggregator restatements, not DORA primaries. — High for the version fact.
• Not cited and deliberately excluded: the widely circulated figure that a 25% increase in AI adoption was associated with a 1.5% decrease in throughput and a 7.2% decrease in stability. It is attributed to DORA 2024, reached here only through aggregators, and was not verified against the 2024 primary. Low — do not cite until read from the 2024 report. The 2025 directional statements above do not depend on it.
• Faros AI, The AI Engineering Report 2026 (2026-04-12) and He, Agarwal, Denisov-Blanch et al., arXiv:2607.01904 (2026-07-02) — Moderate, both already held in this base, cited above only to situate the agreement and the conflict.

Filed by / Routing

Filed by: Stafford (claude-opus-5), daily brief exploratory slot, 2026-10-03 Routing: strategy / market

Tags: paper-2, practitioner, Technology Illusion, Momentum Mirage, COUNTERARGUMENT, Capability, Momentum Source: arXiv (Murphy-Hill, Butler & Savelieva — Microsoft) URL: https://arxiv.org/abs/2607.01418 Date published: 2026-07-01 Date flagged: 2026-10-03

Summary

Emerson Murphy-Hill, Jenna Butler and Alexandra Savelieva study Microsoft's early-2026 rollout of two command-line agentic coding tools — Anthropic's Claude Code and GitHub's Copilot CLI — across "tens of thousands of engineers," over a four-month window. The paper's framing is explicitly an organizational-economics one rather than a tool-evaluation one: "At organizational scale, token spend can run into millions of dollars annually, so misreading adoption, retention, or impact can make a rollout expensive without changing engineering velocity."

Three findings are reported. First use spread primarily through social networks rather than through demographics or mandate. Retention was "associated more with engineers' coding activity than with demographics." And "adopters merged roughly 24% more pull requests than they would have otherwise," with the authors stating that "the lift persists across our four-month window" — which is their answer to the novelty-effect objection. Their conclusion: "These results suggest that CLI coding agents are neither uniformly adopted nor mere novelty effects and that organizations should treat visible peer use as central to rollout strategy."

The sentence that makes this entry more valuable than its headline number is a methodological disclaimer the authors volunteer: "We use merged pull requests as our proxy for output — acknowledging that a merged PR is not the same as the value it delivers." That is the authors of the best-identified study on this substrate declining, in print, the exact inference that the rest of the genre makes silently.

Five Breakpoints Intersection

Technology Illusion — as counter-evidence, and more careful counter-evidence than the rest of the genre: a measured ~24% lift in merged pull requests among adopters, persisting across four months inside a single organization, is evidence that a technology deployment produced sustained movement in a hard output measure rather than being absorbed into existing habits — the outcome Paper 1's Technology Illusion predicts organizations usually fail to get. The entry is tagged as counterargument rather than support because that is what it is, and because the authors' own proxy caveat is the only thing standing between this finding and a straightforward refutation of the breakpoint in this setting.

Momentum Mirage: The paper's own framing supplies the mechanism this breakpoint names — a rollout that can "make a rollout expensive without changing engineering velocity" — and its finding that first use spreads through visible peer use rather than through mandate or governance is a direct observation that adoption momentum in an AI deployment is carried by social reinforcement, which is Paper 1's claim that momentum is the organization's capacity to convert intent into visible progress and reinforce it fast enough to sustain belief.

Strategic Disconnection, Incentive Fragmentation and Process Friction are not tagged. The study measures adoption, retention and output; it contains nothing on outcome definition, on how engineers or their managers are measured, or on handoffs and decision rights.

Four Forces Intersection

Capability: The clearest Capability-as-readiness measurement in this base on an AI deployment — output per adopter rising ~24% and holding for four months inside one firm's own records — and the authors' proxy caveat marks precisely where that Capability reading stops.

Momentum: Retention tracked coding activity rather than demographics, and first use spread through social networks, which together describe an adoption curve sustained by visible use rather than by sponsorship — the force operating as Paper 1 describes it, observed rather than asserted.

No Purpose or Commitment implication. The instrument has no item bearing on either.

Relevance to Brandon's Positioning

This is the highest-quality member of a three-instrument cluster found on 2026-10-03, all reading organizational outcomes off version-control records, and its role in Brandon's argument is specific: it is the one that supplies the discipline rather than the number.

Taken alone, the ~24% lift is counter-evidence to the machine-speed thesis in the one function where AI deployment is deepest. It is within-organization, it survives four months, and the identification is better than anything else on this substrate — which is exactly why Brandon should cite it rather than avoid it. The position worth defending is not that AI deployments do not raise output. It is that output raised at a review gate is not the same thing as organizational capability, and the authors of this paper say so in their own words.

That sentence — "a merged PR is not the same as the value it delivers" — is the citation that holds together today's harder argument. DX reports daily AI users reaching their 10th merged PR in 49 days against 91, and reads it as onboarding time halving. LinearB reports AI-assisted PRs merging within 30 days at 32.7% against 84.5% unassisted. Three independent record-based instruments, one shared left-hand side: the merged pull request. When the most carefully identified of the three states that its own outcome variable is a proxy for output and not for value, the genre's convergence on that proxy becomes the finding — and it is the same structural problem as the Résumé-Panel Consensus pattern this base already tracks, moved from the right-hand side to the left.

Limits.

1. A preprint, not peer-reviewed, filed 2026-07-01 and not located in a journal.
2. The authors are Microsoft researchers studying Microsoft's own rollout of a Microsoft product (Copilot CLI) alongside a partner's. That does not impeach the measurement, which is drawn from internal records, but it bears on what gets published and how it is framed.
3. "Than they would have otherwise" is doing unexamined work. The abstract states the counterfactual without naming the design that produces it; whether adopters are compared to matched non-adopters, to their own pre-adoption trend, or to a modelled baseline is not determinable from the abstract, and the selection problem is the whole question — engineers who adopt a CLI agent early are the engineers whose coding activity predicted retention, by the paper's own second finding.
4. Tens of thousands of engineers at one firm, and that firm is Microsoft. Nothing here transfers to an ordinary enterprise without the transfer being argued: the population is unusually senior in tooling maturity, and the reviewers absorbing the extra PRs are Microsoft reviewers.
5. The full text has not been read here — this entry rests on the abstract and title page as fetched. Every quotation above is from that text; no figure from the body is cited, and none should be until the paper is read in full.

Sources & Confidence

• Emerson Murphy-Hill, Jenna Butler & Alexandra Savelieva, Adoption and Impact of Command-Line AI Coding Agents: A Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI, arXiv:2607.01418, submitted 2026-07-01 (https://arxiv.org/abs/2607.01418) — Moderate. Source of the title, authors, submission date, the tens-of-thousands sample, the four-month window, the social-network spread and activity-based retention findings, the ~24% merged-PR lift, the persistence statement, and the verbatim proxy caveat. Abstract fetched directly; full text not read. Filed moderate rather than high: an unreviewed preprint by researchers employed by the organization studied, with the counterfactual design not stated in the abstract. Against that, the unit of observation is internal version-control records at a scale and with a within-organization frame that no survey in this base matches.
• DX (2025-09-10, 2026-04-09) and LinearB (2026-03-24) — Moderate, cited here only to situate the shared-proxy observation. Both filed separately.

Instrument note (recorded before the finding, per CLAUDE.md). Data source genre: telemetry — internal records at a single firm (Microsoft), covering tens of thousands of engineers over a four-month window in early 2026, combining tool-usage data with merged pull-request counts. Inference method: adopters compared against a stated counterfactual ("than they would have otherwise") whose construction is not specified in the abstract; the outcome is merged pull requests, which the authors explicitly label a proxy for output and distinguish from delivered value; adoption spread and retention modelled against social-network exposure, coding activity and demographics. Observation window: approximately four months, early 2026. Shared left-hand side: the merged pull request is the outcome variable in this entry, in dx-developer-ramp-time-10th-merged-pr-ai-use-2026-04 and in linearb-2026-benchmarks-ai-pr-merge-rate-2026-03. These three are not independent confirmations of anything stated in terms of that variable, and must never be cited as such — the same prohibition CLAUDE.md applies to Babina et al. and Ewens & Giroud, here arising on the left-hand side rather than the right.

Filed by / Routing

Filed by: Claude Code daily scanner (daily scan, 2026-10-03) Routing: strategy

Verification addendum — 2026-10-03 (Stafford, same day, full text read)

Why this addendum exists. The entry above was written from the abstract and says so at limit 5. The 2026-10-03 scan opened a thread recording the full-text fetch as "one cheap item… owed before then [with] an owner and a date," on the grounds that "this base has now watched three windows and one citation hold die for want of exactly that." The full text was fetched from https://arxiv.org/html/2607.01418v1 the same afternoon. It overturns one recorded limit, supplies two things the abstract withheld, and discharges a thread criterion in the negative. Added as a dated addendum rather than by rewriting the entry, so the abstract-level reading and its correction both stay visible.

1. Limit 3 is withdrawn. The counterfactual is specified, and it is better than the entry assumed. Limit 3 said "'than they would have otherwise' is doing unexamined work" and that "the selection problem is the whole question." The paper names two designs, not one:

• Between-person: Bayesian structural time-series via CausalImpact against a synthetic control built from never-users — verbatim, "The synthetic counterfactual is built from a pool of 10 daily-mean regressors: we sampled PR creators who never used either tool, randomly partitioned them into 10 groups, and averaged each group's daily PR-per-engineer series." Result: "CausalImpact estimates a +24.0% lift in PRs/engineer/day over the post-period [95% CI +14.5%, +33.7%], posterior tail-area p<0.001."
• Within-person: a fixed-effects Poisson regression with engineer and week fixed effects, comparing each engineer against their own zero-use weeks — which is the specification that directly addresses the selection concern the entry called decisive.

So the headline number has a stated interval and two designs behind it, one of which holds the engineer constant. This instrument is stronger than this base recorded it as being, and the correction runs in the direction that favours the counterargument. confidence stays Moderate on the preprint and employer-studies-itself grounds, which are untouched.

2. The effect is not monotonic in seniority. It is a C shape, and that is new information. The paper breaks the lift down by career stage (IC4 reference: +21.3% [+19.6%, +23.0%]) and by tenure (5–15 years reference: +21.5% [+19.9%, +23.1%]). In both cuts the extremes gain more than the middle — "junior ICs and senior managers" show the larger lift by career stage, and "low- and high-tenure engineers experienced larger lifts than mid-tenure engineers." The authors flag the <1-year-tenure point estimate as large but caution it may be confounded by onboarding.

That last caution is the most consequential sentence in the full text for this base, and it is a direct collision with the entry filed alongside this one. dx-developer-ramp-time-10th-merged-pr-ai-use-2026-04.md reads a faster new-hire ramp as AI compressing time-to-competence. Murphy-Hill et al., on the same substrate in the same month, treat the equivalent new-hire effect as a confound to be warned about. The same correlation is one instrument's headline finding and the other's threat to validity. A C-shape is also not what the learning-rate channel predicts: Friebel et al. require juniors qualifying as seniors faster, which is a story about the bottom of the distribution; a pattern where junior ICs and senior managers both gain more than mid-career engineers is a task-composition story, not a qualification story.

3. Thread criterion (c) is discharged — negative, and the authors say so in print. The learning-rate thread's counter-evidence criterion was narrowed on 2026-10-03 to independent judgment: unsupervised decision authority, reduced senior review, revert or defect rates, promotion to a sign-off role. The full text contains none of it. No review burden or reviewer load, no revert rate, no defect or incident measure, no code-quality metric, and nothing on whether tool use changed an engineer's authority to approve changes or obtain sign-off. The paper states the gap itself: verbatim, "The pressing open question is now about quality, whether this added throughput yields better software. The field still lacks agreed-upon measures to answer it."

That quotation is more valuable to Brandon than the proxy caveat the entry above features, and should replace it as the headline citation. The proxy caveat ("a merged PR is not the same as the value it delivers") is an author limiting their own paper. This one is the authors of the best-identified instrument on the substrate stating that the field has no agreed measure of the thing everyone is inferring — a claim about the literature, not about one study, made by people with every incentive to report that the rollout worked.

Sources & confidence for this addendum. Murphy-Hill, Butler & Savelieva, arXiv:2607.01418v1, full text at https://arxiv.org/html/2607.01418v1, read 2026-10-03 — Moderate (unreviewed preprint, authors employed by the organization studied; all quotations above taken from that HTML text, including the CausalImpact specification and estimate, the Figure 9 career-stage and Figure 10 tenure reference estimates, the onboarding-confound caution, and the open-question sentence). Figure-level point estimates for the non-reference cells are reported in the paper as figures and are described directionally here rather than quoted as numbers, because they were not read off the plots.

Instrument note, amended. Inference method is no longer "not specified in the abstract": it is CausalImpact BSTS against a 10-pool synthetic control of never-users, plus engineer-and-week fixed-effects Poisson on within-person variation, outcome PRs per engineer per day, window ~4 months in early 2026, population tens of thousands of Microsoft engineers. The shared left-hand side with the DX and LinearB entries is unchanged and the citation prohibition above stands in full — a better design on the same proxy is still the same proxy.

Filed by: Stafford (claude-opus-5), daily brief thread chase, 2026-10-03

Tags: paper-2, practitioner, Momentum Mirage, Capability, COUNTERARGUMENT Source: Michael Ewens & Xavier Giroud (Columbia Business School / NBER), author copy of the September 2026 revision; NBER Working Paper 34162 URL: http://www.columbia.edu/~xg2285/Hierarchy.pdf Date published: null — the author copy is dated September 2026 with no day; NBER's landing page lists issue date August 2025, revision date September 2026. No exact date is stated, so none is invented (hard rule 1). Date flagged: 2026-10-02

Summary

A third version of this paper exists and this base has now held two superseded ones. On 2026-09-25 Stafford reconciled the August 2025 original against an October 2025 revision and recorded the result as hard rule 14. The paper has since been revised again. The author copy at the Columbia URL is now dated September 2026, and NBER's landing page serves its abstract with revision date September 2026. Downloaded and text-extracted in full on 2026-10-02 (4,685 lines, 65 pp).

This is not a cosmetic revision — the measure itself was rebuilt, and the headline descriptive statistic moved by a factor of nearly two. Abstract, verbatim: "We introduce a novel measure of corporate hierarchy for 3,454 U.S. public firms. We construct this measure using online resumes of 18.1 million employees and a network estimation technique that allows us to identify hierarchical layers... Firms have on average six hierarchical layers and a pyramidal organizational structure." The prior two versions reported ten average layers (2,500+ firms / 16M résumés in August 2025; 3,100+ firms / 7M résumés in October 2025). The data provenance is now stated explicitly and differently: résumés come from two named vendors, Revelio and CoreSignal (3,050 firms from Revelio, 404 from CoreSignal), and layers are estimated with the network methodology of Huitfeldt, Kostøl, Nimczik & Weber (2023).

The observation window is no longer pre-generative. The paper states: "online resume data from LinkedIn for 18.1 million workers at U.S. public firms from 2015 to 2024." This base's Résumé-Panel Consensus pattern records this instrument's window as 2010–2018. It now runs through 2024.

The AI result is renumbered, larger, and better powered — and the "10% only" caveat this base has carried for seven weeks is gone from the paper. The AI regression is now Table 10, reporting 6,386 firm-year observations across 1,729 firms (was 4,606/1,250, then 5,505/1,491). Coefficients on log(number of layers): −0.42***, −0.36**, −0.38***, −0.072*** — three at the 1% level and one at 5%, where both prior versions had all four starred at 10% only. The phrases "under-powered" and "significant at the 10% level regardless of the metric" occur zero times in the September 2026 text. The prose now reads simply: "we find evidence for a significant reduction in the number of layers in all four columns." Mean dependent variable is 1.83 (was 2.44, then 2.41), consistent with the move from ten layers to six. Within R² 0.02.

Two disclosures in the new version are worth more to Brandon than the stronger result, and they cut the other way. First, the aggregate trend: "the number of layers has been flat to weakly increasing over the sample period" (2015–2024). Second, the identification base, stated by the authors in a new Section 6.1.1 on within-firm variation: *"During our sample period from 2015–2024, only 19% of firms experienced a change in layers, and this occurred in 6.4% of firm-years. When layers do change, they increase 66% of the time, and the median change is one layer... the within-firm designs later in the paper (the IPO and AI analyses) are identified off the 654 firms whose layer count changes during the sample period."

The shared-right-hand-side problem is unchanged and has acquired a second layer. Table 10's note reads: "The right-hand side variables are various measures of AI adoption based on job posting data, obtained from Babina, Fedyk, He, and Hodson (2024), and each enters lagged by one year." Hard rule 9 holds verbatim. New in this version: footnote 30 states "The evidence in Table 10 is related to the finding of Babina, Fedyk, He, and Hodson (2025) that firms with more AI postings tend to have a higher share of employees in junior vs. senior positions. To the extent that flatter hierarchies have a higher share of junior positions, their finding is consistent with our evidence." The authors now cite Babina as corroboration of a result whose explanatory variable they took from Babina.

Also changed: the COVID finding has lost its sector entirely — "biotech" (Aug 2025) and "pharmaceutical" (Oct 2025) both occur zero times; the result is now stated as flatter firms experiencing a lower stock-market drop, framed as organizational agility. Still absent in all three versions: any span-of-control measure, and any outcome variable linking flattening to subsequent firm performance.

Five Breakpoints Intersection

• Momentum Mirage: Three versions and fifteen months later, the base's flagship flattening instrument still measures the structural artifact of transformation and nothing downstream of it — there is no outcome variable anywhere in the specification linking fewer layers to whether the firm subsequently performed, moved or delivered. The revision enlarges the panel to 3,454 firms and strengthens every AI coefficient while leaving that absence untouched, which is the appearance of accumulating evidence about transformation without any measurement of movement.
• Technology Illusion: The authors' own aggregate statement — layers "flat to weakly increasing" across 2015–2024, with layer changes rising two times out of three when they happen at all — sits inside the single paper most cited for the proposition that AI is flattening organizations. The flattening is a conditional effect among adopters identified off 654 firms, not a trend in the economy, and the gap between those two readings is where a technology narrative is doing work the data does not support.

Four Forces Intersection

• Capability: The authors' mechanism is unchanged across all three versions and remains a Capability claim in the framework's exact sense — layers fall because cheaper knowledge acquisition reduces the need for problem-solver layers, making structure a function of where knowledge sits rather than of intent or follow-through. No Purpose, Commitment or Momentum term enters the specification.

Relevance to Brandon's Positioning

Four concrete consequences, and the first requires editing two drafts before either ships.

(1) A load-bearing sentence in two of Brandon's drafts is now false. drafts/the-flattening-bet.md states that "the flattening estimates in the most-cited paper are significant only at the 10 percent level, a limitation its authors state plainly and its citers rarely repeat," and drafts/four-papers-one-method.md carries the same claim in its citation block. In the current paper three of four coefficients are significant at the 1% level and the authors' under-powered language is gone. The thinness leg the essay leaned on hardest has been removed by the authors. Corrected in place today.

(2) The replacement legs are better, and they come from the authors rather than from Brandon's reading of them. The essay's thesis is that the great-flattening narrative is overbuilt relative to its instruments. It can now say, in the paper's own words, that across 3,454 U.S. public firms from 2015 to 2024 the number of hierarchical layers was "flat to weakly increasing"; that only 19% of firms ever changed layer count and only 6.4% of firm-years saw a change; that when layers did change they went up two times out of three; and that the entire AI finding is identified off 654 firms. That is a far stronger argument than a significance level, it survives the authors strengthening their own result, and it is quotable from the primary.

(3) The Résumé-Panel Consensus pattern loses one of its two defining properties for this member. The pattern's case is that the firm-level AI-and-structure literature is one data genre with pre-generative observation windows. For Ewens & Giroud the window is now 2015–2024, inside the generative era, and the vendors are named (Revelio, CoreSignal). The data-genre half of the pattern is untouched — still résumé-inferred structure, still Babina's postings-based AI variable on the right-hand side — and the shared-RHS rule is now doubly earned, since the authors cite Babina (2025) as consistent with a result built on Babina (2024)'s explanatory variable. The pattern should be amended rather than weakened: its charge is the shared instrument genre, not the dating.

(4) The citation-hygiene rule needs a third entry and a standing route. Three versions of one working paper in fifteen months, each with different headline figures, two of which this base has at some point asserted as current. The practical rule: for any NBER working paper, read the author's own copy and the landing-page revision date on every re-check, never the /revisions/ path, which served the August 2025 text as late as 2026-10-02.

Sources & Confidence

• Ewens M, Giroud X. "Corporate Hierarchy." Author copy dated September 2026, columbia.edu/~xg2285/Hierarchy.pdf, 65 pp, downloaded and text-extracted in full on 2026-10-02 (4,685 lines). Abstract, measure construction (§2 and the LinkedIn data section), §6.1.1 within-firm variation, the AI section, Table 10 with its note and footnote 30, and the Conclusion all read directly. — High (primary, author-hosted), with the standing caveat that a revised working paper remains un-peer-reviewed.
• NBER landing page nber.org/papers/w34162, fetched 2026-10-02: abstract matches the September 2026 author copy (3,454 firms, 18.1 million résumés, six layers); issue date August 2025, revision date September 2026. — High.
• nber.org/system/files/working_papers/w34162/revisions/w34162.rev0.pdf — not re-fetched today. The research-evidence scan fetched it on 2026-10-02 and reported the August 2025 cover date and the 2,500-firm / 16-million-résumé abstract, consistent with Stafford's 2026-09-25 finding that this path serves the original. Treated as confirmation that the path is stale, not as a version of the paper. — High on the staleness, from two independent fetches six days apart.
• Prior-version reconciliation: knowledge-base/ewens-giroud-corporate-hierarchy-october-2025-revision-2025-10.md (Stafford, 2026-09-25) — full side-by-side of the August 2025 and October 2025 texts. — High.
• Babina T, Fedyk A, He A, Hodson J (2024), Journal of Financial Economics 151, 103745 — origin of Table 10's AI-adoption measures. Not read directly. — Unknown pending a read.
• Babina, Fedyk, He & Hodson (2025), cited in footnote 30 on junior-versus-senior employment shares. Not read, not located. — Unknown.
• Huitfeldt, Kostøl, Nimczik & Weber (2023) — the network estimation methodology the layer measure is built on. Not read. — Unknown.

Filed by / Routing

Filed by: Stafford (claude-opus-5), instrument verification / exploratory slot, 2026-10-02 Routing: strategy

Tags: paper-2, Technology Illusion, Capability, COUNTERARGUMENT Source: Harvard Business Review; KPMG LLP newsroom; UT Austin McCombs News URL: https://hbr.org/2026/07/research-why-some-junior-employees-work-well-with-ai-and-others-dont Date published: 2026-07-23 Date flagged: 2026-10-01

Summary

Ashish Agarwal, Anitesh Barua, Wen Wen and Fangchen Song (McCombs School of Business, UT Austin) with Anu Puvvada (KPMG LLP) ran a field study of 523 US-based early-career professionals at KPMG LLP, each with under 18 months' tenure. Participants completed work assignments with an AI agent, the tasks mirroring real client work in their own focus area. The researchers first had the agents complete the same work with no human involvement to establish an AI-only baseline, then measured each professional's collaborative output against it.

The result is a three-way split. AI Amplifiers, 50.1%, beat the AI baseline. AI Delegators, 25.8%, produced results comparable to the AI working alone. AI Apprentices, 24.1%, performed below the AI baseline — that is, a quarter of participants made the output worse than if they had not been involved.

The finding the authors emphasise is that the split was not predicted by individual capability. In their words, "traditional measures of capability didn't explain the gap. Those who outperformed AI looked nearly identical on paper to those who didn't." More pointedly: AI Apprentices "matched Amplifiers and scored higher than Delegators on every foundational skill, yet they landed below the AI baseline." Critical thinking, AI literacy and domain knowledge did not sort people into the three profiles. What distinguished Amplifiers was behaviour during the work: they were "orchestrating the workflow, framing problems to guide the AI, anchoring the work in real domain frameworks, and refining results across multiple rounds."

Five Breakpoints Intersection

Technology Illusion: a quarter of users (24.1%) produced work worse than the tool would have produced unaided, and no measure of their skill predicted that they would. Any adoption metric — seats, logins, usage — counts those 24.1% as successful deployment. This is the breakpoint's definition with a number attached: the technology was deployed, adoption was real, and whether it added or subtracted value was decided after arrival by something nobody was measuring.

No Strategic Disconnection or Incentive Fragmentation mapping. The study is a task-level field experiment inside one employer with no strategy, incentive or structural variable in it, and tagging either would be the speculative tagging this base's rule 4 forbids.

Four Forces Intersection

Capability: the study measures Capability directly and splits it in two. The trait version — what someone knows, scored on foundational skills — did not predict performance at all. The practice version — how work is framed, directed, evaluated and iterated — did. For a framework that treats Capability as "skills, infrastructure, and org capacity to execute," this is a finding about which half of that definition is load-bearing.

Basis for these tags: the KPMG newsroom release and the McCombs News research summary were both read in full; both are first-party accounts by the sponsor and the research institution. The HBR article itself is paywalled and has not been read. Forces tagged from the readable first-party sources, not from the paywalled text.

Relevance to Brandon's Positioning

This is the chase target the 2026-10-01 scan logged and did not run, and running it reverses what the scan expected to find. The scan recorded "KPMG / University of Texas 'AI amplifiers'" as a chase target heard second-hand twice (09-21, 09-29), and flagged that if real it would be "a third instrument putting the AI moderator at individual evaluative skill, alongside Sarkar and Idan & Anand — the strongest live rival to paper-2's organizational thesis."

It is not a third member of that rival. It is a replication of one arm of it against the other arm, and the two arms disagree. Sarkar (SSRN 5713646) finds the AI gain rising with workforce experience (+52% vs +23%). Idan & Anand (arXiv 2605.18143) find GPA predicts nothing (p=0.59) and prior domain knowledge predicts nothing (p=0.2), while an AI Interaction Competence measure predicts the outcome. KPMG × McCombs lands squarely on the Idan & Anand side, on working professionals, n=523, on real client-type work: foundational skill flat across profiles, interaction behaviour predictive. The person-level-moderator challenge in knowledge/thesis-challenges.md has been carrying Sarkar's expertise arm and Idan & Anand's competence arm as a single converging objection. They are not converging; a third instrument now contests the expertise arm.

What it does not do is support Brandon. The surviving mechanism is still person-level — a behaviour rather than a trait. Its value to paper 2 is that a behaviour is trainable, specifiable and assignable, and Idan & Anand already showed a supplied written scaffold substituting for it (mean .38 → .45, spread SD .22 → .14, benefit falling as individual skill rose). The one person-level mechanism that replicates is the one already demonstrated replaceable by a design artifact. That is an argument the paper can make; it is not yet a measurement anyone has run on professionals against an organizational outcome.

Sources & Confidence

• Agarwal, Barua, Wen, Song & Puvvada, Research: Why Some Junior Employees Work Well with AI — and Others Don't, Harvard Business Review, July 2026 — High publisher (HBR is on the high-credibility list), but HBR is editorial review, not peer review, and the article itself is paywalled and unread here.
• KPMG LLP newsroom, Shaping Early-Career Success in the Age of AI, 2026-07-23 — High publisher, interested party. KPMG sponsored the study, supplied the entire sample from its own staff, and a KPMG Studio Leader is a named co-author.
• UT Austin McCombs News, Shaping Early-Career Success in the Age of AI, 2026-07-23 — High.
• Three binding method cautions. (1) One employer. Every participant is a KPMG professional under 18 months' tenure, on KPMG tooling and training. Organizational variance is zero by construction, so this study cannot adjudicate organizational against person-level explanations in either direction — the same objection logged 2026-09-26 as The within-firm estimate cannot carry an organization-level claim. (2) The three-profile split is a function of the baseline. "Below the AI baseline" means below a performance level the researchers produced by asking the agents to work unaided; how hard the agent was made to try sets the 24.1%, and no one outside the study can check it. (3) Early-career only — nothing here speaks to senior professionals, which is precisely the population Sarkar's experience gradient is about, so the contest between the two instruments is not a clean head-to-head.
• Not chased, and not to be cited: a separate "1.4 million interactions" KPMG × UT analysis surfaced in search results alongside this study. It is an earlier, different collaboration and has not been verified. Do not cite it as part of this finding.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-10-01) — chase of the research-evidence scan's 2026-10-01 named chase target Routing: strategy

Tags: paper-2, Incentive Fragmentation, Technology Illusion, Commitment, Capability Source: arXiv (Ummara Mumtaz, Rabi Noor, Summaya Mumtaz) URL: https://arxiv.org/abs/2605.16281 Date published: 2026-04-10 Date flagged: 2026-10-01

Summary

The authors take a snapshot of the AI Incident Database (AIID) containing 1,365 incident records, restrict to incidents dated 2020–2026, and screen down to a final analytic sample of 480 incidents (regex-assisted screening to 1,123; conceptual screening excluding malicious-misuse and pure-misinformation cases to 568; manual removal of scams, impersonation and copyright/IP disputes to 480). Each incident is then coded against nine post-deployment provisions drawn from the EU AI Act, the NIST AI Risk Management Framework and the GDPR, using automated content analysis over incident report texts and structured metadata. Every provision receives one of four classifications — compliant, partially compliant, non-compliant, or insufficient evidence — and compliance is assessed only for the subset of incidents to which a provision is applicable.

The headline results are gaps. 77.1% of incidents show no evidence of EU AI Act post-market monitoring and 99.6% show no documented Data-Protection Impact Assessment. Mapped to NIST GOVERN 4, 91.5% of incidents show accountability structures absent. 9.8% are simultaneously non-compliant under two or more regulatory regimes. The authors also split by detection channel: 24 of 480 incidents (5.0%) were detected through internal monitoring; 456 (95.0%) were identified externally — by journalists, users, researchers, regulators, or through legal proceedings. Incidents detected internally code far more compliant than externally detected ones (87.5% vs 5.3% under the EU AI Act; 95.8% vs 58.1% under NIST), from which the authors conclude that "monitoring capacity is a key condition for effective post-deployment governance." They then propose a four-phase Proactive AI Governance Compliance Framework (PAGCF).

The authors state their own principal limitation plainly: "The AIID captures publicly documented AI failures and therefore probably underrepresents incidents that are detected and resolved internally before becoming public," which "biases our sample toward externally detected cases." They further note that the internally detected subset "may reflect broader organisational capacity rather than monitoring alone — a classic selection-on-observables concern."

Five Breakpoints Intersection

Incentive Fragmentation: 91.5% of 480 publicly documented AI incidents show no accountability structure present under NIST GOVERN 4 — in nine cases out of ten, when an AI system caused a documented harm, no identifiable party's responsibility for it was on the record. This is the outcome-side companion to the Gravitee survey's input-side 7.2% named-accountability prevalence (research-evidence/knowledge-base/, staged 2026-10-01): one instrument asks organizations whether anyone owns the agent, the other reads what was documented after something went wrong, and they return the same shape through completely different data genres.

Technology Illusion: 77.1% of the incidents show no evidence of post-market monitoring, the specific obligation that exists because a deployed AI system's behaviour cannot be established at deployment time — the systems were put into production on the assumption that the organizational condition for running them safely was in place, and in three cases out of four nothing in the record shows it was.

Not tagged Momentum Mirage, deliberately. The 5.0% internal-detection figure is the most quotable number here and it cannot carry that tag — or any population claim — because the authors' own stated selection bias means internally detected and internally resolved incidents largely never enter this register at all. See Sources & Confidence.

Four Forces Intersection

Commitment: accountability structures absent in 91.5% of incidents is a direct measure of the absence of any party whose follow-through was on the hook when the system failed — the Force this base has consistently found hardest to observe from outside, measured here on the failure side rather than by self-report.

Capability: 77.1% without post-market monitoring evidence is an absence of the organizational capacity to observe a deployed system at all, independent of anyone's intent to govern it.

Relevance to Brandon's Positioning

This is the first instrument in the base that measures the accountability gap from the record of failures rather than by asking organizations about themselves, and that is its whole value. Every governance-gap number this base holds — Schellman, Workiva, EMA/Cequence, Grant Thornton, Gravitee — is a survey in which the same respondent supplies both the confidence and the substrate. This one is coded from third-party incident documentation by authors with nothing to sell. It therefore speaks directly to the Confidence Without Substrate pattern's four-times-unmet promotion condition (a disinterested publisher, or a substrate verified independently rather than asked about), but it does not meet it: the pattern requires confidence and substrate measured on the same units, and this instrument has no confidence side and no organizational denominator. It supplies an independently verified substrate for a different population. That distinction should be stated when it is cited, not glossed.

It is also the strongest available fuel for asset 16 ("Borrow the Auditor's Questions"), which has so far been built entirely on what a standard requires. 91.5% is what the absence costs.

Sources & Confidence

• Mumtaz, Noor & Mumtaz, Post-Deployment Accountability in AI Governance, arXiv:2605.16281 (submitted 2026-04-10; v2 2026-07-14; v3 2026-08-10) — Moderate. Unrefereed preprint; arXiv is a peer-track venue but this is not NBER/SSRN economics and carries no institutional affiliation on the listing page. The coding is automated content analysis over incident text, so every figure is a measure of what the public record documents, never of what organizations did. Three revisions in four months: per hard rule 14, any figure here that later disagrees with a cited version is a revision until proven otherwise — v2 was read for this entry and the abstract figures are v3-current.
• Binding caution for any citation. The 5.0% internal-detection figure describes the composition of a public incident register, not a detection rate in any population of organizations. The authors say so. This base has a documented laundering history with exactly this kind of number (the 2026-08-29 ledger traces aggregators converting a "confirmed or suspected" rate into an "88.4% breach rate"). Cite 91.5% and 77.1%, which are properties of the coded sample and survive the selection bias as statements about documented incidents. Treat 87.5% vs 5.3% as uninterpretable for causal purposes: detection channel and documentation evidence are read off the same source text, so an incident caught internally is by construction an incident at an organization that wrote things down.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-10-01) — exploratory query, Stafford's own slot Routing: governance

Tags: paper-2, Technology Illusion, Strategic Disconnection, COUNTERARGUMENT Source: ISS-STOXX (Institutional Shareholder Services), U.S. Compensation & Governance Advisory URL: https://www.iss-stoxx.com/insights/articles/artificial-intelligence-and-governance-is-2026-a-tipping-point-for-turning-awareness-into-action/ Date published: 2026-05-25 Date flagged: 2026-09-30

Summary

ISS's own governance advisory arm, writing in May 2026 on disclosure data through 2025, reports that 22% of S&P 500 companies disclosed board oversight of AI and 24% disclosed frameworks and policies around AI, with the Russell 3000 at 6% on both. Authors Alyce Lomax and Danielle Rizak state that "the lion's share of both S&P 500 and Russell 3000 companies have no directors with clearly disclosed artificial intelligence skills," and, separately, that "companies are much more likely to have directors with AI experience than they are to disclose oversight at the board level."

Those two sentences are the reason this entry exists. Six weeks later, an ISS dataset "compiled in June 2026" was analysed by the Morgan Stanley Institute for Sustainable Investing and reported that 68% of S&P 500 companies have at least one non-executive director with AI skills, with almost 700 such directors, around one-fifth of the S&P 500 total, "identified by ISS as having AI skills." ISS's own earlier disclosure-based counts of the same object — the share of S&P 500 companies with at least one AI-expert director — were 13% (DEF 14As filed September 2022–September 2023, published 2024-03-21) and 20% (2024 data, published 2025-03-19).

68% and "the lion's share have none" cannot both be disclosure counts of the same population six weeks apart. They are not. ISS publishes at least two distinct measures of board AI expertise: a disclosure count, which it reports at 13% → 20% and describes in May 2026 as still leaving most companies at zero; and an identification field in its board-composition dataset, which reports 68%. The second is ISS's own classification of directors, not a transcription of what companies claim about them. ISS says so obliquely in the sentence above and nowhere publishes the criteria.

Five Breakpoints Intersection

Technology Illusion: The instrument the market uses to price board AI competence manufactures a large part of the competence it prices — ISS identifies 68% of S&P 500 boards as holding AI skills while its own disclosure series, run on the same index, reports most companies disclosing none, so a governance signal that moves proxy votes and director recruitment rests on an unpublished vendor taxonomy rather than on anything the company or the director has asserted.

Strategic Disconnection: ISS states that "companies are much more likely to have directors with AI experience than they are to disclose oversight at the board level" — a gap running in the opposite direction to almost every other confidence construct in this base, where the claim exceeds the substrate; here the substrate that a vendor can find exceeds anything the board has committed to writing, so shareholders, directors and the vendor hold three materially different pictures of the same boardroom.

Four Forces Intersection

No direct Force implication — this is an instrument-verification entry about how a governance variable is measured, and the article estimates no outcome against any Force. Recording a Force here would be the guessed-tag failure this base exists to avoid.

Relevance to Brandon's Positioning

This runs the instrument check that the 2026-09-30 research-evidence scan named as load-bearing and scheduled for 2026-11-15, and it comes back adverse to the scan's own strongest new claim. The scan opened a construct it called the reclassified credential — board AI competence is supplied by relabelling existing technologists rather than by acquiring capability — resting on Morgan Stanley's finding that 62% of AI-skilled non-executive directors were appointed in 2022 or earlier, against 63% of the broader board population, and on Morgan Stanley's own inference that companies "do not necessarily require this experience to view a non-executive director as qualified in this area." That inference names companies as the viewer, and the qualification was assigned by ISS. Most of those companies, on ISS's own May 2026 reading, have disclosed nothing about any director's AI skills. No nominating committee was asked and none is on record viewing anyone as anything.

The 62%≈63% similarity is additionally close to a mechanical result rather than a finding: classify sitting directors by biography and the classified subset inherits the population's tenure distribution unless the classifier happens to correlate with tenure. The absence of a difference here is the null, not the result. Any public use of the appointment-vintage line — it is the most quotable sentence in a report governance press reads — is one competent reader away from being taken apart.

What survives, and it is enough to carry a piece. (1) The 24% / 96% / 80% ratio is untouched: three skill categories, one publisher, one taxonomy, one month, definition held fixed — board depth in AI is a quarter of the depth those same boards hold in ordinary technology. (2) The energy-sector zero is untouched as a fact about where the credential appears, though it now describes where ISS finds AI-relevant biography rather than where boards recruited. (3) The relabelling claim survives with its subject changed, and is more interesting for the change: the relabelling is being performed by the proxy-advisory firm whose ratings move votes, not by nominating committees — which is the Technology Illusion applied to the measurement apparatus rather than to a board.

One correction this entry forces on a claim the scan made the same morning. The scan concluded that on board AI competence the self-report and the coded count agree, and that directors are therefore not overconfident. The ISS instrument reports 68% of companies holding at least one AI-skilled director; PwC's 599 directors report 71% saying their boards need to improve their AI skills. The coded count is more generous than the directors' own assessment, not harsher. The scan reached its conclusion by quoting the 24%-with-three-or-more figure and not the 68%-with-one figure, which are the same instrument. Which of the two you quote decides the story, and that is a property of the instrument worth stating whenever it is cited.

Sources & Confidence

• ISS-STOXX, Artificial Intelligence and Governance: Is 2026 a Tipping Point for Turning Awareness into Action?, Lomax & Rizak, 2026-05-25 — the 22% / 24% / 6% figures and both verbatim sentences. Moderate (first-party from the data vendor; not on CLAUDE.md's high-credibility list, and ISS has a commercial interest in the salience of board-composition data).
• ISS-Corporate press release, Roughly 15 Percent of Large U.S. Companies Disclose Board Oversight of AI, 2024-03-21 — 15% board oversight, 13% with at least one director with AI expertise, DEF 14As filed 2022-09 to 2023-09. Moderate.
• ISS-Corporate press release, Roughly One-Third of Large U.S. Companies Now Disclose Board Oversight of AI, 2025-03-19 — 31.6% board oversight as of 2024, 20% disclosing at least one director with AI expertise, 11% explicit full-board or committee oversight, 19 AI shareholder proposals in 2024. Moderate.
• Morgan Stanley Institute for Sustainable Investing, How Boards and Executives Are Governing the Rise of AI, 2026-08-27 — the 68%, 24%, 96%, 80%, ~700 / one-fifth, 62%/63% and committee-distribution figures, retrieved by direct curl (morganstanley.com returns HTTP 503 to the fetch tool) and quoted verbatim from the page text. Moderate. Already held in research-evidence as of 2026-09-30.
• The Conference Board, Governing AI in the Corporation, Jones & Campbell, 2026-04-22 — 2.7% of S&P 500 directors disclosed AI expertise as of December 2025 (1.5% in 2021), 51% technology, 27% cybersecurity. High. Already held: research-evidence knowledge-base/conference-board-governing-ai-in-the-corporation-2026-04.md. Cited here as directionally consistent only — a share of directors (TCB) and a share of companies (ISS) are different objects, and that KB file's own cross-instrument caution forbids mixing the two publishers' series in one sentence. The load-bearing comparison in this entry is ISS-internal and needs no cross-publisher step.

Filed by / Routing

Filed by: Stafford (stafford-research, Wednesday falsification sweep, 2026-09-30) Routing: governance / market — run before any public use of the Morgan Stanley board figures, and before the 2027 proxy-season window is written.

Tags: paper-2, practitioner, Technology Illusion, Capability Source: Journal of Robotic Surgery — Tsuyoshi Uchida et al. (Yamanashi University, Japan). Read alongside Linsky, P. & Wei, B., "Training in robotic thoracic surgery" (review), as the training-structure source. URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC12075338/ Date published: 2025-05-13 Date flagged: 2026-09-29

Summary

A single-center retrospective analysis at Yamanashi University of 154 lung resections performed between November 2018 and May 2024 by four surgeons at varying experience levels, using cumulative sum (CUSUM) analysis to locate the point at which operative performance stabilizes. Reported operational metrics: mean operating time 206.5 minutes, mean console time 153.3 minutes, mean blood loss 23.9 g, complication rate 8.4% (13 patients, none converted to thoracotomy). The lead surgeon showed "a typical upward convex learning curve"; the three surgeons who joined the program subsequently showed "smaller peaks." The paper's institution-level conclusion is that "institutional proficiency did not significantly decline even with new surgeons."

The second source is a training review (Linsky & Wei) that describes how robotic thoracic programs are actually structured. Two features are named: the da Vinci dual console, which lets an instructor "have the same view as the trainee, but essentially have the same hands," permitting intervention without repositioning; and Cerfolio's stepwise protocol, tabulated as a graduated allocation in which trainees progressively perform more of the operation — beginning with trocar placement and advancing through dissection, vessel identification and resection — each step with an allocated timeframe.

The review also compiles cases-to-proficiency figures from the prior literature: robotic (RATS) proficiency "seems to be around 20 cases," with Jang et al. reporting as few as 6; against VATS at "30 cases for safety and 90 for efficiency," and a second estimate of "100–200 for safety and efficiency and more for excellence." These are review-level compilations of other groups' estimates, not primary estimates by either source, and they are not a within-design comparison.

Five Breakpoints Intersection

Technology Illusion: — as the control arm rather than as an instance. The breakpoint holds that technology deployed onto untouched organizational conditions fails; this literature documents the opposite configuration and is useful for exactly that reason. Robotic thoracic surgery was adopted together with a rebuilt instructional structure — dual-console co-presence and a stepwise protocol allocating named operative steps to the trainee — and the field's reported time-to-proficiency is lower than for the predecessor technology, with institutional performance holding as new operators entered. It is evidence about the mechanism's sign, not a case of the breakpoint firing.

Four Forces Intersection

Capability. Uchida's institutional CUSUM is a direct measurement of capability replenishment rather than capability level: it tracks whether an organization's performance degrades as unpracticed operators are absorbed, and reports that it did not. This is the measurement design that the Capability force has lacked in this base, which has until now held Capability almost entirely as a stock-level question.

Relevance to Brandon's Positioning

This does not advance a paper claim as evidence. Its value is methodological and argumentative, and it bears on the learning-rate counterargument that Friebel et al. created against paper-2 Claim 1 on 2026-09-29 (research-evidence scan). Friebel's model makes the rate at which juniors qualify as seniors the only shock that permanently reshapes a hierarchy, and treats that rate as an exogenous technology parameter — which, if true, is a channel running outside all four Forces. Surgical education is the field that has measured time-to-competence as a hard number for two decades, in cases-to-proficiency and CUSUM; it supplies both the instrument design the learning-rate story thread is missing and a reason to doubt the exogeneity assumption, since the structure that the field itself credits for the compression is an instructional protocol, not a device.

Two things it also supplies for asset 2 (The Flattening Bet) and for the learning-rate thread: an organization-level capability-replenishment measurement that exists and is routine in the wrong field, and a contrast with enterprise behaviour — Cerfolio's ladder allocates operative steps to the trainee while the technology makes supervision cheaper, whereas the enterprise pattern recorded in this base (ZipRecruiter 2026: 38% of hiring managers moved basic data processing off entry-level workers onto AI; 31% raised entry-level experience requirements) allocates the steps to the machine and removes them from the junior.

Sources & Confidence

• Uchida, T. et al., "Institutional proficiency and learning curves in robotic-assisted thoracoscopic surgery," Journal of Robotic Surgery, published 2025-05-13. Design, sample, window, metrics and the institutional-proficiency conclusion read from the PMC record (PMC12075338). — Moderate. Peer-reviewed, but single-center, retrospective, n=154, four surgeons, no comparison arm.
• Linsky, P. & Wei, B., "Training in robotic thoracic surgery" (PMC5803121), review, 2018. Dual-console description, Cerfolio stepwise protocol (their Table 2), and the compiled cases-to-proficiency figures. — Moderate. A training review, not an instrument; the case numbers are compilations of third-party estimates.

Instrument discipline (recorded before the finding, per CLAUDE.md rule 9). Data source: operative records at one Japanese university hospital (Uchida); narrative review of training practice (Linsky & Wei). Inference method: CUSUM change-point detection on operative performance, by surgeon and at institution level. Observation window: November 2018 – May 2024, entirely pre-generative for the technology in question. Binding bounds, all of which must travel with any use of this entry: (1) neither source estimates training structure against the curve — the attribution of compressed proficiency to the instructional protocol is the review's framing and nobody's coefficient; (2) the RATS-versus-VATS case numbers come from different studies, eras and case mixes and are not a within-design comparison; (3) surgeons who enter robotic programs are self-selected; (4) teleoperation is not delegation — the RATS trainee performs every motion, which is the opposite of the agentic case, so this is an analogy about instructional structure and never a measurement of AI's effect on anything.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-29) — exploratory slot, executing the method instruction handed forward by the 2026-09-29 research-evidence scan ("query the field that owns the measurement, not the field that owns the question"). Routing: strategy / personal

Tags: paper-2, practitioner, Incentive Fragmentation, Technology Illusion, Commitment, COUNTERARGUMENT Source: Artificial Intelligence Underwriting Company (AIUC) — AIUC-1 standard documentation site URL: https://standard.aiuc-1.com/accountability (requirement text); https://standard.aiuc-1.com/aiuc-1-certification (eligibility and scope); https://aiuc-1.com/ (domain index) Date published: null Date flagged: 2026-09-28

Summary

The 2026-09-28 research-evidence scan ingested AIUC-1 from its landing page and recorded, four separate times, that the per-domain requirement pages remained unopened — noting that the Accountability domain's literal requirement text "is the one remaining thing that could further narrow this reading" and "should be read before any of this is said in public." This entry is that read. Three pages were fetched first-hand: the domain index at aiuc-1.com, the Accountability requirement page at standard.aiuc-1.com/accountability, and the certification-eligibility page at standard.aiuc-1.com/aiuc-1-certification.

Domain E runs E001–E017, with E007 and E014 marked retired (merged into E004 and E017 respectively) and E013 marked supplemental (a quality-management-system requirement). The first four requirements are the material ones and are quoted verbatim. E001: "Document AI failure plan for AI privacy and security breaches assigning accountable owners and establishing notification and remediation with third-party support." E002: "Document AI failure plan for harmful AI outputs that cause significant customer harm assigning accountable owners and establishing remediation with third-party support." E003: "Document AI failure plan for hallucinated AI outputs that cause substantial customer financial loss assigning accountable owners and establishing remediation." E004: "Document which AI system changes across the development & deployment lifecycle require formal review or approval, assign a lead accountable for each, and document their approval." Further requirements cover data-storage practices (E005), upstream-vendor due diligence (E006), internal process reviews (E008), third-party API logging (E009), an AI acceptable use policy (E010), processing locations (E011), applicable law (E012), audit logging for incident investigation (E015), user disclosure (E016) and a transparency policy with model cards (E017).

The eligibility page states, verbatim: "Organizations should pursue AIUC-1 certification if they are: 1. Adopting or developing agentic AI systems built on generative models 2. Applying agentic AI in high-risk use cases, such as customer-facing agents where brand is on the line, agents with access to confidential data, agents handling critical workflows 3. Looking to build trust and demonstrate AI security to potential enterprise customers, internal executives or board members." It adds that "Certified organizations range from seed stage to publicly traded enterprises," and — the sentence that governs everything else here — that certification scope is "Typically limited to specific product(s), not the full organization's products and practices."

Five Breakpoints Intersection

Incentive Fragmentation: The standard requires a named accountable owner per failure class (E001–E003) and a named approval lead per change class (E004), then scopes the certificate to "specific product(s), not the full organization's products and practices" — so the accountability it audits attaches to a product's failure modes rather than to the organization's agent estate, which is precisely the boundary at which ownership of an agent's consequences stops being anybody's scorecard item.

Technology Illusion: An organization can hold an audited, quarterly-retested certificate attesting that accountable owners are documented for three failure classes of one certified product while having no such document for any of the other agents it runs — the assurance artifact is real and its coverage is narrower than the thing a reader assumes it covers.

Four Forces Intersection

Commitment: E001–E004 are the most precise operational specification of follow-through this base holds — they require that the answer to "who is responsible when this fails" be written down, per failure class, before the failure, and re-evidenced to an auditor every quarter; the observation is that this specification exists, is auditable, and is applied at product scope.

Relevance to Brandon's Positioning

This entry corrects two factual claims in the scan's own 2026-09-28 case-file and pattern text and re-specifies a third, and it does so with the standard's own words. The scan wrote that "no requirement visible on the published scope names a role inside the deploying customer organization that must own a deployed agent" — E001–E004 name owners, and E004 names them explicitly "across the development & deployment lifecycle." The scan wrote that the deploying enterprise "is not a party that is audited" — the eligibility text's first criterion is "Adopting or developing," and its third names "internal executives or board members" as the audience, which is internal governance rather than procurement. What survives, and survives more strongly than the version it replaces, is a scope claim: the regime can specify and audit named accountability, and it sells that audit at product scope.

The consequence for Claim 3 is direct and it cuts the other way from the scan's framing. The scan's OPEN QUESTION asked whether organizational accountability goes unassured because it is unmeasurable or because it is uninsurable, and set the discriminating test at the Q4 2026 refresh in January. The requirement text answers it now: the instrument is written, numbered and audited today. The unmeasurable branch is dead. What has not happened is anyone pointing the instrument at an organization's whole agent estate rather than at one shipped product — a scoping and demand question, not a measurement question. That makes Brandon's lead time shorter than the January framing implies, and it hands him a concrete, borrowed, unarguable diagnostic: four numbered requirements from an insurer-funded standard that a client can be asked to produce for its own agents.

Sources & Confidence

• AIUC-1 Accountability requirement page, standard.aiuc-1.com/accountability — Moderate (primary document read first-hand; AIUC is a venture-funded standards-and-insurance startup, not on the high-credibility publisher list, and it sells both the certificate and the insurance)
• AIUC-1 certification-eligibility page, standard.aiuc-1.com/aiuc-1-certification — Moderate (same; eligibility and scope sentences quoted verbatim)
• AIUC-1 domain index, aiuc-1.com — Moderate (domain URLs and Accountability scope sentence)
• published_date stored null: the standard is versioned by quarter only, and the landing page's "Last Updated" timestamp equals the ingest date — the exact shape of Known Failure Mode #1. Refused rather than recorded.
• Not verified and not asserted anywhere above: the widely repeated "51 requirements, 130 controls" count and the ElevenLabs first-policyholder claim, both of which the scan also failed to confirm on AIUC's own pages. The E-series count here is what the Accountability page itself serves.
• Read but not quoted at requirement level: domains A–D and F. The standard.aiuc-1.com/evidence page served only A001–A008 and B001–B004 before truncating, so the evidence specifications for E001–E004 were not obtained and it is not established what an auditor accepts as proof of an "accountable owner."

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-28) — verification chase on the scan's own owed item Routing: strategy / governance

Tags: paper-2, practitioner, Technology Illusion, Process Friction, Capability, COUNTERARGUMENT Source: arXiv (preprint, not peer reviewed) URL: https://arxiv.org/abs/2608.31017 Date published: 2026-08-31 Date flagged: 2026-09-27

Summary

Sebastian Fox, Luke Markham, Ryan Lail and Michael Karotsieris audited three commercial ambient AI scribes on the same 142 consultations — 565 notes drawn from recorded UK primary-care and US ambulatory encounters plus authored scenarios. The paper opens on the premise it exists to test: "Ambient AI scribes draft clinical notes under the reassurance that a clinician signs every note."

The instrument is stated in full in the abstract and is unusual enough to record before the finding. Twelve discovery passes proposed 13,678 candidate errors; the 5,898 that cleared an importance filter went to an adversarial panel of two models from different families, each instructed to refute what it could, and 618 survived. On that count, one note in three — 31.3% [27.0, 35.6] — carries a verified failure. The failures concentrate in "allergy and medication information, invented patient identity, and history written up as examination on telephone consultations that can contain none." No product was given a patient record; setting aside the two classes a record would have prefilled (invented identity and dates), the rate is 24.8% [20.8, 29.0] — the authors' own netted-out figure and the one to quote. One failure mode fell outside their scheme and was taken from published scribe-error taxonomies: "a treatment the clinician retracts, recorded as delivered care."

The verification chain has a human layer, and its size should be stated whenever the headline is: two clinicians adjudicated blind, disjoint samples — a physician author upheld 20 of 21 findings (95.2% [77.3, 99.2]) and an independent clinician, not an author, 12 of 12 ([75.8, 100]); both judged every sampled refusal genuine. So the 618 survivors were found by language models and adjudicated by language models, with human confirmation on 33 sampled findings in total.

Five Breakpoints Intersection

Technology Illusion: The deployment's sole safeguard is the clinician signature, and this audit prices what that signature is being asked to absorb — a verified failure in roughly one note in four after netting out test-harness artifacts, concentrated in allergy and medication fields. A control whose required catch rate is a quarter of all outputs, asserted rather than measured, is the Technology Illusion in its exact form: the organizational condition the technology assumed was solved (a reviewer with the time and retained skill to catch a fabricated examination) was never specified, resourced, or tested.

Process Friction: The failure classes named here are the ones plausibility-based reading cannot catch — "history written up as examination on telephone consultations that can contain none" and a retracted treatment recorded as delivered care both produce clinically coherent prose. Catching them requires re-deriving the note against the encounter, which is the documentation work the deployment was bought to remove; the control therefore reinstates the friction the technology was justified by eliminating.

Four Forces Intersection

Capability: The paper measures the volume and the class of error a reviewer must detect, which is the denominator the Capability force has lacked in every human-in-the-loop argument in this base — review capacity has been discussed as an unpriced cost and is here given a workload figure, though not a catch rate.

Commitment is deliberately not tagged. No organization's deployment decisions, review policies or staffing appear anywhere in the paper; tagging it would be the speculative exemption CLAUDE.md hard rule 4 forbids.

Relevance to Brandon's Positioning

This is the missing quantity under asset 5 ("The Human-in-the-Loop Control That Isn't"), and it arrives in the same clinical domain as the base's two strongest deployment instruments (Rotenstein's five health systems, Ohde's three). Asset 5 currently argues from two customer-support field experiments that "human in the loop" collapses two designs and calls both a control. It has never had a number for what the control is required to catch. It now has one, in a setting where the control is named in the product's own marketing.

The bound that must travel with every citation, because it is the difference between a strong argument and an overclaim: this paper audits scribe drafts, not signed notes. It does not measure whether clinicians caught these errors, so it does not demonstrate that the control failed. It establishes the control's workload and the adversarial character of that workload, and nothing about its performance. Cited as "one note in three slips through," it is wrong. Cited as "the signature is being asked to catch a verified failure in roughly one note in four, including fabricated examinations that read as normal," it is exactly right and it is new.

Set beside the "mitigation depreciates" challenge (logged 2026-09-25 on Budzyń et al.), the two form the strongest version of the prescriptive problem this base has assembled: the control's workload is high and adversarial (this paper), and the reviewer's unassisted capability declines with exposure to the thing being reviewed (Budzyń). Neither measures review quality, which remains the discriminating test.

Sources & Confidence

• Fox S, Markham L, Lail R, Karotsieris M, One note in three: a verified census of three deployed AI scribes, and the instrument that counted it, arXiv:2608.31017, submitted 2026-08-31; cs.CL, cs.AI, cs.CY — Moderate. Abstract read verbatim from the arXiv listing; full text not read. Not peer reviewed and no venue named in the comments field. Author affiliations are not established here and are deliberately not guessed.
• Not established, and not to be asserted: per-product breakdown across the three scribes (the abstract reports a pooled rate only), the products' identities, the UK/US split of the 142 consultations, how many of the 565 notes were authored scenarios rather than recorded encounters, and any catch rate by reviewing clinicians — Unknown
• Instrument caution, per CLAUDE.md hard rule 9: discovery and adjudication are both language-model steps (twelve discovery passes; a two-family adversarial panel). Human adjudication covers 33 findings across two blind samples, one by an author. This is a model-adjudicated census with a small human validation layer, not a clinician chart review, and the 31.3% should never be cited as though clinicians counted it.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-27 — surfaced by the exploratory slot's retrieval search, not by the target) Routing: strategy / governance

Tags: paper-2, Technology Illusion, Momentum Mirage, Capability Source: arXiv working paper (econ.GN) — Chatterji, Holtz, Rakholia, Tambe & Weeratunga (OpenAI; Columbia Business School; Wharton) URL: https://arxiv.org/abs/2608.12236 Date published: 2026-08-12 Date flagged: 2026-09-26

Instrument — recorded before the finding (mandatory for structural instruments)

Data source. Internal ChatGPT Enterprise account records — OpenAI's centrally administered workplace product — linked to (a) message-level usage telemetry, (b) inferred worker job-title classes and seniority levels from administrative account metadata, (c) an LLM task classification of message content, and (d) public-company financial statement data for the subset of U.S.-based enterprise customers that can be matched. The worker-level sample at the six-month adoption horizon covers over 1,500 organizations and over 17 million messages.

Inference method. Descriptive decomposition of usage growth into within- and between-firm components on an organization-week panel; cross-sectional comparison of adopter and non-adopter public firms on financial characteristics; and linear-probability regressions of adoption on lagged firm characteristics and on employee-normalized stocks of SG&A, R&D and capitalized software (log one plus the stock), used as proxies for accumulated organizational and intangible capital. The authors state explicitly that these are conditional associations and "should not be interpreted causally." There is no instrument, no experiment and no identification strategy anywhere in the paper.

Observation window. Usage data through March 2026; the token-growth series runs June 2025 – March 2026; adoption cohorts from January 2024. This is the first firm-level AI instrument in this base whose measurement window is entirely inside the generative era and extends into 2026.

Vendor position, stated first because it bounds every use. Four of five authors are OpenAI (Chatterji, Rakholia, Weeratunga; Holtz and Tambe are OpenAI paid contractors with Columbia and Wharton affiliations). This is the vendor studying its own product using its own customer records. Working paper, not peer reviewed, and marked "results are subject to change."

Summary

The paper documents four facts about enterprise generative-AI adoption. First, usage grew rapidly on both margins: aggregate output tokens produced by ChatGPT Enterprise customers grew roughly sevenfold between June 2025 and March 2026, and roughly fourfold within the fixed cohort of firms that had already adopted by June 2025 — so about half the growth came from intensification inside existing customers rather than from new logos.

Second, adoption is concentrated among large, resource-rich firms. Among U.S.-based public companies, median revenue is $2,275.1M for adopters against $209.6M for non-adopters, median employment 2,934 against 424, median market value $4,997.2M against $316.4M, and median R&D expense $113.1M against $9.9M. Adopters are also more R&D- and SG&A-intensive. In the regression, a one-log-point increase in lagged revenue per employee is associated with roughly 0.4 to 0.9 percentage points higher adoption probability. The authors read the pattern as showing that "early enterprise AI adoption is associated with greater prior investment in intangible and organizational capabilities," consistent with the view "that generative AI, like other general purpose technologies, depends on complementary capabilities."

Third, use is spread across the hierarchy rather than concentrated at one level — at the average firm after six months, managers and directors are about 24% of weekly active users, individual contributors and professionals 15%, senior ICs and principals 14%, executives about 10%, and early-career workers and trainees 7% — while intensity runs the other way, with early-career workers sending "many more messages than more senior employees" and marketing and communications workers sending more than executives. Fourth, use spans many knowledge-work tasks — writing, technical work, communication and information synthesis — rather than concentrating in a single workflow.

The authors' own conclusion is the sentence worth quoting: "adoption is only the beginning of deployment. The rapid adoption of generative AI by firms should therefore not be equated with immediate productivity transformation." They attribute the gap to a slow process of co-invention "in which firms discover use cases, invest in complements, and reorganize production." Their stated limitations include that the task results "do not measure downstream work products, productivity effects, or changes in organizational routines," and they name the missing work directly: "Future work should connect enterprise AI telemetry to measures" of outcomes.

Five Breakpoints Intersection

Technology Illusion: The paper puts accumulated intangible and organizational capital — stocks of SG&A, R&D and capitalized software per employee — on the right-hand side of an adoption probability and finds it predicts who deploys, while measuring no value variable of any kind; the firm that sells the technology, holding 17 million messages across 1,500+ organizations, states that adoption "should not be equated with immediate productivity transformation." That is the antecedent of Technology Illusion evidenced on the vendor's own telemetry and against the vendor's commercial interest.

Momentum Mirage: Aggregate output tokens grew roughly sevenfold in nine months — the single most quotable "enterprise AI is scaling" figure this base holds — and it appears in the same paper whose authors decline to infer transformation from it and who report that firms "are still actively learning how to integrate AI into organizational workflows"; volume growth is the progress metric the market reads, and the producer of the metric explicitly refuses the inference.

Four Forces Intersection

Capability: Adoption is predicted by pre-existing, employee-normalized stocks of SG&A, R&D and capitalized software — the authors' explicit proxies for "accumulated investment in organizational capabilities" — which makes Capability an accumulated stock that sets who can absorb the technology at all, measured here on financial statements rather than inferred from hiring text.

No Purpose, Commitment or Momentum tag: the paper contains no measure of stated intent, of decision-maker behaviour under tradeoff, or of sustained transformation energy. Honest omission rather than inference.

Relevance to Brandon's Positioning

This is the cross-genre corroboration the Schubert entry needed and could not supply itself. The scan's 2026-09-26 entry (schubert-organizational-technology-ladders-2026-08.md) evidences prior organizational condition → faster adoption from Lightcast job postings, the genre governed by the ESCALATED Résumé-Panel Consensus pattern and one this base is saturated in. Chatterji et al. reach a compatible conclusion — pre-existing intangible and organizational capital predicts who adopts — from usage telemetry joined to statutory financial statements, which shares no measurement, no vendor and no inference method with postings. On the pattern's own terms this is the first firm-level adoption result here that is not a member of the résumé/postings genre, and it is the first whose window runs into 2026.

But it corroborates the antecedent while supplying a rival reading of the mechanism, and that is the more useful contribution. Schubert's story is that a prior adaptation built capability that lowered the cost of the next adoption. Chatterji et al. find adopters are simply an order of magnitude larger and richer on every margin — 11x median revenue, 7x median employment, 11x median R&D. The simplest account of that is resources, not accumulated organizational learning, and the paper is explicitly non-causal so it cannot separate them. Schubert's IV is designed to, which is why the two belong in a draft together: one has identification and a saturated instrument, the other has a clean instrument and no identification. Neither alone establishes that organizational readiness rather than spending power drives absorption, and a referee will ask.

Most consequentially for paper-2's Claim 5: the paper is an authoritative statement that the outcome half has not been run, by the people best placed to run it. OpenAI holds the telemetry, has already linked it to public-company financials, and reports only descriptive facts — naming the outcome linkage as "future work." The story thread opened 2026-09-26 ("The ladder's second half") exists to catch exactly that result; this entry establishes that the data infrastructure for it now exists in identified hands, which makes the thread's criterion (a) a live watch rather than a speculative one.

And the quotable line runs against the seller's interest, which is the test this base already applies to Faros and the Tandem scribe study. "Adoption is only the beginning of deployment" is Brandon's Technology Illusion argument in OpenAI's own words. It is stronger for him than any consultancy survey on the same point, and he should cite it with the vendor affiliation attached rather than in spite of it.

Sources & Confidence

• Chatterji, A., Holtz, D., Rakholia, N., Tambe, P. & Weeratunga, G. "How Organizations Use AI: Evidence from ChatGPT." arXiv:2608.12236, last updated 13 August 2026 (submitted 12 August 2026), econ.GN. — Moderate. Working paper, not peer reviewed, marked "results are subject to change."
• Five limits, and the first two bound every citation. (1) Vendor research on the vendor's own product. Four of five authors are OpenAI; Holtz and Tambe contributed as paid OpenAI contractors. The descriptive growth figures are the ones most exposed to that interest; the "adoption ≠ productivity transformation" conclusion is the one least exposed, and is the one worth citing. (2) Nothing here is causal and the authors say so — "should not be interpreted causally." No outcome variable appears anywhere in the paper. (3) The adopter population is ChatGPT Enterprise customers, not AI adopters generally; firms standardising on Copilot, Gemini, Claude or in-house models are non-adopters by construction, so the adopter/non-adopter financial gap partly measures who buys this product. (4) Job titles and seniority are inferred from administrative account metadata and the authors note they are "incomplete and do not provide denominators for the full workforce in each role" — the 24%/15%/14%/10%/7% seniority shares are shares of observed weekly active users, not of headcount. (5) The extensive and intensive margins point in opposite directions on seniority and must not be collapsed: early-career workers are the smallest share of active users (7%) and the highest-intensity users. A citation saying "AI use is concentrated among junior staff" would be false on the first margin and true on the second.
• Second vendor-telemetry instrument in this base, and not an independent confirmation of the genre's validity. This base already holds Yu, Chen, Hu, Suri & Counts (yu-chen-hu-suri-counts-genai-productivity-communication-telemetry-2026-08.md), M365 Copilot telemetry across 11 companies and 40,164 enabled users. Two vendors measuring usage of their own products with their own logs share a structural interest even though they share no data; treat them as one genre with two members, per the 2026-08-24 shared-right-hand-side rule.
• Comparison entries: schubert-organizational-technology-ladders-2026-08.md (research-evidence, postings/IV) and bloom-sadun-van-reenen-americans-do-it-better-2012-02.md (blind management scores × IT → productivity) — High (American Economic Review).

Filed by / Routing

Filed by: Stafford (claude-opus-5), daily brief 2026-09-26 — exploratory slot Routing: strategy

Tags: paper-2, Technology Illusion, Capability, COUNTERARGUMENT Source: The Lancet Gastroenterology & Hepatology URL: https://pubmed.ncbi.nlm.nih.gov/40816301/ Date published: 2025-08-12 Date flagged: 2026-09-25

Summary

Budzyń and 23 co-authors — including Michał F Kamiński, Mette Kalager, Michael Bretthauer and Yuichi Mori — report a retrospective, observational study at four endoscopy centres in Poland participating in the ACCEPT (Artificial Intelligence in Colonoscopy for Cancer Prevention) trial. The centres introduced AI polyp-detection tools at the end of 2021, after which examinations were assigned to AI-assisted or unassisted by date. The authors compared the quality of unassisted colonoscopy in the three months before AI implementation with the three months after it. Their stated question: "It is not known if continuous exposure to artificial intelligence (AI) changes endoscopists' behaviour when conducting colonoscopy. We assessed how endoscopists who regularly used AI performed colonoscopy when AI was not in use."

Between 8 September 2021 and 9 March 2022, 1,443 patients underwent non-AI-assisted colonoscopy — 795 before and 648 after AI introduction (median age 61, IQR 45–70; 58.7% female). The primary outcome, adenoma detection rate of standard unassisted colonoscopy, fell from 28.4% (226/795) to 22.4% (145/648), an absolute difference of −6.0% (95% CI −10.5 to −1.6; p=0.0089). In multivariable logistic regression, three factors were independently associated with detection: exposure to AI (odds ratio 0.69, 95% CI 0.53–0.89), male patient sex (1.78, 1.38–2.30) and patient age ≥60 (3.60, 2.74–4.72). Funding: European Commission and the Japan Society for the Promotion of Science.

The authors' own interpretation is hedged and should be quoted as they wrote it: "Continuous exposure to AI might reduce the ADR of standard non-AI assisted colonoscopy, suggesting a negative effect on endoscopist behaviour." The design is before-and-after within the same centres with no contemporaneous unexposed control group, the post-exposure window is three months, and the abstract does not report the number of endoscopists — the unit of observation is the patient, not the clinician.

Five Breakpoints Intersection

• Technology Illusion: The deployment improved the measured outcome while the tool was running and degraded the same outcome by 6 percentage points when it was not — direct evidence that a technology can be installed on top of a human capability it silently consumes, and that the assisted number alone cannot tell an organization whether its underlying capacity rose or fell.

Four Forces Intersection

• Capability: The paper is the first real-world clinical measurement in this base of a professional capability declining as a function of using the system that was meant to support it, which is the empirical claim the Mitchell/Ghosh/Passi position paper (arXiv:2608.23642) rests on and does not itself supply.

Relevance to Brandon's Positioning

This is the single most load-bearing unread citation in the 2026-09-25 scan, and it was routed to Brandon as "read the deskilling citations before recommending human review as a mitigation." Read in the primary, it does two things and only two. It supports the mechanism: measured patient-outcome degradation of unassisted performance after AI introduction, at four centres, in a real clinical service. It does not support the load the commentary literature puts on it: the exposure window is three months, not "extended use"; the design has no contemporaneous control, so a general secular or pandemic-period decline across September 2021 to March 2022 is not excluded; and the outcome is detection while performing the task unassisted, which is not the same construct as the capacity to review and overrule a machine's output — the capability every governance source in this base actually depends on. For Brandon's assets this cuts cleanly: it strengthens asset 5 (The Human-in-the-Loop Control That Isn't) by giving the argument a clinical outcome rather than a laboratory proxy, and it puts a real amendment on asset 7 (The Ladder With No Verification Rung), whose proposed verification rung is exactly the capability this study shows depreciating. Cite it as the best available measurement of automation-induced skill loss tied to a patient outcome, and cite the three-month window in the same sentence.

Sources & Confidence

• Budzyń K, Romańczyk M, Kitala D, et al. "Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study." The Lancet Gastroenterology & Hepatology 2025;10(10). DOI 10.1016/S2468-1253(25)00133-5. First published 2025-08-12. Structured abstract read in full via the Europe PMC REST API (EXT_ID:40816301, resultType=core); full text not read — thelancet.com is subscription-gated. — High (publisher and author group), with the design caveats above carried in every citation.
• Mitchell M, Ghosh A, Passi S. AI Agents Push Humans Out of the Loop, arXiv:2608.23642 — the source that names this citation; held in research-evidence/knowledge-base/mitchell-ai-agents-push-humans-out-of-the-loop-2026-08.md. — Moderate (unrefereed position paper, no data of its own).

Filed by / Routing

Filed by: Stafford (claude-opus-5), citation chase, 2026-09-25 Routing: personal / governance

Tags: paper-2, Momentum Mirage, Capability, COUNTERARGUMENT Source: Michael Ewens & Xavier Giroud (Columbia Business School / NBER), author copy of the October 2025 revision; NBER Working Paper 34162 URL: http://www.columbia.edu/~xg2285/Hierarchy.pdf Date published: null — the author copy is dated October 2025 with no day; NBER lists issue date August 2025, revision date October 2025. No exact date is stated, so none is invented (hard rule 1). Date flagged: 2026-09-25

Summary

Two versions of this paper exist and the evidence base has been holding the older one. Both were read in full today, cover to table, and the differences are recorded here rather than characterised from a summary.

August 2025 (NBER WP 34162, original): "We introduce a novel measure of corporate hierarchies for over 2,500 U.S. public firms. This measure is obtained from online resumes of 16 million employees..."; the Covid layer-expansion result is reported for biotech companies; the volatility fact is "more stable cash flows"; Table 12 (firm response to AI) reports coefficients −0.50, −0.62, −0.58, −0.13 on 4,606 observations across 1,250 firms, all starred at the 10% level only.

October 2025 (revision; author copy, and the abstract NBER's landing page now serves): "We introduce a novel measure of corporate hierarchies for over 3,100 U.S. public firms. This measure is obtained from online resumes of 7 million employees..."; the Covid result is reported for pharmaceutical companies; the volatility fact becomes "operating asset volatility"; Table 12 reports coefficients −0.61, −0.75, −0.68, −0.14 on 5,505 observations across 1,491 firms, again all starred at the 10% level only, with standard errors 0.34, 0.40, 0.36 and 0.074. Mean dependent variable moves 2.44 → 2.41; the reported fit line changes from R2 0.99 to Within R2 0.01.

Unchanged across both versions: the average firm has ten hierarchical layers; the AI right-hand-side variables are taken from Babina, Fedyk, He and Hodson (2024) rather than independently constructed; the authors' stated limitation that "the firm fixed effects reduce the power of our tests due to the limited within-firm variation in hierarchical layers"; the long-differences robustness check that "yields estimates that are similar to those in Table 12"; the conclusion that "companies flattened their hierarchies following the adoption of artificial intelligence (AI) technologies"; and the absence of any span-of-control measure and of any outcome variable linking flattening to subsequent firm performance.

Access note, load-bearing for anyone re-checking this. https://www.nber.org/system/files/working_papers/w34162/revisions/w34162.rev0.pdf — the URL held in the base and queued by the 2026-09-25 scan as the route to the revision — still serves the August 2025 text (cover page dated August 2025, abstract reading 2,500 firms / 16 million resumes / biotech), while the NBER landing page abstract for the same paper serves the October revision. https://www.nber.org/system/files/working_papers/w34162/w34162.pdf returns HTTP 403 to this container. The working route to the revision is the author copy at columbia.edu/~xg2285/Hierarchy.pdf (HTTP 200).

Five Breakpoints Intersection

• Momentum Mirage: The revision enlarges the firm panel and raises every AI coefficient by roughly 20% while adding no outcome variable whatsoever — fifteen months and a substantially rebuilt sample later, the base's flagship flattening instrument still measures the structural artifact of transformation at scale and says nothing about whether any of it converted into movement.

Four Forces Intersection

• Capability: The authors' mechanism is unchanged and remains a Capability claim in the framework's exact sense — layer counts fall because cheaper knowledge acquisition reduces the need for problem-solver layers, making structure a function of where knowledge sits rather than of intent or follow-through.

Relevance to Brandon's Positioning

Three concrete consequences. (1) A live citation error is corrected. The base's existing KB entry (research-evidence/knowledge-base/ewens-giroud-corporate-hierarchy-2025-08.md) carries a "Citation warning, verified today" dated 2026-08-14 asserting that secondhand accounts reporting 3,100 firms, 7 million resumes and pharmaceutical are fabrications and that the correct figures are 2,500+/16 million/biotech. Those "fabricated" figures are the authors' own October 2025 revision. The warning was a false positive produced by comparing a summary of the new version against a read of the old one, and it should be replaced rather than deleted — the fabrication-detection instinct was right seven times and wrong here, and the correction is the useful record. (2) Two of Brandon's drafts cite superseded numbers. drafts/the-flattening-bet.md and drafts/four-papers-one-method.md both cite 16 million resumes and the 4,606/1,250 Table 12 figures. (3) The instrument gets broader and thinner at the same time, and that is the reading worth having. Firm coverage rises from 2,500+ to 3,100+ while the resume count falls from 16 million to 7 million — average resumes per firm drops from roughly 6,400 to roughly 2,260, about a 65% thinning, for a measure that infers hierarchical layers from a network of observed job transitions. The result survives and gets larger; the evidentiary base under each firm gets sparser; the authors' own power limitation is unchanged. Hard rule 9 is untouched: Table 12's right-hand side is still Babina et al., so these two papers remain one instrument and must never be cited as independent confirmations.

Sources & Confidence

• Ewens M, Giroud X. "Corporate Hierarchy." Author copy dated October 2025, columbia.edu/~xg2285/Hierarchy.pdf, 65 pages, downloaded and text-extracted in full on 2026-09-25. Abstract, Section 8 AI discussion, Table 12 and Conclusion read directly. — High (primary, author-hosted), with the standing caveat that a revised working paper has still not been peer-reviewed.
• Ewens M, Giroud X. "Corporate Hierarchy." NBER Working Paper 34162, August 2025, served at nber.org/system/files/working_papers/w34162/revisions/w34162.rev0.pdf, 65 pages, downloaded and text-extracted in full on 2026-09-25 for the side-by-side. — High (primary).
• NBER landing page nber.org/papers/w34162, fetched 2026-09-25: abstract matches the October revision; issue date August 2025, revision date October 2025. — High.
• Babina T, Fedyk A, He A, Hodson J (2024), Journal of Financial Economics 151, 103745 — origin of the AI-adoption measures on Table 12's right-hand side in both versions. Not read directly. — Unknown pending a read.

Filed by / Routing

Filed by: Stafford (claude-opus-5), exploratory slot / instrument verification, 2026-09-25 Routing: strategy

Tags: paper-2, practitioner, Technology Illusion, Incentive Fragmentation, Momentum Mirage, Capability, Commitment, COUNTERARGUMENT Source: JAMA URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC13044793/ Date published: 2026-04-01 Date flagged: 2026-09-24

One source, two files — never cite as two. The abstract-level entry for this paper lives in research-evidence/knowledge-base/rotenstein-jama-ai-scribes-clinician-time-multisite-2026-04.md, built 2026-08-27 from the structured abstract. This file is the full-text read of the same paper, made on 2026-09-24 to answer a question the abstract could not: does the base's strongest instrument report its results by health system? It adds Methods, Sensitivity, Discussion and Limitations content that the abstract does not contain, and corrects nothing in the earlier entry. The two files are one source and one instrument.

Summary

Rotenstein, Holmgren, Thombley et al. (32 authors; senior author Rebecca G. Mishuris, Chief Health Information Officer, Mass General Brigham; Julia Adler-Milstein and David W. Bates among the co-authors) report a multisite longitudinal cohort study of AI scribe adoption at five US academic health systems, published in JAMA 335(16):1408–1417, doi:10.1001/jama.2026.2253, first publication 2026-04-01 (PMID 41920565; PMC13044793). The Methods name the five institutions: Mass General Brigham, Emory Healthcare, the University of California San Francisco, Yale New Haven Health, and the University of California Davis. The sample is 8,581 ambulatory clinicians — 1,809 who received AI scribe access and 6,772 who did not — across 181,273 clinician-month observations, with outcomes drawn from Epic's Signal database rather than from self-report. Adoption was opt-in at four of the five sites. Headline results: total electronic health record (EHR) time fell 3.3% (about 13 minutes per 8 scheduled patient hours), documentation time fell 11.7% (16.0 minutes, 95% CI 13.7–18.3), weekly visits rose 0.49 (95% CI 0.17–0.81), and EHR time outside scheduled work hours did not change significantly.

The methodological facts that this file exists to record. The estimator is a difference-in-differences ordinary least-squares model with clinician fixed effects and month fixed effects, standard errors clustered at the clinician level. Because every clinician sits in one institution, clinician fixed effects absorb all time-invariant differences between the five sites: between-organization variation is removed from the estimate by construction rather than reported. Subgroup analyses were run by interacting adoption with specialty, role, sex, and AI scribe use intensity — four clinician characteristics, and no organizational characteristic. Site enters the analysis exactly twice, both times as a robustness device: "To ensure results were not driven by a single site, we repeated our difference-in-differences models excluding each site sequentially," and in the Limitations, "While many of the control observations derived from 1 institution (Mass General Brigham) given data collection specifications, the sensitivity analysis excluding data from Mass General Brigham suggested that inclusion of data from this site only makes the presented estimates more conservative." The direction of one site's influence on the pooled estimate is reported; no magnitude, and no per-site estimate, appears anywhere.

The organizational variation is documented, and it is documented as a nuisance. The Methods state that the sites' rollouts differed on the most consequential organizational choice available — who was allowed to have the tool: "sites defined the clinicians who were eligible for gaining access to the tool (eg, ranging from no restrictions on eligibility to restrictions based on clinical volume or eligibility based on leadership nomination)." Vendors differed too (Ambience, Nuance DAX Copilot, or Abridge, individually or in combination), on a common Epic EHR. These institutional characteristics and rollout designs are tabulated in eTable 1 in Supplement 1 — which this read could not obtain (the PMC supplement link returns a download interstitial). The Discussion states the pooling rationale plainly: "Given their derivation from pooled data from 5 health systems, these estimates are less likely to be biased by site-specific implementation details and represent generalizable results regarding the relative time savings associated with clinicians adopting AI scribes in real-world settings." The Limitations restate it: "Implementation details varied across the study sites; thus, estimates reflect averages and assessments of real-world implementation experiences rather than measurements of maximal efficacy of AI scribes."

Five Breakpoints Intersection

Technology Illusion (new in this file): The paper's stated justification for pooling — that pooled estimates are "less likely to be biased by site-specific implementation details" — treats how each of five organizations chose to deploy the tool as a source of bias to be averaged out in order to recover the technology's effect, which is the Technology Illusion premise expressed as a research design: the object of measurement is the capability, and the organization enters the model only as error.

Incentive Fragmentation (new in this file): At one or more of the five sites, eligibility for the tool was set by "leadership nomination" — access to a documentation-burden technology allocated by discretionary managerial grant rather than by where the documentation burden actually sat — while other sites imposed no eligibility restriction at all; the paper records this three-way difference in access rules and never puts it on the same page as an outcome.

Momentum Mirage (carried, not new): Across 1,809 adopters and a study period running June 2023 to August 2025, EHR time outside scheduled work hours did not change significantly — the measure closest to the clinician-burnout problem the deployments were justified by — while programme-visible metrics (sites live, clinicians onboarded) advanced.

Four Forces Intersection

Capability: The paper measures the capability question in its flow-and-readiness sense and finds the gain concentrated in clinicians who used the tool in 50% or more of visits, locating the realised benefit in depth of use rather than in breadth of access — access is not adoption, measured from system data rather than survey.

Commitment: Five organizations each decided who would be eligible for the tool, with rules ranging from open access to leadership nomination, and the study design then removes those decisions from the estimate; the Commitment variable — whether decision-makers acted consistently with the stated aim of reducing documentation burden when choosing who got the tool — is documented in a supplement and never tested against an outcome.

Relevance to Brandon's Positioning

This is the single most consequential thing established about Brandon's own evidence base in several weeks, and it cuts two ways at once. It settles the question Stafford's 2026-09-23 brief asked and pre-registered as its next chase: the strongest instrument in the base does not report results by health system. But the answer is more useful than either branch that brief predicted. The paper did not overlook the site dimension — it named five organizations, documented that their access rules differed, ran a leave-one-site-out analysis, reported the direction in which one site moved the pooled estimate, absorbed all between-site variation into clinician fixed effects, analysed heterogeneity on four person-level characteristics and none organizational, and described site-specific implementation as bias that pooling removes. The organizational variance Brandon's thesis is about was in these authors' hands, was computed at least five times, and was reported as a robustness check.

The bearing on paper-2's Claim 5 is direct and partly corrective of Stafford's own 2026-09-23 reasoning. That brief argued against the scan's dispersion restatement partly on the grounds that "Rotenstein itself contains a dispersion result running the other way" — heavy users at roughly 1.6×/1.7×/2.0× the all-adopter estimates. That argument must now be weakened by its author. The person-level subgroup estimates are the only kind of subgroup this paper estimated; their existence is a fact about the paper's design, not evidence that organizational variance is small. This instrument cannot speak to organizational variance in either direction, and it must not be cited on either side of the dispersion question again.

What it does supply is the cheapest available route to the observation the base has now recorded as missing for two consecutive days. The decomposition does not require a new study: it requires five named health systems, whose informatics leadership are the paper's own co-authors, to re-run an existing model with a site interaction, or to publish eTable 1 against the outcome table. That converts asset 14 from an essay resting on three second-hand figures into an argument with a named, reachable instrument and a specific unasked question behind it.

Sources & Confidence

• Rotenstein LS, Holmgren AJ, Thombley R, et al., "Changes in Clinician Time Expenditure and Visit Quantity With Adoption of Artificial Intelligence-Powered Scribes: A Multisite Study," JAMA 335(16):1408–1417, doi:10.1001/jama.2026.2253, first publication 2026-04-01 — High (peer-reviewed; full text read directly at PMC13044793, Methods/Results/Discussion/Limitations; all quotations verbatim from that text)
• Europe PMC REST API record for PMID 41920565 — High (used for the citation, DOI, PMCID, 32-author list and complete affiliation list)
• eTable 1 and eMethods in Supplement 1 — NOT READ. The PMC supplement link (NIHMS2166399-supplement-supplement.pdf) returns a download interstitial rather than the file. Everything stated here about eTable 1 is the main text's description of it, quoted. Its contents are unverified and must not be characterised further until obtained — Unknown
• Secondary coverage cross-check: HIT Consultant (2026-04-01) lists the five sites correctly; a general web search returned a site list substituting NYU for Emory, and separately asserted that the supplement contains "site-specific estimates and random effects models." Both are wrong or unsupported — the Methods name Emory, and no random-effects or site-specific estimate is described anywhere in the main text. Recorded as the seventh instance in this base of a search intermediary manufacturing a methodological detail — Low

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-24 — exploratory slot, pre-registered on 2026-09-23) Routing: strategy / governance

Tags: paper-2 Technology Illusion Momentum Mirage Capability COUNTERARGUMENT Source: npj Digital Medicine (Nature Portfolio) URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC13172454/ Date published: 2026-03-23 Date flagged: 2026-09-23

Summary

A Perspective by Ohde JW, Thompson A, Liu Z, Rost LM, Overgaard JD, Tan JYE, Overgaard SM, Sarmiento RFR, Ke Y, Liew JCK, Ong JCL and Liu N (npj Digital Medicine 2026;9, doi:10.1038/s41746-026-02554-0; PMCID PMC13172454). It argues that ambient AI scribes were tested in low-acuity ambulatory settings and that extending them into diverse care settings raises unresolved clinical, technical and ethical problems. It is a Perspective, not an empirical study: it reports no new data and has no sample of its own.

Its load-bearing passage for this base is a compilation of deployment outcomes at three named US health systems, each running an ambient-scribe deployment, each measured against clinician time. Mass General Brigham: "median total EHR time reduced by 5.6 min per appointment, with the biggest gains in specialty practices and heavy EHR users." The Permanente Medical Group, described as the largest implementation to date: "minimal time savings of only 18 s per appointment compared to non-users," alongside 84% of clinicians reporting a positive impact on visit interactions and 56% of patients reporting a positive impact on visit quality. Intermountain Health: a "matched cohort study reported no statistically significant productivity gains."

The authors then explain the spread in a single sentence: "This heterogeneity may be explained by study design–related factors, including variation in the types of ambient AI scribes evaluated, insufficient sample sizes, and the absence of standardized protocols for outcome measurement." They attribute the variation entirely to measurement and design, and name no organizational characteristic of any of the three systems. The paper reports no variance decomposition, no intraclass correlation, and no within-versus-between-organization analysis. Its own recommendation is that the most consistent value of these tools lies in reduced cognitive load, patient engagement and clinician well-being rather than in time saved.

Five Breakpoints Intersection

Technology Illusion: Three health systems deployed the same class of capability against the same stated problem (documentation burden) and got results differing by roughly nineteen-fold in the measure the investment was justified by — 5.6 minutes per appointment at one system against 18 seconds at another and a null at a third — which is direct evidence that the technology does not carry its own outcome and that something about the receiving organization or the measurement of it determines what arrives.

Momentum Mirage: The Permanente deployment records 2.5 million uses, 84% clinician-reported positive impact and 56% patient-reported positive impact against 18 seconds of measured time saved per appointment, so the adoption and sentiment metrics move at full scale while the outcome the deployment was justified by is close to zero.

Four Forces Intersection

Capability: The Mass General Brigham result concentrates in "specialty practices and heavy EHR users" — the gain lands where existing practice and usage intensity already suited it, which locates the determinant in the organization's and the user's prior condition rather than in the tool.

No Purpose, Commitment or Momentum tags — the Perspective reports no evidence on stated intent, follow-through or sustained energy at any of the three organizations, and tagging them would be guessing.

Relevance to Brandon's Positioning

This is the strongest available answer to the question the 2026-09-23 research-evidence scan raised in its Open Question and declared unanswered: whether anyone reports the spread in AI outcomes across organizations running the same intervention. The scan's claim — "no instrument anywhere decomposes within-firm against between-firm variance in AI outcomes" — survives on its literal terms, because this paper runs no decomposition. But the scan built its Open Question and its proposed client diagnostic on the premise that the between-organization spread is unavailable, and that premise is wrong: three named systems, one intervention class, one function, one health economy, already compiled in a peer-reviewed venue.

The more valuable finding is what the authors do with it. They attribute a nineteen-fold spread across three named organizations exclusively to study-design factors — different products, small samples, no standardized outcome protocol — and consider no organizational variable at all. That is a live rival explanation for Brandon's central claim, stated by domain experts in a Nature-portfolio journal, and it is not weak: different scribe products and different outcome definitions genuinely could produce a spread this size with no organizational variance whatsoever. Neither explanation has been tested. The honest position, and the publishable one, is the symmetry: the same dispersion supports an organizational reading and a measurement reading, nobody has run the observation that separates them, and the field has defaulted to the measurement reading without testing it. For paper-2 Claim 5 this is a caution against the dispersion-shaped restatement proposed by the 09-23 scan, which would rest a core claim on exactly the inference this paper declines to make. Weaker than Rotenstein (JAMA, 1,809 adopters, 26 months, EHR-derived) as evidence; stronger than anything in the base as a map of the disagreement.

Sources & Confidence

• Ohde JW, Thompson A, Liu Z, et al., "Barriers and opportunities of scaling ambient AI scribes for clinical documentation across diverse healthcare settings," npj Digital Medicine, 2026-03-23, doi:10.1038/s41746-026-02554-0, PMCID PMC13172454 — High as to what the Perspective says (Nature Portfolio, peer-reviewed, full text read directly via PMC after nature.com returned an IdP auth redirect). Moderate as to the three deployment figures themselves: each is reported here at second hand against the Perspective's own references 3, 11 and 12, and none of the three underlying primaries was read for this entry. Do not cite 5.6 min / 18 s / null to the primary studies without obtaining them.
• Secondary corroboration of the Permanente figure and its scale (7,000 physicians, 2.5M uses in one year) appears in NEJM Catalyst CAT.25.0040 and in Kaiser Permanente Division of Research and Becker's coverage — Moderate, not read in full, listed as a chase target rather than as support.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-23 — Wednesday falsification sweep) Routing: strategy

Tags: paper-2, Strategic Disconnection, Momentum Mirage, Purpose, COUNTERARGUMENT Source: Julie Wulf, Harvard University — Harvard Business School Working Paper 12-087 URL: https://www.hbs.edu/ris/Publication%20Files/12-087_bc50bde2-3016-457a-9bee-dc988cb1056b.pdf Date published: 2012-04-09 Date flagged: 2026-09-22

Instrument (recorded before the finding, per CLAUDE.md hard rule 9)

Data source: a large-sample panel of reporting relationships, job descriptions and compensation structures for over 300 large U.S. firms, described by the author as covering 1986–2006 (the delayering and span results are reported on 1986–1998/1999 cuts). The panel's construction and sampling are documented in Rajan & Wulf (2006), which was not read here — this file must not assert how the panel was assembled.

Inference method: hierarchy is taken from firm-reported reporting lines and job descriptions, not inferred by a researcher from résumés or job postings. Depth is measured as "the average number of hierarchical positions between the CEO and division managers," with division managers defined as "the lowest-level managers with profit-center responsibility." The author states the panel can track a given division-manager position over time: "our data allow us to track the reporting relationship of a given division-manager position over time and to trace its movement up the hierarchy." Decision rights are not observed directly; relative pay is used as their proxy, which is the paper's central inferential step and its main vulnerability.

Observation window: 1986–2006 for the panel; the supporting CEO interviews (seven firms) and CEO time-use survey were run 2010–2011. The author states the two sit in different periods: "Our quantitative and qualitative data are from two distinct time periods." Entirely pre-generative. There is no AI variable anywhere in this paper.

Not independent of: Rajan & Wulf (2006), Guadalupe & Wulf (2010) and Guadalupe, Li & Wulf (2011) — this article is a synthesis of the author's own prior results on the same underlying panel. It is one instrument and one research group, not four corroborating studies, and must never be cited as though those papers independently confirm each other. It is independent of Smeets & Warzynski (2006): different country, different firms, different decade, different research group.

Summary

Wulf asks whether the delayering that management consultants and the business press spent decades urging on large firms actually did what it was sold as doing. She states the sales pitch precisely: flattening "typically refers to the elimination of layers in a firm's organizational hierarchy, and the broadening of managers' spans of control," and "its alleged benefits flow primarily from changes in internal governance: by pushing decisions downward, firms not only enhance customer and market responsiveness, but also improve accountability and morale." She notes that the Boston Consulting Group applied for a trademark on the term "delayering" in 2005, and that on the question of whether it worked, "no one has addressed these questions rigorously and empirically."

The flattening itself is confirmed. Firms did delayer: the number of positions between division heads and the CEO "declined by about 25 percent, from 1.6 in 1986 to 1.2 in 1999"; firms with a COO fell from 55 percent in 1986 to 45 percent in 1998; CEO span of control rose "from 4.5 in 1986 to almost 7 in 1999," and on a smaller later sample "more than doubled from 4.7 in the second half of the 1980s to 9.8 approximately 20 years later." Firms also shifted management pay toward performance-based compensation, which the author reads as consistent with some genuine delegation.

The governance outcome went the other way. The abstract states it: "while firms have delayered, flattened firms can exhibit more control and decision-making at the top. Managers take note. Flattening can lead to exactly the opposite effects from what it promises to do." Three strands support it. (1) Composition: CEOs "dramatically increased the number of corporate staff (or functional managers) that comprise their top executive teams" — CFO, CHRO, CIO, CMO, General Counsel. (2) Pay: "division manager pay declines as functional managers join the executive team," and critically "this relationship is driven by 'product' functional managers who perform activities that are typically the responsibility of division managers (e.g., marketing and R&D) and not the administrative functions." Her reading: "functional managers at the top made decisions or performed activities that were previously performed by the division manager. This finding is directly at odds with what we think of as flattening." (3) What CEOs say and do: in interviews, "the most consistent message from the CEOs was that they had changed the structure of their executive team and broadened their span of control to 'get closer to the businesses.' They or their predecessors had eliminated the COO position in pursuit of more control, more oversight and more involvement in business operations and decisions, not less." Time-use data agrees: CEOs with broader spans "spend more time in meetings and less time alone," and "a COO acts as a substitute for CEO attention to internal functions/operations."

Her own summary declines the tidy version: *"Standard classifications of 'centralization' or 'decentralization' are too simple to represent accurately these changes in internal governance. Firms are doing both."* The consequence she draws for practitioners is a mismatch: "a manager may flatten structure to push decisions down and then hire and develop division managers suited to 'being the boss.' But if flattening actually pushes decisions up, the division managers are now out of sync with the organization: they don't have autonomy to make decisions and there is a mismatch between managerial talent and decision rights."

Five Breakpoints Intersection

Strategic Disconnection: The stated purpose of the intervention and its measured effect ran in opposite directions across 300+ firms — flattening was sold and bought as "pushing decisions downward" to "improve accountability and morale," and what the data show is that decision rights moved up, to functional managers at the top, leaving division managers, in Wulf's words, "out of sync with the organization: they don't have autonomy to make decisions and there is a mismatch between managerial talent and decision rights." That is a stated direction and an operational reality that never met, measured on firm-reported structures rather than diagnosed after a failure.

Momentum Mirage: The metric that certifies a delayering succeeded is the one that moved — depth fell 25 percent, COOs disappeared, spans nearly doubled — while the governance change the exercise was undertaken to produce did not happen at all. An organization reading its own layer count in 1999 would have seen a completed transformation; the decision rights it was transforming to relocate had moved the other way, and nothing in the layer count could show that.

Considered and deliberately not tagged — Incentive Fragmentation. The pay finding (division-manager pay falling as product-functional managers join the top team) is the most quantitatively specific thing in the paper and the temptation is to read it as misaligned incentives between two groups. It is declined: Wulf uses relative pay as a proxy for decision rights, not as an incentive-conflict finding, and the paper makes no claim that the pay movement misaligned anyone's behaviour or caused any outcome. Tagging it would convert her instrument into a finding, which is the speculative tagging CLAUDE.md prohibits.

Four Forces Intersection

Purpose: The clarity-of-intent failure here is not that the organizations failed to communicate a goal — it is that the goal they announced and the goal their CEOs describe in interviews are different goals. The public rationale was delegation and responsiveness; the CEOs' own account is that they delayered "in pursuit of more control, more oversight and more involvement in business operations and decisions, not less." An intervention whose sponsors privately want the opposite of what it publicly promises cannot produce a shared definition of success, and the division managers left holding the mismatch are the measured consequence.

Considered and not tagged — Capability. The architecture question (whether firms had the machinery to run at the new spans) is the natural Four Forces reading of a delayering study, and it is the one Smeets & Warzynski support. Wulf measures decision rights and pay, not supervisory capacity, and nothing in the paper speaks to whether the flattened firms could execute at their new shape.

Relevance to Brandon's Positioning

This is the second independent member of the construct today's 12:00 scan identified and declined to open as a pattern — that delayering relocates the work the removed layer carried rather than eliminating it, and that the relocation is invisible in the layer count that certifies the flattening. The scan had one member (Smeets & Warzynski's informal assistant-manager tier, one Scandinavian firm, 2001) and correctly said one firm is an observation. Wulf supplies 300+ U.S. firms over two decades, on the same genre of instrument — employer-reported chain of command, not résumé inference — from a different research group. The two members disagree about the direction of the relocation, and that is the finding rather than a problem with it: in Smeets & Warzynski the supervisory work reappears below, as an unofficial tier the firm declined to number; in Wulf the decision rights migrate up, into a C-suite that grew functional specialists. Displacement is what both measure. Where it lands is contingent.

It bears directly on Brandon's published position, and in his favour. Paper 4 (The AI Mirror, published 2026-05-26) asserts of the stewardship functions: "Those functions require carriers. They may be redesigned. They may be concentrated in fewer people than before. But they do not disappear because routing got cheaper," and warns that restructurings "correctly identify routing work that technology can absorb, then cut the human layer that also carried informal interpretation, escalation and judgment." Wulf is that claim measured on the previous shock: the layer went, the work concentrated in fewer people at the top, and the org chart recorded only the first half. Brandon published the mechanism as an assertion in May; it now has a 300-firm precedent from 2012.

It also corrects the restructuring asset 2 is about to undergo. The 2026-09-18 amendment to The Flattening Bet proposed re-centring the essay on a fork — coordination-cost theory predicts hierarchy shrinks, monitorability theory predicts it re-sorts — and called the two "distinguishable with data organizations already publish." Wulf ran a version of that test on the previous episode and the answer was both, simultaneously: depth fell and authority re-sorted upward, in the same firms in the same period, which is why she refuses the centralization/decentralization vocabulary outright. Meta's Project OT (research-evidence, 2026-08-26) already showed both running at once in one company. Two independent instruments now say the fork is not a fork, and an essay built on it would be knocked over by anyone who knows this literature. What survives, and is better, is Wulf's measurement: relative pay movement between levels as the observable proxy for where decision rights went.

Finally, it is a live counterargument and is tagged as one. If the previous delayering shock left stewardship strengthened at the top — more functional oversight, more CEO involvement, more internal meeting time — then Paper 4's prediction that AI-era delayering strands stewardship needs a stated reason why this episode differs. Logged in knowledge/thesis-challenges.md, 2026-09-22.

Sources & Confidence

• Julie Wulf, The Flattened Firm — Not as Advertised, Harvard Business School Working Paper 12-087, dated 09 April 2012 (text dated March 2012). Full PDF read directly from hbs.edu and extracted with PyMuPDF; every quotation above is from that text. Moderate-to-high — Harvard, and the author is the principal investigator of the underlying panel, but this is an unrefereed working paper written as a practitioner-facing synthesis of the author's own prior results, not a new estimation. It reports no new regressions, no standard errors and no sample sizes for the claims it summarizes.
• Underlying results are attributed by the author to Rajan & Wulf (2006) (Review of Economics and Statistics 88(4):759–773; NBER WP 9633, April 2003, revised June 2020), Guadalupe & Wulf (2010) and Guadalupe, Li & Wulf (2011). None of these was read here. The NBER abstract page for w9633 was read and confirms the 300+ firm panel, the rise in direct CEO reports, the fall in levels between division heads and CEO, and that "flattening cannot be characterized simply as centralization or decentralization." High for the citation facts; the findings are cited as Wulf reports them, and page-level claims must not be attributed to the journal articles.
• Smeets & Warzynski (2006), research-evidence knowledge-base/smeets-warzynski-corporate-hierarchy-span-of-control-2006-02.md — the independent second member of the displacement construct. Moderate.
• Brandon Freitag, The AI Mirror (Paper 4, published 2026-05-26), §"Routing and Stewardship Are Not the Same Thing" — papers/paper-4-the-ai-mirror.md, canonical text supplied 2026-08-30. High.
• Boston Consulting Group's delayering claim is quoted by Wulf from "Global Delayering for Competitive Advantage," BCG Opportunities for Action, April 2006; not read here, cited only as Wulf cites it. Low as a standalone reference.

Filed by / Routing

Filed by: Stafford (claude-opus-5), daily brief, exploratory slot, 2026-09-22 Routing: strategy

Tags: paper-2, practitioner, Incentive Fragmentation, Strategic Disconnection, Momentum Mirage, Commitment, Purpose Source: Catchpoint (annual SRE Survey; report authored by Leo Vasiliou with Kurt Andersen and contributing practitioners) URL: https://resources.catchpoint.com/hubfs/Website%20Assets%20-%20Briefs,%20EBooks,%20etc/The%20SRE%20Report%202025%20Catchpoint.pdf Date published: 2025-01-13 Date flagged: 2026-09-21

Instrument — recorded before the findings, per CLAUDE.md hard rule 9. Data source: opt-in online survey promoted by an observability vendor to a self-selected reliability audience. Inference method: single-respondent self-report throughout; no telemetry, no administrative data, no organization-level verification. Observation window: survey open six weeks during July and August 2024; report published 13 January 2025. Sample: n = 301 worldwide. Composition (per the report's own demographics): 68% North America; 40% technology-platform or as-a-service providers; roughly half the sample carries managerial responsibility (25% individual contributor, 24% up to one level of direct reports, the remainder deeper). Catchpoint sells Internet Performance Monitoring — the tooling questions are vendor-interested and are not used here. One respondent per organization, so every figure below is a person's report of their own organization, not a count of anything.

Summary

Catchpoint's seventh annual SRE Report presents survey data across seven "insights." Two are load-bearing for this base. Insight III — "The Danger of Unstable Organizational Priorities" reports that a majority of respondents feel pressure to put delivery ahead of reliability: asked "Do you feel pressured to place release schedules or deadlines as the priority — instead of reliability as the priority?", 41% answer "often" or "always." The report calls that figure "especially concerning" and frames it as "a classic case of agility versus stability: businesses want updates, new features, and revenue growth, whereas practitioners prioritize reliability and resilience." It further reports that the pressure tracks priority instability — "when production pressure to perform is more frequent, the less stable an organization's priorities appear to be."

The same insight records a gap the report names but does not resolve. 53% agree that "the challenges and problems of product reliability are addressed by the organization", and the report states that the majority also believe OKRs are clearly communicated and that stability is the predominant team posture — "despite" which "the majority still indicated feeling pressured to prioritize release schedules over reliability." In Insight I, asked which practices the organization should prioritize adopting over the next twelve months, 41% name site reliability engineering itself and 40% name service- or experience-level objectives (SLOs/XLOs) — the two top answers, from a population that self-identifies as reliability practitioners. Insight II reports that toil rose for the first time in five years, against the expectation that AI would reduce it; the report's hypothesis is that "the free time created by expediting valuable activities may end up being filled with toilsome tasks."

Error budgets appear in the report only as an illustration. A burndown-chart page states that "performance and error budgets are inherently part of SLOs and XLOs" and that "visualizing error budgets helps stakeholders understand risk, honor SLOs, and balance agility with stability." The report contains no measurement of error-budget adoption or usage. Separately, in a signed "View from the field" essay, Dave O'Connor (Google SRE 2004–2021, now a reliability consultant) writes that "Lore is Mobile — the taste-makers who initially either sponsored or created an SRE group are increasingly likely to have moved on, for various — possibly unfortunate — reasons. This leaves org-wide priorities in flux, and it's entirely possible for SRE groups to not be aware that the fundamental attitude to Reliability in their organisation has either shifted — or could shift — at any moment." He adds that product organizations increasingly hold "a calm, holistic picture of their need for investment in Reliability" but with "less of an appetite to do the hard work of adjusting the organisation's capabilities to suit."

Five Breakpoints Intersection

• Incentive Fragmentation: 41% of reliability practitioners report being pressured "often" or "always" to put release schedules ahead of reliability — the two parties either side of the delivery/reliability interface are answering to different numbers, reported from inside the one discipline that built a mechanism (the error budget) specifically to make that trade explicit rather than contested.
• Strategic Disconnection: the report's own "where it gets complicated" finding — 53% agree reliability problems are addressed by the organization while a majority simultaneously report deadline pressure overriding reliability — is stated direction and operating reality diverging inside the same respondent's answers, and the report names the contradiction without explaining it.
• Momentum Mirage: toil rose for the first time in five years during the period AI tooling was expected to reduce it, and the report's own hypothesis is that freed capacity refilled with low-value work — activity sustained, the outcome it was supposed to buy absent.

Technology Illusion was considered and deliberately not tagged. The obvious reading — organizations bought observability tooling instead of fixing the incentive split — is available but is not measured anywhere in this instrument, and the publisher sells that tooling, which makes a speculative tag here worse than useless.

Four Forces Intersection

• Commitment: the Force asks whether decision-makers act consistently with stated direction under pressure. This instrument measures exactly that pressure and finds it wins 41% of the time, "often or always," on practitioners' own account.
• Purpose: O'Connor's "Lore is Mobile" names a decay mechanism — the sponsor who established the reliability posture leaves, and the organization's actual stance on reliability can shift without the group that implements it being told.

Momentum and Capability are not tagged: nothing here measures energy over time or organizational capacity, and the toil finding is a single-year self-report.

Relevance to Brandon's Positioning

This is the first external instrument in the base on SRE, error budgets or SLOs — the territory the 2026-09-18 challenge "Google is a special case — your best evidence doesn't transfer" names as a zero-file gap that the objection wins by default on. It does not close the challenge, but it changes its terms in Brandon's favour in three specific ways, and it removes one thing he was about to cite.

First, it substitutes primary evidence for the figure Stafford could not verify. The 2026-09-20 chase recorded a secondary attribution — "only 20% of SREs regularly use error budgets," credited to a 2021 SRE report — and flagged it as the strongest single piece of evidence for the objection if it verified. It does not verify here: the 2025 edition measures error-budget adoption nowhere, and the figure traces only to Isaac Sacolick's InfoWorld piece of 9 August 2021, absent from both Catchpoint's own 2021 highlights and APMdigest's summary of the same report. Do not cite the 20%. What this report supplies instead is better sourced and makes a smaller claim: in 2024, asked what their organization should adopt in the next twelve months, the top two answers from reliability practitioners are SRE itself (41%) and SLOs/XLOs (40%) — a decade after the practice was published, the discipline's own members are still nominating it as a thing to acquire.

Second, it puts a number on the mechanism's failure in the field rather than its absence. Brandon's two-variable diagnostic asks for a legible felt consequence and a sponsor with authority to install the measure. The 41%-under-deadline-pressure figure says the consequence is present and legible in these organizations — these are reliability professionals who know when the site is slow — and the trade still resolves against them, often or always. That is the objection's mirror image: the transfer problem is not that ordinary organizations lack Google's instrumentation, it is that having the measure does not settle whose number wins. That is Incentive Fragmentation stated in the vocabulary of the field Brandon is about to argue in front of.

Third, it names a decay term nobody in this base had. The diagnostic's variable (2) treats sponsorship as a binary — is there a sponsor with enough influence. O'Connor, seventeen years inside Google SRE, says sponsorship is mobile: the taste-makers move on and the group implementing the practice may not learn that the organization's posture changed. That is not Stafford's timing conjecture — which is about installing the measure early, before both sides have separate leadership and separate metrics — but it is adjacent and it is more testable: it predicts that interface measures decay on sponsor turnover, not on age. Practitioner testimony in a vendor report, not evidence; recorded as a named claim with an author attached, which is more than the conjecture has.

Weaker than what the base already holds, in one respect that must travel with every use. This is a 301-person opt-in vendor survey with one respondent per organization, fielded in mid-2024 and published in January 2025 — it sits well below the Faros telemetry and NBER 35275 in instrument quality, and it measures perceptions rather than behaviour. It earns its place because the base held nothing at all in this territory, not because it is strong.

Sources & Confidence

• Catchpoint, The SRE Report 2025 (7th edition), published 2025-01-13; survey fielded July–August 2024, n=301 — Moderate. Vendor-published, opt-in sample, self-report throughout; methodology (sample size, fielding window, demographics) is disclosed in full in the report's own "Demographics and Meta" section, read directly from the PDF.
• Catchpoint press release, "The SRE Report 2025: Highlighting Critical Trends in Site Reliability Engineering" (2025-01-13), for the publication date — Moderate.
• Isaac Sacolick, "How SLOs and error budgets improve app reliability," InfoWorld, 2021-08-09, carrier of the unverified "20% of SREs regularly use error budgets" figure — Low as a source for that figure; recorded here so the next run does not re-chase it.
• Catchpoint, "SRE Report 2021: The Highlights"; APMdigest, "2021 SRE Report Reveals Need to Expand Boundaries of Observability" — both checked for the 20% figure, neither carries it.

Filed by / Routing

Filed by: Stafford (daily brief, exploratory/verification slot, 2026-09-21) Routing: strategy / governance — STAGE-CANDIDATE for research-evidence; appended to the "Google is a special case" challenge in knowledge/thesis-challenges.md.

Tags: paper-2, practitioner, Strategic Disconnection, Technology Illusion, Purpose Source: arXiv preprint (cs.CL) — Hongliu Cao; accepted at REALM @ EMNLP 2026 URL: https://arxiv.org/abs/2609.14400 Date published: 2026-09-13 Date flagged: 2026-09-19

Summary

Agent benchmarks score an LLM agent against a written policy by comparing its trajectory to a single "gold" trajectory. Cao audits two domains of τ²-bench and argues the premise does not hold, because the policies themselves are defective: they contain "silence, ambiguity, or contradiction, admitting multiple defensible readings that a single gold trajectory cannot capture." Where a policy is silent or self-contradictory, more than one action is justifiable, and the benchmark records the agent's choice among defensible readings as an error.

The consequences Cao reports are about measurement reliability rather than agent capability. Affected tasks produce inconsistent scoring patterns that hit different models differently and reduce consistency across repeated trials of the same model — so a share of the variance that benchmark consumers read as a capability difference is a property of the policy text. Cao's second structural claim is a conjunction: exploitability emerges where "policy ambiguity and tool permissiveness" both exist. Neither alone produces the failure; an ambiguous rule with no tool to act on it is inert, and a permissive tool under an unambiguous rule is constrained.

The paper's conclusion is stated as a general bound rather than a benchmark fix: "policy specification quality sets the ceiling on evaluation quality."

Five Breakpoints Intersection

Strategic Disconnection: Cao's finding that written policy "admit[s] multiple defensible readings that a single gold trajectory cannot capture" is the mechanism of stated direction failing to determine operational behaviour while the measurement system continues to score as though it did — the organization believes it has specified what it wants, the evaluation reports alignment or its absence, and neither is reading the same rule the actor read.

Technology Illusion: The conjunction result — exploitability requires "policy ambiguity and tool permissiveness" together — is a direct statement that capability granted on top of an unresolved specification is where the failure lives, and Cao's title makes the diagnostic consequence explicit: the resulting failure is recorded as an agent error, so the organization fixes the deployment and leaves the broken condition exactly as it was.

Four Forces Intersection

Purpose: "Policy specification quality sets the ceiling on evaluation quality" is a bound on everything downstream set by clarity of intent alone — no amount of capability, monitoring or enforcement raises the ceiling that the written rule establishes.

Relevance to Brandon's Positioning

This is the second source in one week, from an unrelated literature, locating the binding variable in the precision of the control's wording rather than in its presence — the first being Wadi & Ma (arXiv 2609.17989), whose Study 2 shows a disclosure control inverting under the label "Promoted by the platform" (74.2% against a 54.2% no-label baseline) while holding almost completely under "Sponsored" (1.4%). Cao arrives at the same place from benchmark validity: ambiguity is what makes a rule exploitable, and the exploit is then misattributed to the agent. Neither is a marker-earning convergence — the designs and questions differ too much — but the agreement is worth stating, because it cuts against a framing this base has been assembling all week.

Specifically, it bounds the "counting controls measures presence, not effect" claim in the 2026-09-19 scan brief and paper-2 case file. That claim is right about counting and wrong about the implied remedy being unavailable. The remedy in both papers is the same and it is cheap: write the rule so it admits one reading. For Paper 2 this belongs with Claim 4 (the failure produces no visible artifact) as the thing that makes the artifact visible, and it is the first constructive finding attached to that claim rather than another instance of it.

For the consulting practice it changes asset 5's audit card rather than adding an essay — the question a client can answer in a first meeting is not "do you have an AI policy" but "how many defensible readings does this paragraph have, and who tested that." Cao also supplies the reason the answer is never known: the field's own evaluation apparatus treats the policy as ground truth, so every existing instrument scores the reader and none scores the rule.

Limits, and the first is scope. This is a benchmark-reliability audit of two domains of one benchmark suite, not an organizational study — no firm, no employee, no outcome, and the "policies" are synthetic evaluation artifacts written for customer-support simulation, not real corporate control language. The number of tasks, policies and models is not stated in the abstract or listing metadata. A workshop-accepted preprint. It establishes that specification defects corrupt agent evaluation; it establishes nothing about the ambiguity rate of real organizational policy, which nobody in this base has measured.

Sources & Confidence

• Hongliu Cao, "Policy Loopholes in Agent Evaluation: When Policy Ambiguity Masquerades as Agent Error," arXiv:2609.14400 (cs.CL), submitted 2026-09-13; accepted at REALM @ EMNLP 2026. Abstract and listing metadata read; full text not retrieved. Author affiliation is not stated on the arXiv listing. — Moderate. Unrefereed preprint with workshop acceptance; single author; no affiliation verifiable from the listing.
• Davood Wadi & Yu Ma, "Whom Do AI Agents Work For?," arXiv:2609.17989, 2026-09-16 (cross-reference; primary entry held in research-evidence). Study 2 cell values verified against the paper's own Table 13 by Stafford on 2026-09-19. — Moderate.

Instrument. Benchmark audit — manual inspection of policy text in two τ²-bench domains, plus re-scoring analysis across models and repeated trials. Data source: synthetic evaluation policies and agent trajectories, not organizational documents. Inference method: qualitative identification of silent/contradictory clauses, with scoring-inconsistency evidence as the consequence. Observation window: not applicable. Never cite as evidence about real corporate policy ambiguity.

Filed by / Routing

Filed by: stafford (daily brief, 2026-09-19 — exploratory slot) Routing: governance / strategy

Tags: paper-2, Strategic Disconnection, COUNTERARGUMENT Source: Administrative Science Quarterly — Warren Boeker (University of Washington) URL: https://doi.org/10.2307/2393919 Date published: 1997-06-01 Date flagged: 2026-09-18

READ DEPTH — declared before the finding

The paper itself has not been read. Citation verified against multiple independent secondary listings; the finding below is taken from the abstract as summarized in search results, not from the primary text. It must not become load-bearing in any draft until the paper is read. Filed now because it is prior art for a mechanism Stafford proposed to Brandon as novel on 2026-09-18, and the correction should not wait. Minor discrepancy across listings on the closing page (213–236 vs 213–237).

Instrument (per CLAUDE.md rule 9, as far as the abstract establishes it). Data source: movement of top managers across firms in the semiconductor industry. Inference method: executive migration events linked to subsequent product-market entry by the receiving firm. Observation window: 18 years, pre-generative and pre-internet, in one industry. Sample size not established from the abstract. Outcome variable is product-market entry — a strategic action, not a performance result.

Summary

Boeker examines whether the movement of senior executives between organizations produces strategic change at the receiving firm. Over an eighteen-year period in the semiconductor industry, the results "support the argument that executive migration brings managers into the organization with prior exposure to different products and strategies, which in turn is reflected in subsequent product-market entry decisions by the executive's new firm."

The mechanism is that an arriving executive carries the strategic repertoire of the firm they came from, and the receiving organization's subsequent strategy moves toward that repertoire.

Five Breakpoints Intersection

Strategic Disconnection: The paper establishes that a receiving firm's strategic direction shifts toward the products and strategies of an arriving executive's previous employer — direction imported from another organization's conditions rather than derived from this one's, which is a measured mechanism for how stated strategy comes loose from local circumstances.

Four Forces Intersection

No Forces tags — the paper has not been read. Honest nulls beat guessed tags (CLAUDE.md rule 2).

Relevance to Brandon's Positioning

This is prior art for the "imported executive playbook" mechanism, and Stafford proposed that mechanism to Brandon as a novel testable claim earlier the same day. The correction matters more than the find.

Brandon's observation from inside Google Cloud — that senior leaders "tend to run their own playbooks that they brought with them from previous companies," with Oracle and Microsoft playbooks prevalent — describes what Boeker measured in 1997. The mechanism is not new. What Brandon adds is not the phenomenon but the consequence he attaches to it: that the imported playbook is the leadership-layer instance of bolting on, chosen because redesigning for local conditions requires understanding and authority the arriving leader does not yet have. Boeker establishes that strategy follows the migrating executive; he does not claim it is a failure mode. That framing is Brandon's and it remains his.

The specific instrument proposed on 2026-09-18 still appears not to exist, and this is the genuinely useful result of the search. Boeker measures individual executive movement against strategic action. The adjacent literature measures top-management-team faultlines and CEO successor origin (inside / contender / outsider). Nothing located measures the count of senior hires drawn from an arriving leader's prior employer, which is the tighter proxy for a playbook being installed with its operators, and nothing compares high-import against low-import within outside hires — the design that would separate "imported playbook" from "arrived at a burning building."

And that confound is now confirmed rather than conjectured. The outside-versus-inside CEO literature reports that firms often become more profitable after appointing external CEOs specifically when the appointment follows a period of poor performance — so outside hiring is selected on prior trouble, exactly as Stafford warned. Any test that does not hold that constant will produce the predicted correlation for the wrong reason.

Sources & Confidence

• Boeker, W. (1997). "Executive Migration and Strategic Change: The Effect of Top Manager Movement on Product-Market Entry." Administrative Science Quarterly, 42, 213–236. doi:10.2307/2393919. — Moderate, and only because the citation is confirmed across independent listings; the paper is unread and the confidence is on the citation, not on the finding. ASQ is a top-tier peer-reviewed venue, which is why it is filed at all.
• Adjacent, located but not verified and deliberately not filed: the executive-migration-and-international-M&A literature (prior-firm acquisition experience shaping current-firm host-country selection, with the effect stronger where prior and current industries are similar); the top-management-team faultlines literature (Georgakakis, Greve & Ruigrok, and a CEO-successor-origin study on Chinese firms); and Barry, J.W., "Executive Mobility in the United States, 1920–2023" — a long-run executive mobility dataset that is the most likely data source if this test is ever built. All unread; recorded so a future slot does not re-chase them.

Filed by / Routing

Filed by: stafford (consult session, 2026-09-18) Routing: strategy

Tags: paper-2, paper-5, practitioner, Incentive Fragmentation, Process Friction, Momentum Mirage, Purpose, Commitment Source: Primary practitioner testimony — Brandon Freitag (Stafford consult session) URL: null Date published: null (engagement periods not established) Date flagged: 2026-09-18

PUBLICATION CONSTRAINT — read first

Governed by CLAUDE.md hard rule 13. Four client organizations are named here. Three of the four are recorded with unflattering outcomes, and all were commercial relationships of Brandon's employer. No client name leaves this file without Brandon's explicit clearance.

One of them is already solved. The KeyBank engagement appears in papers/paper-5-before-it-breaks.md (Discipline 2) as "a CISO at a financial institution" — mechanism intact, client unidentifiable. That is the template for all four, and it is Brandon's own precedent, not a rule imposed on him.

Why this file matters more than its length suggests

This is the closest thing to a sample Brandon personally owns. His team covered hundreds of engagements at an average win rate of around 60% — multiple organizations, observed over a period, through one consistent analytical lens, by the person who built the lens. Nothing else in this base has that shape.

Two limits bound every use, and they are severe. (1) The team sat in pre-sales, so "we didn't get to stick around fully until the customer landed on success" — outcomes are partial, second-hand, or learned informally after the fact. (2) There is no systematic record. The cases below are those Brandon recalls, which is a recall-biased sample of a real population, not the population. Treat "hundreds of engagements" as context for the pattern claim and never as a denominator.

The pattern claim

The majority of customers faced some breakpoints, and in Brandon's assessment this is "very common in large enterprises that have been around for a while."

The age-and-size conditioning is the substantive claim here, and it is the same one the framework makes about AI-native versus legacy organizations — the greenfield organization designs the function in; the established one must redesign or bolt on, and bolts on. That Brandon reached the same conditioning independently from hundreds of pre-generative cloud engagements is worth noting, because it means the AI-native claim has a prior shock behind it from his own practice rather than only from the literature.

The cases

Nokia — datacenter exit, one-year target

Wanted to exit multiple datacenters within a year. Last Brandon heard, it had been three years and they still weren't done.

KeyBank — datacenter takeout, CISO veto at launch

The CISO vetoed approval of any changes to a newly launched datacenter takeout until all of his security mandates were met. He had engaged another consulting firm in parallel and produced the veto as a surprise at launch. Already published de-identified in Paper 5, Discipline 2, where the incentive analysis is also given: his metrics were tied to security posture, a successful migration required accepting transition risk, the two were incompatible, and "nobody had a conversation about it before the kickoff."

Read alongside the DaVita security impasse, this is the same mechanism twice — a security leader whose sole lever is veto and whose measure is posture, blocking a migration that required accepting transition risk. Two independent instances of the CISO pathology, in different sectors, one of them already in print.

Box.com — dev environment migration, under 90 days

Development environments fully migrated to cloud in under 90 days. Brandon's stated cause: it took commitment from the top of the customer organization to pull off. The base's clearest positive case with a duration attached.

A south-central bank — the negative case

Appeared aligned to avoid the issues: a very clear vision communicated broadly through the organization, with multiple senior stakeholders aligned, attributed by Brandon to strong CEO and CIO direction. He also recalls a small number of sessions in which all key C-level and VP stakeholders were deeply engaged and working toward a common vision.

His own summary, and the framework should adopt it unchanged: it is possible, and it is rare.

Outcome, confirmed 2026-09-18: as far as Brandon is aware it was largely a success. No timeline available, so it cannot join the duration table below. The negative case therefore holds — it is a case the framework declines to claim, with a known good outcome, and it is load-bearing in the provenance challenge (knowledge/thesis-challenges.md, 2026-09-18).

The timeline pattern — the most usable finding in this file

Two independent enterprise programmes with one-year executive timelines ran to three-plus years. The one case that moved fast moved very fast, and the cause Brandon names is top-level commitment. This base is starved of outcome variables and has almost no transformation durations measured from inside with a stated expectation to compare against. These are three.

What it is not. Three recalled cases are not a distribution, the 1-year/3-year coincidence across two of them is not a rate, and the Box contrast is a single case whose distinguishing condition is Brandon's attribution rather than a measurement. The honest published form is a shape, not a statistic — enterprise programmes given one-year executive timelines took three years or more, except where commitment was held at the top — and even that needs the sample described as recalled cases from one practitioner's portfolio.

Five Breakpoints Intersection

Incentive Fragmentation: The KeyBank CISO's metrics were tied to security posture while the migration required accepting transition risk — two incompatible objectives inside one organization with no conversation before kickoff, which is the breakpoint in its cleanest published form and is already in Paper 5.

Momentum Mirage: Nokia's one-year datacenter exit reaching year three without completion is the gap between announced ambition and sustained execution, measured in elapsed time rather than in sentiment.

Four Forces Intersection

Purpose: The south-central bank's distinguishing feature was a clear vision communicated broadly, with senior stakeholders aligned — Purpose present as a designed and maintained condition rather than as a stated intention, and it is the only case in the portfolio Brandon names as having avoided the pattern.

Commitment: Box.com's sub-90-day migration is attributed by Brandon directly to commitment held at the top of the customer organization, which is the Force operating as the discriminating variable between the fast case and the slow ones.

Evidence gaps

• No file in either repository mentions Nokia or KeyBank, and Box.com appears once. The four cases Brandon most readily recalls from hundreds of engagements were entirely unrecorded before today.
• There is no systematic record of the engagement portfolio. If any artifact survives — win/loss notes, QBR decks, account retrospectives — it would convert recall-biased anecdote into something with a denominator, and it is the single highest-value acquisition available to this base. See Open Questions.

Open questions for Brandon

1. Does any written record of the engagement portfolio still exist and is any of it retrievable? Even a partial list with outcomes would transform this from recall into data. Note the obvious constraint: material belonging to a current or former employer is not yours to extract, so the realistic version is probably your own recollection structured against a list of accounts rather than anything retrieved.
2. Of the engagements you recall, roughly what share had a named executive sponsor who stayed through delivery? Sponsor turnover is Paper 3's mechanism 1 and it has no evidence behind it anywhere in this base.
3. ANSWERED 2026-09-18 — largely a success as far as Brandon is aware; no timeline. Remaining question: is there any way to establish the timeline? It is the only case in the portfolio where alignment was present from the start, so its duration is the natural comparator for the one-year-to-three-year pattern and would turn a three-row table into a genuine contrast.
4. How many of the recalled cases involved a security veto specifically? Two are already in hand (DaVita, KeyBank). A third would make the CISO pathology a pattern in Brandon's own portfolio rather than a pair.

Sources & Confidence

• Brandon Freitag, Stafford consult session, 2026-09-18. Low as evidence, Moderate-to-High as illustration. Recalled cases from a pre-sales vantage with partial outcome visibility; no contemporaneous record; recall-biased selection from the underlying population.
• papers/paper-5-before-it-breaks.md, Discipline 2 — the KeyBank case, published de-identified with its incentive analysis. High, and the only element of this file already cleared for public use in that form.

Filed by / Routing

Filed by: stafford (consult session, 2026-09-18) Routing: strategy / market

Tags: paper-2, paper-3, practitioner, Strategic Disconnection, Incentive Fragmentation, Process Friction, Technology Illusion, Momentum Mirage, Purpose, Commitment, Momentum, COUNTERARGUMENT Source: Primary practitioner testimony — Brandon Freitag (Stafford consult session) URL: null Date published: null (periods of employment not established) Date flagged: 2026-09-18

PUBLICATION CONSTRAINT — read first

Governed by CLAUDE.md hard rule 13. Captured named and at full fidelity because the analysis needs the facts. Nothing here is publication-ready.

Named-individual risk is higher in this file than in any other in the base. It contains a named, currently-serving senior executive of a named employer with a characterization of his leadership, and it contains claims about Brandon's current employer including an assertion about teams being dissolved. None of that may appear in any published form, de-identified or otherwise, without Brandon's explicit clearance on the specific passage — de-identification does not protect a claim where the role plus the sector plus the period identifies one person. The publishable unit is the mechanism; these specifics are private evidence for it.

The model for doing this correctly already exists in Brandon's own work: the KeyBank engagement appears in papers/paper-5-before-it-breaks.md (Discipline 2) as "a CISO at a financial institution" with the employer removed and the mechanism intact. That is the template.

Provenance — how the framework was actually built

Brandon did not arrive at DaVita with the Five Breakpoints, the Four Forces, or a full understanding of the GRIP levels. He arrived mid-MBA, after an original career in manufacturing and process engineering, a shift to software engineering (Motorola), and twenty years in technical sales (Oracle, VMware, Google). In his words, not a blank slate — but without the framework.

The Five Breakpoints and the GRIP levels gained clarity while he was at DaVita, through three inputs running together: observing the failures directly, re-reading his prior career for parallels, and MBA coursework supplying the theory to draw those parallels with.

This should be recorded plainly rather than tidied, because its epistemic status is specific and a reviewer will ask. The framework was abducted from one organization, then retrospectively validated against the author's own prior experience. It was not derived independently and then tested. That is how most practitioner frameworks are built and it is not disqualifying — but a framework constructed by reading back through one's own history will tend to find itself everywhere it looks, and the honest defence is to say so rather than to imply a cleaner derivation. See the challenge logged in knowledge/thesis-challenges.md on 2026-09-18.

The five at DaVita — the originating case

Incentive Fragmentation. Each peer carried a single distinct metric: VP of Dev — speed. Ops — reliability. Security — risk and breach management. Finance — fiduciary risk management. And the detail that explains the whole decision structure: the CIO held security and finance closest, as "the things that could get him fired the fastest."

This is the most analytically valuable sentence in the account and it is not in the framework. It says metrics are not held with equal weight — they are ordered by how quickly each can end the holder's tenure, and that ordering, not the metric list, is the operative incentive structure. It predicts which conflicts get resolved and which are left standing, and it explains why security's veto was decisive while the interface between Dev and Ops was not. Candidate framework addition: the termination-risk ordering of metrics. Stafford's formulation from Brandon's observation; see Open Questions.

Strategic Disconnection. Visible inside Brandon's own organization, not only across it: an individual several levels down stating "I don't care about the why — just tell me what I need to get done", and a director reporting to him who persisted in a direction contradicting the one Brandon was driving, whom Brandon ultimately had to fire. Across the organization: the CIO needed a new patient system to meet specific business needs, and "much of the middle to low part of the organization didn't understand and definitely didn't agree/align on how that should happen." Brandon's own summary of the failure: the importance of the patient-system effort was not communicated clearly and repeatedly through the entire organization.

Process Friction. The mechanism was team topology. Teams were formed by technology layer — a VM team, an OS team, a database team, a security team, a monitoring team. Standing up a system for South America composed of five VMs took months, because there was "only a collection of teams with narrow scope and no interest on how things fit together." Rework was, in his word, immense.

Technology Illusion. Partly visible in his selling career and fully visible at DaVita from the buying side: he was accountable for delivering against business problems, while "99% of salespeople talked about their product." Products were a necessary component; the real difficulty was making it all work, which most often became an operational failure — people, process, culture.

Momentum Mirage. Effort concentrated on selecting the products and the solution and on the launch, after which it "faded into the background." The clarifying moment was Simon Sinek's treatment of the Diffusion of Innovation (Rogers' curve): all effort is aimed broadly at a launch, while only around 16% of an organization is initially interested in change, so the approach has to be different. On the Cloud programme specifically: the launch was broad instead of focused, and the operational change — people, process, culture — needed to sustain it was absent.

The five at Google — the more valuable half

Brandon went to Google to help customers avoid the Five Breakpoints and found the breakpoints inside Google as well. This is the framework's strongest available answer to its most natural objection — that it is a theory of badly-run organizations — and it is the opposite of what a Google alum is expected to say.

Strategic Disconnection. Google Cloud needed to contribute cash flow to Alphabet, which was facing regulatory headwinds and intensifying competition in Ads — and that was never made clear. The Cloud CEO, arriving from a product-development leadership role at Oracle, "instead of knitting together a vision and strategy… defaulted to having the key 3 product areas/products that were important to sell each year." Each area then formed its own interpretation, and the resulting efforts conflicted and competed with one another.

Incentive Fragmentation. The customer-facing organization was split between pre-sales and post-sales, each with its own executive and its own metrics, with partners attached to post-sales as a third split. Sales closed the signature and passed it "over the wall." Post-sales then took weeks to redo discovery and write a statement of work, and sometimes had to report that what sales had sold was not possible to deliver.

Process Friction. Sales organizations formed by vertical and by region, then overlay organizations for technical and product depth, then solutions engineers attempting to knit products into repeatable solutions, then further overlays — each with its own incentive. Taking a customer from "Hello" to success was confusing internally and externally, "because each team was only focused on a small chunk and nobody knew how to do the entire process."

Technology Illusion. Externally, "every single day": the CEO pointed at products, sales leadership built incentives to push those products, and customers were told the technology was what they needed. Internally, and this is live right now: the answer offered is to use the range of Google AI products, with "no training, reuse, etc. or any operational thought being given to how to do things well."

Momentum Mirage. New internal websites and tools are announced; those who launch them are rewarded; the work then fades, because maintaining it is not rewarded.

Absent or failed? — ANSWERED 2026-09-18, and the answer is worse than either option

Brandon's answer: Google is a large organization with "a success story in SRE but a failure story in Google Cloud." Google Cloud had access to the information and to the people who made SRE work — and "sold" SRE to customers as a Google Cloud differentiator — but chose, by omission or commission, not to implement it itself.

That is not "the remedy lived in another building." The remedy was marketed by the failing organization as a reason to buy from it. The knowledge was available, the practitioners were available, and the practice was commercially valuable enough to put in front of customers. What did not happen was adoption by the sellers.

Candidate name, Stafford's formulation from Brandon's observation: the unpracticed remedy. An organization that sells a practice as its differentiator and does not run it internally. This is distinct from Technology Illusion — nobody here mistook a product for the work; they understood the work well enough to sell it and still did not do it — and it is not a capability or awareness gap. It is a choice under incentives, which makes it a sharper instance of the framework's central argument than a straightforward Illusion case would be. Not yet a pattern: one organization, one observer.

What would count as a second instance — asked by Brandon, 2026-09-18, and the two candidates establish different things. (a) Another Google organization outside the sales org is not independent — same company, same leadership, same incentive environment — so it cannot make this a pattern under the framework's own escalation rules. It would establish something else worth having: that the unpracticed remedy is a property of the company rather than a quirk of one division, which strengthens the case and sharpens the mechanism. (b) A different company makes it a candidate pattern, and that is what the framework needs.

Where to look for (b), in descending order of value. Brandon's own client portfolio is the best source and has not been searched for this — hundreds of engagements, and any client that sold a practice it did not run internally is a second instance he already has. Then: AI vendors selling AI transformation while their own adoption is shallow, which is live, on-thesis for Paper 2, and testable against public statements versus disclosed practice; consultancies selling transformation and change management, measured against their own; and security vendors with public breach histories.

The mechanism Brandon offers for why, and it is the most portable idea in this section. Thomas Kurian and various senior Google Cloud leaders "tend to run their own playbooks that they brought with them from previous companies" — Oracle and Microsoft playbooks prevalent. He also notes recurring "Google Cloud's culture needs to change" pronouncements from leaders over time.

This closes a loop with Brandon's own principle from earlier in the same session. He argued that designing a function in at the outset is easy while retrofitting requires rare quantities of will, so bolting on becomes the default. An imported executive playbook is that default operating at the leadership layer. A leader arriving in an established organization either redesigns their approach for local conditions — which requires understanding what is already there and the authority to change measures — or runs the playbook they already have, which requires neither. The playbook is the bolt-on.

And the "culture needs to change" pronouncements are the tell: leadership naming culture as the problem while being the vector by which a different culture is imported. Maps to Strategic Disconnection — the operating model in use is inherited from another company's strategic conditions rather than derived from this one's, which is why proximity to SRE was irrelevant.

Testable implication — first version proposed, objected to by Brandon, and redesigned 2026-09-18. The first version was organizations filling senior roles externally show more imported-playbook failure than those promoting internally. Brandon's objection: no company publishes its failures; they are excused and papered over.

The objection is right about the mechanism and wrong about the design, and the corrected version is better than both. The design never needed self-reported failure: both variables are already public. Outside-versus-inside appointment is disclosed in proxy statements, bios and press releases; outcomes are observable as returns, forced turnover, divestitures and restatements. This is precisely asset 4's logic — the ex-ante measure is already filed — applied to a different layer, and it is the fourth instance this week of a solution already present in Brandon's own files.

But Brandon's skepticism points at a real problem that is worse than publication bias, and it is selection rather than suppression. Outside CEOs are disproportionately hired when the organization is already in trouble. A raw outside-versus-inside comparison therefore confounds imported playbook with arrived at a burning building, and would produce the predicted correlation for entirely the wrong reason. That, not disclosure, is why this literature is hard.

Redesign, and it addresses both problems at once. Outside-versus-inside is a proxy for unfamiliar, not for imported. A tighter and still-public measure: the count of senior hires from the leader's prior employer in the N months after arrival. That measures the playbook being installed with its operators, which is far closer to the mechanism, and it is visible in public bios and appointment announcements. It also supplies the discriminator: compare high-import against low-import within outside hires only, which holds the burning-building selection roughly constant. Secondary observable proxies: reorganizations mirroring the prior company's structure, and adoption of its vocabulary and frameworks.

Caution, per CLAUDE.md rule 9. Executive-bio data is résumé-panel-adjacent, the genre this base is already over-weighted in. This is a different use — tracking executive co-movement rather than inferring organizational structure — so it does not compound the Résumé-Panel Consensus problem, but the instrument block must say so explicitly rather than leaving a reader to notice.

Still unevidenced. This base holds no outside-CEO-versus-inside-CEO literature and no instrument of this design.

Publication note. Thomas Kurian is named here and is a real, currently-serving, identifiable executive; the characterization is one former employee's observation. This passage may not be published in any form, de-identified or otherwise — "the CEO of a major cloud provider who arrived from Oracle" identifies one person. The mechanism (imported executive playbooks; the unpracticed remedy) is publishable; the instance is not.

The superseded inference, retained so the record shows the correction. Every breakpoint above sits in the commercial and go-to-market organization, not in production engineering. The mechanisms this base has been treating as Google's contribution — error budgets, blameless postmortems — live in SRE and production engineering, a different part of the company. So the provisional reading is that Google's mechanisms were absent where the breakpoints appeared rather than tried and failed there — which would mean the breakpoints and the remedies never met, and neither constitutes evidence about the other. Stafford inferred absent and that was wrong — Brandon's answer above establishes that the knowledge and the people were available inside Google Cloud and the practice was sold by it. The correct reading is available and not adopted, which is a stronger finding than either option originally offered.

The stewardship finding, and it is the sharpest thing in the account

Brandon reports that every team inside Google he is aware of — including his own, currently — that carried any passion or capacity for arguing that success required more than the technology or the product has been dissolved. His own vision had to be taken external, to Substack and workthatholds.com, "because leadership wasn't able to and/or interested in the type of change and success that I was pointing to."

If accurate, this is Paper 2's Stewardship Gap observed not as an absence but as an active removal — the function that would address the breakpoints being repeatedly eliminated by the organization that needs it. That is a materially stronger claim than "nobody owns it," and it supplies the Stewardship Gap with a mechanism it currently lacks: stewardship roles are not merely unstaffed, they are selected against.

Handle with maximum care. This concerns Brandon's current employer and his current position, it is a single insider's account with no external corroboration, and "every team I know of" is a sampling claim from one vantage. It is the highest-legal-risk statement in this base. It is recorded because it is important, not because it is publishable.

Where Brandon contributed to a breakpoint

Twenty years in technical sales, contributing to the Technology Illusion — in his words regrettably, though not purposefully.

This is undersold in his own telling and it is the strongest credibility instrument available to him. The author of a framework naming Technology Illusion as a primary failure pattern spent two decades on the selling side of it, and can describe the mechanism from inside because he operated it. That is the same move as Paper 3's public retirement of the 70% figure, executed at career scale, and it converts a framework that could read as an indictment of other people into one whose author is implicated in it.

He separately notes that, as a manager at Google, once it became clear that his views on change, success and leadership were not being accepted and that he was wanted to execute the product-sales path, "it was the beginning of my exit because I couldn't bring myself to operate that way." Recorded as context on his departure rather than as an instance of causing a breakpoint — it describes a refusal, not a contribution, and should not be presented as the answer to the question.

Five Breakpoints Intersection

All five, at both organizations, with a named mechanism for each — set out above rather than repeated here. This is the only source in the base evidencing all five breakpoints in a single organization, and it does so for two.

Four Forces Intersection

Purpose: "I don't care about the why — just tell me what I need to get done", and a Cloud CEO substituting an annual list of three products to sell for a vision and strategy, are the same failure at two altitudes — direction that was never made to mean anything to the people executing it.

Commitment: The CIO's ordering of metrics by termination risk is Commitment structured by personal exposure rather than by stated direction — follow-through allocated to whichever obligation could end his tenure fastest.

Momentum: Rewarding launch and not maintenance, at Google, is a designed decay — Momentum failing because the incentive system stops paying at exactly the point where sustaining effort begins.

Evidence gaps this account exposed

• Conway's Law and Team Topologies: zero files across both repositories (conway 0, team topologies 0, skelton 0). The DaVita Process Friction story — teams formed by technology layer, five VMs taking months — is the canonical illustration of a literature the base does not hold at all. This is the nearest named prior art to the Process Friction breakpoint and it should be acquired before Paper 2 goes further.
• Diffusion of Innovation is held (rogers 3, sinek 6, diffusion of innovation 4, early adopter 5), and Rogers 2003 is already cited in Brandon's published papers — but the base nowhere records that the Sinek framing of it was the originating moment for Momentum Mirage. The citation is present and its role in the framework's construction is not.
• GRIP appears in 4 files; its relationship to the Five Breakpoints and its own provenance are not established here and should be.

Open questions for Brandon

1. ANSWERED 2026-09-18 — see Absent or failed? above. Neither: available and not adopted. Follow-on question now open — is the unpracticed remedy a pattern or a single observation? Any second organization that sold a practice it did not run would make it one, and it would be a strong addition to the framework.
2. The termination-risk ordering of metrics — is this already part of GRIP or the Four Forces, or is it a genuine addition? It is the most portable idea in the DaVita account and it is not in the published framework as far as this base records.
3. Did the framework arrive as five, or accrete? Which came first, and was there ever a sixth that was dropped? A framework that arrived whole and one that converged over time carry different burdens of explanation.
4. GRIP — provenance, definition, and its relationship to the Five Breakpoints. It is referenced in 4 files and defined in none that this session located.

Sources & Confidence

• Brandon Freitag, Stafford consult session, 2026-09-18 — direct testimony, recorded in his terms with verbatim phrases marked. Low-to-Moderate as evidence, High as illustration and as provenance. Uncorroborated, retrospective, from the framework's author about his own employers. Two of the three organizations discussed are former employers and one is current.
• papers/paper-5-before-it-breaks.md, Discipline 2 — the KeyBank engagement, already published in de-identified form. High, and the working template for translating any of this material.

Filed by / Routing

Filed by: stafford (consult session, 2026-09-18) Routing: strategy / personal

Addendum 2026-10-10 (Stafford, /consult) — the Sinek mechanism was dropped from the published framework, and open question 3 now has a candidate answer

The provenance note above says the citation is present and its role is not. Today's reading sharpens that, and the sharper version is a content loss rather than an attribution gap.

This account records the clarifying moment as Sinek's treatment of Rogers' curve: "all effort is aimed broadly at a launch, while only around 16% of an organization is initially interested in change, so the approach has to be different." That sentence contains a mechanism and a prescription. The mechanism is that initial interest is a small minority; the prescription is that a broad launch is therefore the wrong shape.

Neither survives into the published Momentum Mirage. Checked against knowledge-base/five-breakpoints-source.md today: the Momentum Mirage section describes reinforcement decaying after leadership attention shifts, and its self-check is "what would happen if you stopped actively managing this transformation for 90 days?" There is no 16%, no adopter distribution, and no claim that the launch should have been narrow. The breakpoint names the symptom the Sinek insight explains and drops the explanation. On the Cloud programme this account even records the diagnosis in the original terms — "the launch was broad instead of focused" — and the published framework cannot say that, because it has no concept of launch shape.

Why this matters now. Greg Satell (Fast Company, 2026-10-07) independently arrives at the same prescription and states it as a design rule: shape networks rather than opinions, secure a local majority first, never operate from the minority. A competitor is publishing the prescription that Brandon's own origin story contains and his published framework omits. Brandon has the better diagnosis of why transformations fade; Satell has the operative instruction Brandon arrived at first and did not print.

Open question 3 — "Did the framework arrive as five, or accrete? ... was there ever a sixth that was dropped?" — now has a candidate, and it is a precondition rather than a sixth failure mode. All five breakpoints are states of a transformation already underway; none can be detected before launch, and none says do not launch yet. Satell's second prescription — "change isn't always possible. The time isn't always right... pushing for change when it is unlikely to succeed is more likely to put you deeper in the hole" — names a failure the framework has no slot for: launching into conditions that cannot carry it. Whether that is a sixth breakpoint, a Force precondition, or a separate readiness layer is Brandon's call, and it bears directly on growth edge #2's unanswered request for an ex-ante measure — a readiness instrument is exactly an attempt to detect this before launch.

One more thing worth recording, because it is the first of its kind in this base. Satell's third prescription is, structurally, a Commitment measurement design. The Four Forces have no instrument anywhere in this base (asset 3, The Missing Instrument; growth edge #2, unanswered since 2026-08-29; The McAdam standard challenge, logged today). A dilemma action works by taking a stated commitment, constructing an act consistent with it, and forcing the holder to either honour it or visibly break it — which converts an unobservable into an observable. His stated design condition is the operative part: "key to the success of a dilemma action is that it is seen as a constructive act rooted in a shared value." n=1, uncontrolled, author as protagonist — worthless as evidence and genuinely interesting as instrument design.

Filed by: Stafford (/consult, 2026-10-10). Routing: strategy (the dropped mechanism is a Paper 2 / asset 12 input), personal (open question 3 has a candidate answer).

Tags: paper-2, paper-3, practitioner, Incentive Fragmentation, Process Friction, Commitment, Purpose Source: Primary practitioner testimony — Brandon Freitag (Stafford consult session) URL: null — see Provenance and evidentiary status below Date published: null (period of employment not established; see Open Questions) Date flagged: 2026-09-18

Provenance and evidentiary status — read before any use

This is not an ingested source and must not be treated as one. It is first-person, uncorroborated testimony from the author of the framework it supports, recorded from a Stafford consult session on 2026-09-18 and not verified against any external record.

PUBLICATION CONSTRAINT (Brandon's direction, 2026-09-18) — read before any use in a draft. "There is my own story with all of its details and then there is that which is publishable — to avoid sounding like grievance and/or to avoid legal issues." This file is captured at full fidelity deliberately, because the analysis needs the facts. It is not publication-ready and no part of it may be treated as such. Nothing here reaches a draft, a post, a talk or a client deck without Brandon's explicit clearance on that specific passage.

Highest-risk material, named so it is never moved by accident. Three classes, in descending order of exposure. (1) Statements about identifiable individuals' honesty or motive — that the CIO "portrayed them as green to those outside of IT", and that Brandon was given "promises that weren't true." These concern real, role-identifiable people at a named employer and must never appear attached to any identifiable party in any form. (2) Named-employer performance claims — ~3 outages per week; three years and tens of millions on a failing implementation; the security posture denying the patient system access to patient data. Defensible as personal recollection, damaging as published assertion, and unverifiable by any reader. (3) Role-identifiable individuals — "the CIO", "the VP of Dev", "the lead security architect" are each a single identifiable person at a named company in a known period.

The working convention going forward: capture named and full internally; translate once, at drafting time, into a de-identified form (sector, scale, role-shape, no employer); and treat the mechanism as the publishable unit and the specifics as the private evidence for it. Do the translation deliberately rather than under deadline — the grievance register and the legal exposure both enter when specifics are carried across late.

Three further constraints follow.

1. Never stage this to research-evidence. That pipeline requires a reachable URL and feeds the public dashboard. An unverifiable self-report from the framework's own author entering the evidence index would contaminate exactly the thing the base's anti-patterns exist to protect. This file is Stafford-local, permanently.
2. It is an existence proof and an illustration, never a measurement. No sample, no outcome variable, no comparison group, no external corroboration. It shows a mechanism is possible and gives it a concrete shape. It cannot establish frequency, magnitude or generality.
3. The author has an interest in the conclusion. The case, as told, supports the teller's published framework and casts a former employer's decisions unfavourably. That does not make it false — practitioner testimony is legitimate evidence and Brandon's is unusually structured — but any public use must disclose the relationship and should prefer the reading in which the other parties acted reasonably (see The reading to prefer).

Origin — why both organizations matter, in Brandon's compression

Recorded verbatim in substance, 2026-09-18: DaVita helped him see the Five Breakpoints, because all five existed there. Google was where he went to help customers avoid the Five Breakpoints — and he then discovered the breakpoints existed at Google as well.

Both halves carry weight and the second is the more valuable. The first establishes DaVita as the framework's originating case, which the corpus nowhere records. The second is the framework's best available defence against its most natural objection — that Five Breakpoints is a theory of badly-run organizations. The exemplar that supplies this base's only working model of interface accountability also carried the breakpoints, observed from inside by the same person. That is the opposite of what a Google alum is expected to say, which is exactly why it is worth saying, and it means the framework's claim is about organizations as such rather than about weak ones. Detail not yet elicited — see Open Questions.

Summary

The operator is constant across both cases. Brandon Freitag ran DevOps and SRE at DaVita, in a senior director role filling a lane a VP had held (books/built-to-be-replaced.md), having come from Google where the equivalent practices originated. Same person, same practices, two organizations, opposite outcomes — which is the structure that makes this worth recording.

Google, as he describes it. Security was not a central gatekeeper holding sole accountability. The security organization stayed central to designing the approach, while enablement work focused on getting every team to understand and own its part, integrating the function into the system "so it wasn't an added check or bolt-on." Those who circumvented or ignored the processes and systems were held responsible. Alongside this, the organization practised blameless postmortems: where something happened that the system did not catch or should have caught, the focus was on fixing the system so it could not repeat, and the individual was not blamed.

DaVita, as he describes it. The CIO measured Dev on speed and Ops on stability — directly opposed metrics — and inserted Brandon's organization between the two, expecting it to fix both. The CIO had, in Brandon's account, no interest in establishing a new culture or new measures. The result was not reconciliation: his group "ended up being the blaming point for both orgs to point to." Separately and consistently, the same CIO aligned a primary goal and metric per role and declined to reconcile the conflicts when those goals proved mutually exclusive.

Correction recorded 2026-09-18, same session — the first version of this file got the DaVita case materially wrong and the corrected facts change its mechanism. Neither function was performing. Ops averaged roughly three outages per week, and Ops was the CIO's own former organization. Dev had spent three years and tens of millions of dollars attempting to implement the new system and was failing. The CIO "portrayed them as green to those outside of IT." Brandon was recruited specifically because Google Cloud "had the tools and approach needed to reach success" — but in his account the CIO and the VP of Dev "had bought into the Technology Illusion — that it was just the tech that was needed." What the CIO wanted delivered in one year was not achievable on that basis; Brandon left, and the organization took over three years to get there.

Brandon's account of why Google's version worked is deliberately not "Google is special." He attributes it to Benjamin Treynor Sloss having both a real consequence (global scale, where outages were not excusable) and the influence to install the mechanism — specifically the error budget. His claim about why it does not transfer is that most organizations have a weaker felt consequence (though he notes tolerance for downtime has shrunk substantially over time) and, more importantly, have incentive fragmentation that they do not see the need to address.

Five Breakpoints Intersection

Incentive Fragmentation: The DaVita case is this breakpoint in its clearest observed form — Dev rewarded for speed and Ops rewarded for stability, with no measure spanning the handoff between them, so the conflict between the two objectives had no owner and no instrument, and was absorbed by the team standing at the interface rather than resolved by anyone holding either metric.

Process Friction: Inserting a third organization between two functions with opposed measures did not remove the friction at the handoff; it relocated the friction onto the inserted party, which had responsibility for the joint outcome and no authority over either input — structural friction that blocked execution while both contributing functions reported healthy numbers.

Four Forces Intersection

Commitment: In Brandon's account the CIO's follow-through failed structurally rather than attitudinally — the decision that was declined was not "should we do DevOps" but "should the measures change," and refusing the second while funding the first is the shape of stated direction not surviving contact with what it would cost.

Purpose: Both organizations had clear local purposes (ship faster; stay up) and no stated purpose for the interface between them, so the inserted function had no definition of success it could hold either party to.

The analytical structure this case produced

Developed in the 2026-09-18 session, from the case rather than imposed on it. The case is Brandon's; the formalizations below are Stafford's and should be attributed accordingly.

1. The accountability/sanction grid. Two variables — whether accountability for a joint outcome is concentrated or distributed, and whether a real sanction is available to the party holding it — produce four cells:

Brandon has occupied the blame-sink cell directly and observed the department-of-No cell. The blame-sink cell is the worst of the four and the least written about, plausibly because its occupants leave rather than publish.

2. The CISO pathology is structurally predicted, not cultural. A party holding sole accountability for an outcome everyone produces has exactly one lever — veto — because it cannot make anyone do the work, only stop things. It also violates the third condition below: the veto reduces the CISO's personal exposure, so the sanction becomes an instrument for managing the sanction-holder's risk rather than the organization's. "Department of No" follows from the accountability structure and would reproduce with a different person in the chair.

3. Three design responses to a joint outcome, not two.

• Decompose per role. Every box gets a number; the interface goes unowned. — DaVita.
• Measure the interface with a group sanction. — Google SRE error budgets.
• Dissolve the interface by embedding the function into the work. — Google security enablement, in Brandon's phrase "not an added check or bolt-on." There is no handoff left to measure because the function is inside the work rather than downstream of it. This is what "shift left" denotes once the jargon is removed, and the base holds nothing on it.

4. Three conditions for an interface measure that works. (a) It spans the interface rather than the boxes either side; (b) the consequence falls jointly on everyone at the interface; (c) the party imposing the consequence gains nothing from imposing it. Condition (c) is the one most often missed and is what separates a shared sanction from an armed turf war.

5. The blameless boundary. Brandon's two statements — blameless postmortems, and those who circumvent are held responsible — are not in tension; together they locate the sanction precisely. Blameless applies to failure inside the system; accountability applies to stepping outside it. The reason this is the right line is that circumvention is the one act that is both individually attributable and destructive of the information system itself — you cannot be blameless about the act that makes failure invisible.

6. Blamelessness is a monitoring decision, not a cultural one. Punishing incidents does not reduce incidents; it reduces reported incidents. Blamelessness trades the ability to sanction individual failure for the preservation of information about failure. Stated that way it is a design choice with a cost rather than a value statement — and it has an unstated precondition: you can only sanction circumvention if you can detect it. Where bypass is invisible, the only enforceable sanction is on outcomes, which is blame, which destroys the reporting. Organizations earn the right to stop blaming people by being able to see what happened, which inverts the usual framing of surveillance and humane practice as opposites.

7. Consequence must be unarguable, not merely severe. Google's consequence was immediate, external and not renegotiable — the site is down. Most organizational failures produce consequences that are delayed, diffuse and attributable to whoever one chooses, which is what happened at DaVita. The transferable test is therefore not is the consequence large enough but is the failure legible enough that nobody can argue about whether it occurred. A sanction attached to an arguable failure is an opening negotiating position, not a sanction.

8. The concealment mechanism — and this item is a correction. Stafford's first reading of this case was that incentive fragmentation is invisible from the top by construction: the decomposition that creates an interface conflict is the same decomposition that hides it from anyone above, so the person who can see it has no instrument and the person who could act has no visibility. Brandon rejected that reading on the facts and he is right. Ops was running ~3 outages a week; Dev was three years and tens of millions into a failing implementation. Nothing was green. The CIO could see both, and Ops was his own former organization.

What actually happened is not invisibility but impression management with a personal stake. The failures were visible inside IT and were represented as healthy outside it, by an executive whose own prior organization was one of the two failing functions. That is a choice under incentives, not a knowledge failure — and it is a materially stronger finding than the one it replaces, because it is Paper 3's published thesis operating on its own author (see Relevance to Brandon's Positioning).

The general mechanism is retained, demoted to an untested hypothesis, and must not be attached to this case. It remains plausible that in organizations where the decomposed metrics are healthy, an interface conflict is structurally invisible from above. This case is not evidence for it. Anyone reaching for that argument needs a different instance — one where the roll-up genuinely looked fine — and this base does not currently hold one.

9. The timing hypothesis — testable, and currently unevidenced. Treynor Sloss installed the error budget before there was an established Ops organization with its own leadership and its own defended metric; Brandon arrived at DaVita after Dev and Ops were two organizations with two opposed measures and two leaders whose performance depended on them. Stated first as a gate and corrected by Brandon to a gradient, which is the right shape. The gate version — the window closes — implies designed-in-or-fail. Brandon's version: designing it in at the outset is always easier; retrofitting remains possible but requires sponsorship, desire, will and effort in quantities that are rare. So the prediction is not that late attempts fail but that they succeed only where an unusual amount of will is present, which is why bolt-on becomes the default rather than a considered choice.

Why bolt-on is the default, mechanically. Redesign requires someone to give up a metric, and the party standing at the interface has no authority to make that happen. Bolt-on is the only option available to someone without the authority to change measures — the same structure as the CISO veto, where the available lever determines the strategy.

Brandon's generalization, which is larger than DevOps and connects to the main thesis. This is why AI-native organizations move quickly and legacy ones do not: the greenfield organization designs the function in, while the legacy organization must rethink and redesign or bolt on — and bolts on. The DevOps timing claim is therefore not a separate hypothesis; it is an instance of the AI-native-versus-legacy claim the framework already makes, which means DevOps history is available as a prior shock supplying evidence for it.

Still unevidenced, and checkable against the DORA/Accelerate programme, which this base does not hold.

10. Every actor behaved rationally given the measure they were held to, and the failure was guaranteed anyway. This is the essay's spine and its defence against reading as grievance.

Five actors, five rational responses to their own measures, one certain failure, and no villain. That is papers/paper-3-designed-to-stall.md's "These are not character flaws in individual leaders" instantiated five times inside one organization.

And the security requirement was not a compromise but a contradiction. The Cloud was to run the core patient system; the security posture denied it access to core patient data. The programme was defined as impossible by a model nobody owned the job of reconciling with it — incentive fragmentation producing a logical impossibility rather than merely friction. It is also bolt-on as the default in its purest form: the cloud was made to conform to the datacenter's perimeter model because the only person positioned to object had veto power and no redesign authority.

Relevance to Brandon's Positioning

Paper 3 already contains the argument and does not contain the case, and that is the actionable finding here.

papers/paper-3-designed-to-stall.md line 32: "The strongest predictor of whether a leader does the design work is whether the structure penalizes skipping it." That is the consequence/sanction argument, already published, and Brandon's Treynor Sloss reading is a worked instance of it.

The same paper, line 76, warns that the function it proposes must avoid becoming "a delivery arm for leadership-defined initiatives without the mandate or standing to challenge senior leaders," which "over time… comes to accommodate existing silos rather than address them." That is a description of the role Brandon occupied at DaVita, published as a warning, with the author's own case as its best available illustration and no reference to it.

So the gap is not a missing argument. It is that the strongest evidence for an argument already in print is the author's own career, and it appeared in this knowledge base exactly once before today — in books/built-to-be-replaced.md, on an unrelated point about vendor relationships.

And the match is far closer than "the argument is published" — Paper 3's Career calculus section describes this case in detail without knowing it. "A real transformation effort means visible risk: political conflict, short-term disruption that looks like underperformance before it looks like progress and the possibility of being the person associated with a failed initiative." Then the alternatives: "a technology investment that demonstrates activity without requiring process change" — which is precisely what Google Cloud plus a senior hire from Google was purchased to be — and "The board's visibility into the difference is limited in the early years," which is the concealment, in Brandon's own published words.

So Paper 3 predicted Brandon's own experience more accurately than Stafford's analysis of it did, and the paper does not cite the case. This is the fourth instance in five days of a needed citation already sitting in Brandon's own corpus doing a different job — and the most striking, because what was already written is not a supporting reference but the correct analysis of the author's own career, stated generally.

The outcome numbers are rare and should not be lost: a one-year executive expectation against an actual three-plus years, with the practitioner leaving in between. This base holds very few actual transformation durations measured from inside, and none previously with a stated expectation to compare against.

The sharpest finding of the whole session, and it is about the corpus rather than the case. papers/paper-5-before-it-breaks.md closes with Eight questions before the launch. Question 8 reads: "Have you confirmed you have been given the standing to do this work, or is the first task renegotiating the role itself?" That is precisely the question Brandon says he lacked the insight to ask. Question 2 — "Have you mapped every leader with veto power and resolved the incentive conflicts before the announcement?" — is the security architect, the VP of Dev and the Change Board, in order. Discipline 2's worked example is a CISO whose metrics were tied to security posture blocking a datacenter migration at a financial institution; Discipline 4 is "test organizational readiness before you deploy the technology," which is the Technology Illusion purchase exactly.

So the corpus already contains the complete diagnosis of this case, as general advice, with the case that produced it unwritten. That is a different problem from the retrieval pattern logged 09-14 → 09-18: the papers carry the lessons and not the evidence. It is also precisely what growth edge #3 in knowledge/brandon-development.md has been asking for since 2026-08-29 — for each mechanism, what measured case would show it working or failing — and the answer for Paper 5's disciplines 2, 4 and 8 has been in Brandon's own history the entire time.

Secondary positioning value: the case supplies Paper 2's Claim 1 discussion with a concrete instance of the unbundling mechanism (a structure imported for one function while the function that made it work was left behind), and it is the practitioner anchor for the dissolve the interface option, which the framework does not currently name.

Evidence gaps this case exposed

Base coverage checked across both repositories on 2026-09-18. All counts are of files mentioning the term, not of substantive treatment.

• DevOps / SRE literature: effectively absent. SRE 0, site reliability 0, error budget 0, SLO 0, service level objective 0, Forsgren 0, shift left 0. The eight DORA hits are the Faros finding plus an unrelated Deloitte executive named Dora. This is the best-documented case of interface accountability working at scale and the base does not hold it.
• Safety-science / just-culture literature: half present. Edmondson 10 and psychological safety 17 — the speaking-up half. James Reason 0, Sidney Dekker 0, Swiss cheese 0, crew resource management 0, near miss 0, Safety-II 0, just culture 1. The error-versus-violation distinction — item 5 above — is canonical in this literature and roughly forty years old. Brandon derived the boundary from practice without the citations; a reviewer from a safety-critical industry would find this in the first paragraph.
• Team-production economics: absent. Alchian 0, Demsetz 0, Holmström's moral hazard in teams 0, residual claim 0, free-riding 0, Lazear 0, tournament 0.
• Implementation fidelity: adjacent but unconnected. implementation science 9, Helfrich 9, Weiner 9, Caci 7 — all recruited against the unfalsifiability challenge. fidelity 4, core component 1, active ingredient 0. That literature's concept of adopting an intervention without its active ingredients is the precise diagnostic for what happened at DaVita, and in this base it is doing an unrelated job. Third instance this week of a needed citation already being in Brandon's own files.

The reading to prefer

Revised after the 2026-09-18 correction. The earlier version of this section advised the reading in which the CIO could not see the conflict. That reading is factually false and must not be used — he could see it, and he represented it as healthy to people outside IT.

The grievance risk is unchanged and remains the whole risk of publication. A piece about a CIO who wronged the author dies on contact, because a reader with no independent access cannot adjudicate it.

But the non-grievance reading is not "he was blind" — it is that his incentive was universal even though his choice was bad. His own former organization was one of the two failing functions. Naming the problem properly meant surfacing that. Buying technology — Google Cloud, and a senior hire from Google — demonstrated decisive action without requiring him to do that. The reader identifies not with blindness but with the reluctance, which is far more common and far more uncomfortable to recognize in oneself.

Brandon has already written this reading, generally, in Paper 3, and it is the sentence the piece should be built on: "These are not character flaws in individual leaders. They are predictable responses to a system that rewards visible action over structural design, distributes consequences across a longer time horizon than tenure and treats incremental motion and substantive change as equivalent for purposes of board reporting and analyst calls."

Same discipline on the one-metric-per-role account: state what the CIO did, not why.

Open questions for Brandon

1. ANSWERED 2026-09-18, and the answer is a third option neither framing anticipated: he asked, and was neither refused nor supported — he was told to escalate, by the only person who could have decided.

   Brandon entered believing he had full sponsorship from the CIO to do what was necessary. The "Aha!" moment was an impasse with the lead security architect (reporting to the CISO, who reported to the CIO). Security's position: Cloud fully isolated from the internet; all traffic backflowing through the datacenter; every concentric "moat" around datacenter systems preserved; Cloud treated as the outermost ring of the existing security model. The consequence, in Brandon's words: the Cloud "which was supposed to run the core patient system — wouldn't have access to the core patient or business data."

   Two further blockers, each a rational output of its owner's measure. The VP of Dev refused to implement code reviews, producing deployment errors and multi-hour deployments. Ops insisted on retaining a Change Board that took three to four weeks to approve and schedule a change. Against those three constraints a one-year target was not achievable; hitting it required the metrics, culture and process to change.

   Brandon took it to the CIO — leading with the security blocker — and the CIO said: "You'll need to escalate that." In Brandon's reading: he would not act, and his expectations would not move either. So Brandon lost, and left; the organization reached the outcome in over three years.

   Why that sentence is the most valuable artifact in this case. The CISO was the CIO's own direct report and the security architect was two levels down — the conflict was entirely inside the CIO's own organization. Telling Brandon to escalate it means directing him to take it to the person who just spoke. The CIO was the escalation path, and used the vocabulary of escalation to decline to be it. It sounds like enablement and functions as refusal, and the deadline does not move. That is how the blame-sink cell is created — not by an org design at hiring, but by a single non-decision at the moment of impasse: accountability retained on the appointee, authority withheld, expectation unchanged. It also instantiates papers/paper-4-the-ai-mirror.md line 33 — "Either someone owns the escalation path or the system runs without one until something breaks visibly" — discovered at the impasse rather than at the design.

2. Interview conditions — Brandon's own assessment, recorded in his terms. He states he was "fooled" during the interview: he "should have pressed harder," did not then have the insight he has now, and was also given promises that were not true. Two readings are available and this base cannot adjudicate between them on one account. (a) Misrepresentation. (b) A sincere promise made before its price was known — "full sponsorship" only became expensive when it meant overruling his own CISO on a patient-data security posture inside a healthcare company, which is genuinely hard, and he may not have anticipated it when he said yes. Reading (b) is the more useful and more defensible in print, and it generalizes: sponsorship promises are unreliable not because sponsors are dishonest but because the promise is made before the price is known. Note against it that the same CIO demonstrably managed external representation (portraying Dev and Ops as green outside IT), so a mix is plausible.

3. Dates and duration for both roles. Not recorded here because they were not established and must not be guessed. They matter substantively: the timing hypothesis (item 9) depends on when Brandon arrived relative to Dev and Ops calcifying into separate organizations.

4. How did Google avoid blameless degrading into consequence-free? The known failure mode is excellent postmortems accumulating while the same incident class recurs. The discriminating question is narrow — did the action item ship, and did anyone check? If Google avoided this, the mechanism that prevented it is the transferable part, not the blamelessness.

5. Was there any measure spanning the Dev/Ops interface at DaVita at any point, even an unenforced one? Determines whether this is a case of no interface measure or an interface measure with no sanction, which are different cells in the grid.

Sources & Confidence

• Brandon Freitag, Stafford consult session, 2026-09-18 — direct testimony on both the Google and DaVita cases. Low-to-Moderate as evidence, High as illustration. Uncorroborated, retrospective, from an interested party; see Provenance above. Recorded because the alternative is that it stays unwritten, not because it clears the base's ordinary bar.
• books/built-to-be-replaced.md — Brandon's own manuscript, corroborating the DaVita senior-director role filling a lane a VP had held. High for the role fact; the manuscript is pre-publication and status should be confirmed before any public quotation.
• papers/paper-3-designed-to-stall.md, lines 32 and 76 — published, and the anchor for the positioning claim above. High.
• A Stafford claim corrected during the session, recorded so it is not repeated. Stafford asserted that shared interface measures necessarily reintroduce free-riding. That was wrong as stated and Brandon's DevOps/SRE counterexample was correct. The accurate claim is narrower: the impossibility applies to shared outcomes under budget balance among the agents alone, and the standard resolution — a group penalty imposed by a party who collects nothing from it — is precisely what an error budget implements. The corrected version is what appears in item 4 above. (The formal result is cited from Stafford's own knowledge, is not held in this base, and must be verified against the primary text before any published use.)

Filed by / Routing

Filed by: stafford (consult session, 2026-09-18) Routing: strategy / personal

Tags: paper-2, Incentive Fragmentation, Strategic Disconnection, Commitment, Purpose, COUNTERARGUMENT Source: arXiv (econ.TH) — Shota Ichihashi (Queen's University), Fei Li (University of North Carolina at Chapel Hill), Dihan Zou (University of North Carolina at Chapel Hill) URL: https://arxiv.org/abs/2609.10499 Date published: 2026-09-09 Date flagged: 2026-09-18

Summary

A robust-contracting model of peer monitoring in knowledge work. A principal leads agents who work on "different but related tasks"; by working on their own task each agent acquires information useful for evaluating peers. The authors solve for contracts under which effort by every agent is the unique rationalizable outcome — a stronger requirement than ordinary equilibrium, meaning no agent can talk himself into shirking by any consistent belief about what others will do.

The central result is a derived structure: "The optimal contract induces a hierarchy of monitoring authority: agents easier for the principal to monitor directly are assigned greater authority to evaluate other agents." Monitoring runs in a chain — the principal first disciplines the agent whose task she can see most easily, and that agent then carries evaluative authority over the next. The ordering principle is monitorability from the principal, not seniority, expertise, task centrality, or information volume.

Two further results bound it. "Task diversity improves information value but weakens peer monitoring" — spreading work across more distinct problems yields better information but degrades agents' ability to judge one another, and "this tension distorts optimal public task assignment toward the principal's core expertise." The principal narrows what the organization publicly works on so she can keep monitoring it. The proposed remedy is "confidential random assignment," which "relaxes this distortion by decoupling information production from the perceived monitoring structure" — i.e. the binding constraint turns out to be the monitoring structure agents perceive, not the one that exists.

Five Breakpoints Intersection

Incentive Fragmentation: The model derives that evaluative authority over peers is allocated by how easily the principal can see an agent's own task rather than by that agent's competence to judge others — so the person with the most influence over a colleague's assessed performance is whoever the boss happens to have the clearest view of, which is a structural misallocation of judgment produced by the incentive design itself and not by anyone's intent.

Strategic Disconnection: The result that the monitoring/diversity tension "distorts optimal public task assignment toward the principal's core expertise" is a formal derivation of the work portfolio bending toward what the leader can observe rather than toward what the organization should be doing — vague or narrowed purpose arising as an artifact of the monitoring architecture, which is a mechanism for Strategic Disconnection this base has not previously held in derived form.

Four Forces Intersection

Commitment: Hierarchical monitoring authority is the entire mechanism by which the contract makes effort the unique rationalizable outcome — follow-through is produced by the observation chain, not by disposition, so removing or re-ordering the chain removes the structure that holds commitment under pressure.

Purpose: The task-assignment distortion is Purpose being set by observability — the principal's ability to monitor determines what the organization publicly takes on, so clarity of intent is downstream of the monitoring design rather than upstream of it.

Relevance to Brandon's Positioning

This is the counterargument to paper-2 Claim 1 that today's scan was reaching for, and it is a materially better citation than the one the scan used. Claim 1 holds that traditional hierarchy was an information routing protocol, not a purpose/commitment/capability system. This paper derives a hierarchy whose ordering principle is monitorability — that is, a hierarchy that exists in order to enforce commitment, with routing as a by-product. That is a direct hit on Claim 1's separation, and it lands because the object is an actual hierarchy of authority.

By contrast, the paper the scan conceded Claim 1 on does not support the concession. Cheng, Dogan & Yildirim (NBER WP 34259, staged in research-evidence today) model a three-worker sequential workflow in which each worker observes his predecessor. That monitoring is lateral — peer-to-peer along a process handoff — and the paper's own "middle worker" is a task position, a limit the scan's KB entry states correctly and then argues past. Lateral peer observation inside a workflow is not evidence about what a hierarchy was for, so conceding Claim 1 on that citation invites the reply "that is a process, not an org chart" and makes a correct concession look careless. Concede Claim 1 — cite this paper for it.

The genuinely new argument is a rival prediction about the delayering wave, and Brandon is positioned to make it first. Everything in this base that derives hierarchy from coordination cost (Farach; and empirically Ewens & Giroud, Babina et al.) predicts hierarchy shrinks as AI drives coordination cost down — the flattening thesis. This paper's ordering principle predicts something different: since AI changes monitorability unevenly, making some work radically more observable (logged, traced, diffable) and some far less (absorbed into an opaque model call), hierarchy should re-sort rather than shrink, with authority migrating toward whoever's work AI made legible. Flattening and re-sorting are distinguishable predictions, which is worth more to asset 2 (The Flattening Bet) than another confirmation would be.

Sources & Confidence

• Ichihashi, S., Li, F. & Zou, D., "Monitoring Hierarchies in Knowledge Production," arXiv:2609.10499 [econ.TH], v1 submitted 2026-09-09. Abstract read verbatim from the arXiv listing; model structure, the monitoring-chain result and the incentive-cost expression read from the arXiv HTML rendering. — Moderate. arXiv preprint, nine days old, not peer-reviewed; treated as this base treats other econ.TH preprints (cf. Farach, Bartolucci & Vivo), which is as argumentative machinery rather than evidence.

Instrument discipline (recorded before the finding, per CLAUDE.md rule 9). Data source: none — this is an analytical model with no dataset, no sample and no estimation. Inference method: robust mechanism design; optimal contracts solved under iterated elimination of strictly dominated strategies, tasks located on the unit interval with Markov/Poisson state dynamics. Observation window: not applicable. It states what an optimizing principal would design, never that any organization does it. It therefore cannot be cited as evidence that hierarchies are in fact ordered by monitorability — only that a hierarchy ordered that way is what optimal monitoring design produces. Two further bounds: the model never mentions AI or automation, so the re-sorting prediction above is Brandon's extension and must be labelled as his; and "hierarchy" here is an ordering of evaluative authority, which need not coincide with reporting lines.

Filed by / Routing

Filed by: stafford (daily brief, 2026-09-18) Routing: strategy

Tags: paper-2, Incentive Fragmentation, COUNTERARGUMENT Source: Academy of Management Journal URL: https://journals.aom.org/doi/abs/10.5465/255378 Date published: 1975-12-01 Date flagged: 2026-09-17

Summary

Kerr's article collects illustrations — from society generally and from organizations in particular — of reward systems that pay off for one behavior while the rewarder hopes for a different one. The organizing claim is behavioral and stated in the abstract: most organisms seek information about which activities are actually rewarded and then do, or appear to do, those things, often to the exclusion of activities that are not rewarded. The title is the mechanism: the gap is not between what the organization says and what people believe, but between what the organization says it wants and what its reward system pays for.

Kerr's account of why such systems persist is structural rather than moral. Reward structures get built on criteria that are measurable and visible, while the objectives organizations actually care most about tend to be subjective and hard to measure — so organizations reward what they can see most clearly even when it is not what matters most. His running example is academic promotion and tenure, where the stated route is research and teaching and the paid route is publication, with the predictable effect on how faculty allocate time. He also notes the less comfortable possibility: that in some cases organizations are in fact getting the behavior they want and are simply not honest about wanting it.

Primary not read in this session. The openly posted copy (ou.edu) is an image scan with no text layer and the publisher page returns 403. The summary above is assembled from the published abstract and multiple independent secondary descriptions (Semantic Scholar, AOM listing, the 2017 Academy of Management Perspectives retrospective). Nothing here is quoted verbatim from the article body, and no finding is claimed from inside it.

Five Breakpoints Intersection

Incentive Fragmentation: Kerr's title thesis — organizations reward A while hoping for B, and people do A — is the breakpoint's mechanism named and published in a top management journal in 1975, thirty years before Brandon's framework states it. His stated cause, that reward structures are built on measurable and visible criteria while the objectives that matter are subjective and hard to measure, is also the framework's account of why incentive misalignment survives contact with a well-communicated strategy.

No other breakpoint maps. This entry is filed as prior art on a breakpoint, not as empirical evidence for one: it is an illustrative essay with no sample, no measurement and no test.

Four Forces Intersection

No Forces tags. The primary was not readable in this session (image scan, no text layer; publisher 403), and this base does not guess Forces tags on unread sources.

Relevance to Brandon's Positioning

This is a liability entry before it is an asset one. Paper 2's case file recorded on 2026-09-17 that Technology Illusion has named prior art in Toyama's Law of Amplification and resolved to concede the priority and claim narrower ground. Kerr establishes that the same exposure exists for Incentive Fragmentation, and it is older, more canonical and far more likely to be known to any management-trained reader — AMJ 1975 is standard organizational-behavior curriculum in a way a 2011 iConference paper is not. A reader who has met Kerr will read Incentive Fragmentation as Kerr relabeled unless Brandon says otherwise first.

The constructive half is specific. Kerr diagnoses nothing: he illustrates a folly and explains why it recurs, and his practical advice bottoms out at look at what you are actually paying for. He offers no instrument, no taxonomy, and no account of how incentive misalignment interacts with the other four failure patterns — because he is not describing a system of failures, he is describing one. Brandon's published self-checks (see knowledge-base/five-breakpoints-source.md and papers/paper-5-before-it-breaks.md) are the diagnostic layer Kerr lacks, and asset #4 in knowledge/asset-suggestions.md proposes scoring exactly this construct from comp plans and proxy statements — which is Kerr's folly made measurable from residue rather than asserted from anecdote. The defensible claim is extension and instrumentation, not discovery, and it should be written that way before a reviewer writes it.

Sources & Confidence

• Kerr, S., "On the Folly of Rewarding A, While Hoping for B," Academy of Management Journal 18(4), 1975, pp. 769–783. DOI 10.5465/255378. High as a source (top-tier peer-reviewed management journal, heavily cited), with an absolute caveat on content: the primary was not read in this session and the entry claims nothing from inside the article body. Issue month (December 1975) is the standard citation for AMJ 18(4); day-level precision is not established and the day is set to the 1st by convention.
• Open scan (no text layer): https://www.ou.edu/russell/UGcomp/Kerr.pdf — Moderate (unofficial host, unverified fidelity to the published text).
• Kerr, S., "40 Years (and Counting): Steve Kerr Reflections on the 'Folly'," Academy of Management Perspectives (2017) — unread; logged as a citation-chase lead only. Title verified via ResearchGate listing; content not established.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-17) — own exploratory slot, prior-art sweep on the framework's own breakpoints. Routing: strategy

Tags: paper-2, Momentum Mirage, Strategic Disconnection, COUNTERARGUMENT Source: Harvard Business Review URL: https://hbr.org/1995/03/leading-change-why-transformation-efforts-fail-2 Date published: 1995-03-01 Date flagged: 2026-09-17

Summary

Kotter's article reports observations of more than a hundred companies attempting significant change and argues that transformation efforts fail for a small number of recurring, avoidable reasons rather than for idiosyncratic ones. He enumerates eight errors: failing to establish a sufficient sense of urgency; failing to form a strong enough guiding coalition; lacking a clear vision; under-communicating the vision; failing to remove obstacles that block the new vision; failing to plan for and create short-term wins; declaring victory too soon; and neglecting to anchor the changes in the organization's culture.

The seventh error is the one this base is filing the article for. Kotter's argument is that visible early progress gets read as completion — clear signs of improvement are announced as victory, the effort is treated as finished, and the change never reaches the depth that would make it stick. His corrective is that leaders of successful efforts use early wins as leverage on larger structural problems rather than as evidence the work is done; he reports peak change arriving around year five, several years after the first visible wins. The eighth error is the same problem at the end of the sequence: change that is not anchored in culture and in succession decisions decays once the pressure is removed.

Primary not read in full in this session. The HBR page is paywalled and two openly posted PDF copies (UCSF; Lighthouse Consultants) are image scans or subset-font reprints with no extractable text layer. The account above is assembled from the abstract-level record and multiple independent secondary summaries of the eight errors. Nothing is quoted verbatim from the article body, and no finding is claimed from inside it.

Five Breakpoints Intersection

Momentum Mirage: Kotter's seventh error — declaring victory too soon, in which visible early progress is mistaken for completed change and the effort stops short of the depth that would make it durable — is the breakpoint's mechanism named in the most widely read change-management article ever published, in 1995. The framework's "appearance of progress without actual movement" and Kotter's "declaring victory too soon" are the same observation, and his is thirty years prior.

Strategic Disconnection: Kotter's third and fourth errors — proceeding without a clear vision, and under-communicating the vision that exists — are the antecedent of the breakpoint's "vague purpose producing the illusion of alignment," though Kotter frames the failure as insufficient transmission of a vision rather than, as Brandon's Paper 5 does, as agreement that survives only because the goal is too unspecific to be disagreed with.

Filed as prior art on two breakpoints, not as empirical evidence for either: the article is a practitioner synthesis of unstructured observation with no disclosed sample frame, no measurement and no comparison group.

Four Forces Intersection

No Forces tags. The primary was not readable in this session (paywalled; open copies have no text layer), and this base does not guess Forces tags on unread sources. Noted for a future read: the eight errors are prescriptive-sequence claims and would, if read, most plausibly implicate Momentum and Purpose.

Relevance to Brandon's Positioning

This is the most dangerous single entry in this base for Brandon's priority claim, and it is dangerous precisely because it is famous. Toyama (2011) is obscure enough that conceding it reads as scholarship. Kotter 1995 is the article every executive sponsor and every MBA-trained reader has already met — "declaring victory too soon" is not a passage a reviewer has to find; it is a phrase in general circulation. Momentum Mirage will be read as Kotter's seventh error renamed unless Paper 2 addresses the lineage first.

What survives the concession is real and is worth more than the priority claim. Kotter offers a prescriptive sequence — eight things to do in order — and no diagnostic: nothing in the article tells a leader which of the eight errors their own organization is currently committing, and the eight are framed as a checklist of leadership omissions rather than as a system of interacting structural conditions. Brandon's framework is the inverse: five failure patterns with per-breakpoint self-checks that can be run against an organization as it stands, plus a mechanism layer (Four Forces) explaining why each one recurs. Kotter says what to do; Five Breakpoints says what is wrong. The other half Kotter has no account of at all is speed — his own timeline is years, with peak change at year five, which is the strongest available support for the genuinely period-shaped half of the thesis: nothing in the 1995 canon describes these failure patterns operating at machine speed, because in 1995 they could not.

Sources & Confidence

• Kotter, J. P., "Leading Change: Why Transformation Efforts Fail," Harvard Business Review 73(2), March–April 1995, pp. 59–67. High as a source (HBR is on the high-credibility list; the article is among the most cited in the change-management literature), with an absolute caveat on content: the primary was not read in this session and the entry claims nothing from inside the article body. Citation (volume 73, issue 2, pages 59–67, March–April 1995) cross-checked across three independent bibliographic records.
• Open copies, both without extractable text: https://fcm.ucsf.edu/sites/g/files/tkssra17611/files/Kotter_WhyTransformationEffortsFail.pdf and http://www.lighthouseconsultants.co.uk/wp-content/uploads/2010/08/Kotter-Leading-Change-Why-transformation-efforts-fail.pdf — Moderate (unofficial hosts, fidelity unverified).
• Kotter, J. P., Leading Change (Harvard Business School Press, 1996) — the book-length version; unread, logged as a lead. Note that the book and the article are frequently conflated in citation; this base cites the 1995 article.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-17) — own exploratory slot, prior-art sweep on the framework's own breakpoints. Routing: strategy

Tags: paper-2, practitioner, Technology Illusion, Process Friction, Momentum Mirage, Capability, Commitment, COUNTERARGUMENT Source: Automatica (Journal of IFAC), Pergamon Press / International Federation of Automatic Control URL: https://doi.org/10.1016/0005-1098(83)90046-8 (open copy read in full for this entry: https://ckrybus.com/static/papers/Bainbridge_1983_Automatica.pdf) Date published: 1983-11-01 Date flagged: 2026-09-16

Date note: Automatica Vol. 19, No. 6, pp. 775–779, 1983. A print issue carries no day-level date; the month is the issue month and the year is certain. Received 16 December 1982, revised 23 May 1983; originally presented at the IFAC/IFIP/IFORS/IEA Conference on Analysis, Design and Evaluation of Man-Machine Systems, Baden-Baden, September 1982.

Summary

Lisanne Bainbridge (Department of Psychology, University College London) argues that automating an industrial process expands rather than eliminates the problems associated with the human operator. Her framing claim is stated in the introduction: the growth of interest in human factors among engineers "reflects the irony that the more advanced a control system is, so the more crucial may be the contribution of the human operator." The paper is an argumentative review of the process-control and flight-deck literatures, not a new empirical study; its evidence is other people's experiments, cited by name.

Two ironies organize the argument. The first is about who designs the residual job: "the designer who tries to eliminate the operator still leaves the operator to do the tasks which the designer cannot think how to automate," which "means that the operator can be left with an arbitrary collection of tasks, and little thought may have been given to providing support for them." The second is about monitoring: "the automatic control system has been put in because it can do the job better than the operator, but yet the operator is being asked to monitor that it is working effectively." Bainbridge then derives why that assignment can be structurally impossible rather than merely difficult: "if the decisions can be fully specified then a computer can make them more quickly, taking into account more dimensions and using more accurately specified criteria than a human operator can. There is therefore no way in which the human operator can check in real-time that the computer is following its rules correctly." The operator can therefore only judge at a meta-level whether the machine's decisions are "acceptable" — and "if the computer is being used to make the decisions because human judgement and intuitive reasoning are not adequate in this context, then which of the decisions is to be accepted? The human monitor has been given an impossible task."

Supporting mechanisms, each attributed. Manual skill decays without use, so "a formerly experienced operator who has been monitoring an automated process may now be an inexperienced one" (Edwards & Lees, 1974). Long-term process knowledge "develops only through use and feedback about its effectiveness," so operators taught it theoretically "will probably not understand much of it… and they will not remember much of it." Sustained monitoring of a quiet channel is bounded by vigilance: citing Mackworth (1950), "it is impossible for even a highly motivated human being to maintain effective visual attention towards a source of information on which very little happens, for more than about half an hour," from which Bainbridge concludes "it is humanly impossible to carry out the basic function of monitoring for unlikely abnormalities." Enforcing attention by requiring a log fails because "people can write down numbers without noticing what they are." She reports, via Ephrath (1980), "a study in which system performance was worse with computer aiding, because the operator made the decisions anyway, and checking the computer added to his workload." Her closing irony inverts the training assumption: "it is the most successful automated systems, with rare need for manual intervention, which may need the greatest investment in human operator training."

Five Breakpoints Intersection

Technology Illusion: Bainbridge's second irony is the breakpoint stated as a design law rather than a post-mortem — a control system is installed because it outperforms the operator, and the same operator is then made responsible for monitoring it, which she shows is not a control at all where the machine's rule-following cannot be checked in real time ("the human monitor has been given an impossible task"). This is the "human in the loop" governance assertion refuted forty-three years before the agent-governance market began asserting it.

Process Friction: The first irony names the mechanism by which automation manufactures structural friction — "the designer who tries to eliminate the operator still leaves the operator to do the tasks which the designer cannot think how to automate," producing "an arbitrary collection of tasks" for which no support was designed. The residual human job is a leftover, not a designed role, which is precisely the condition under which execution blocks.

Momentum Mirage: The log-keeping observation — attention is enforced by requiring the operator to record readings, and "people can write down numbers without noticing what they are" — is the appearance of oversight fully decoupled from oversight, measured at the level of the individual compliance artifact. Every completed review record in an enterprise AI workflow is the same artifact.

Four Forces Intersection

Capability: Bainbridge's central mechanism is capability decay caused by the deployment itself — skills "deteriorate when they are not used," so the person the design relies on for recovery is systematically less able to recover the longer the system runs well. Capability here is not a stock the organization has or lacks; it is a flow the technology consumes.

Commitment: The stated direction ("a human supervises the automation") survives on paper and is abandoned in operation for reasons the paper names structurally rather than morally — vigilance limits at roughly half an hour, and the impossibility of real-time rule-checking. Follow-through fails without anyone deciding to stop following through.

Relevance to Brandon's Positioning

This is the paper that tests — and breaks — the Previous-Shock Archive's own falsifiable prediction, and it is therefore more valuable to Brandon as a problem than as a supporting citation. The pattern (research-evidence knowledge/emerging-patterns.md, 2026-09-11, advanced 09-16) predicts that paper-2's mechanism-shaped claims are answerable from the pre-AI information-technology literature while the period-shaped Claim 3 — "each breakpoint manifests differently and more dangerously in AI-native orgs" — should be unreachable, because "its 'more dangerously' clause is a comparison to the current technology and no pre-AI paper can supply it." Bainbridge supplies it in her second paragraph: the more advanced the control system, the more crucial and less supportable the human contribution. The comparison is monotone in automation capability, which makes it reachable from any point on the curve, including 1983. The prediction fails, and the scan's own less comfortable alternative reading — that the pipeline finds whatever it goes looking for — is now the live one.

It also does three things for the published position. (1) It is a third literature (human factors / cognitive systems engineering), not economics of organization and not clinical informatics, so it satisfies the pattern's stated escalation rule on the same day it falsifies the pattern's prediction. (2) It supplies the prior statement of the Agachan supervisor result the 2026-09-15 window is built on — "a supervisor pays for itself when it can verify and becomes a liability when it can only opine" is Bainbridge's real-time-checkability condition, restated for agents; the window's pitch is stronger citing a 1983 antecedent and weaker if a reader supplies it first. (3) It answers the research-evidence scan's 2026-09-16 Open Question from the other direction: the scan observed that Walsh's 96% decline rate proves load is sufficient but not necessary for abandonment of a review step, and Bainbridge/Mackworth names the necessary condition it was looking for — a low event rate, not a high workload. Weaker than the base's randomized entries as evidence (it is a review essay with no data of its own); stronger than anything in the base as a statement of the mechanism.

Binding discipline: this enters on the Bloom/Garicano/Sadun/Van Reenen footing per sources.md caution 2 — a pre-AI result, cited as a pre-AI result, never as an AI finding. Bainbridge's systems are deterministic rule-followers; her real-time-checkability argument rests on the machine applying specified rules faster than a human can trace them, which is a different unverifiability from the one generative AI creates (plausible output with no rule to trace at all). The conclusion rhymes; the mechanism does not. Say so before a referee does.

Sources & Confidence

• Bainbridge L. "Ironies of Automation." Automatica 1983;19(6):775–779. DOI 10.1016/0005-1098(83)90046-8. High — IFAC's journal, peer-reviewed, ~1,800 citations by 2016 and rising. Full text read for this entry; all quotations verified against it.
• Mackworth (1950), Edwards & Lees (1974), Wiener & Curry (1980), Ephrath (1980) — cited within Bainbridge and not independently verified here. Low as standalone citations. Ephrath (1980) in particular appears in her reference list as "Verbal presentation. NATO Symposium on Human Detection and Diagnosis of System Failures, Roskilde, Denmark" — it is not a retrievable publication and must never be cited as a source by this base, only as a pointer.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-16 — Wednesday falsification sweep, exploratory slot) Routing: strategy (the 09-15 supervisor-pattern window needs this citation); governance (the "human in the loop" control claim)

Tags: paper-2, practitioner, Incentive Fragmentation, Technology Illusion Source: International Journal of Human-Computer Studies (Academic Press / Elsevier) URL: https://doi.org/10.1006/ijhc.1999.0349 Date published: 2000-04-01 Date flagged: 2026-09-16

Date note: IJHCS Vol. 52, No. 4, pp. 701–717, April 2000. Print issue — no day-level date; month and year are the issue's.

Summary

Linda J. Skitka, Kathleen L. Mosier and Mark Burdick test whether a purely organizational variable — social accountability — reduces the two error classes that automated decision aids introduce. Their definitions are the standard ones the automation-bias literature still uses: omission errors are failures to respond to a system irregularity because the automated device did not detect or flag it; commission errors occur when people follow an automated directive despite contradictory information from other, more reliable sources, because they either fail to check that information or discount it. The authors' framing claim is that automated aids, "generally introduced to guard against human error, … can fundamentally change how people approach their work, which in turn can lead to new and different kinds of error."

The manipulation is pre-decisional accountability: participants were told before performing the task either that they would be accountable for their accuracy, or accountable for their overall performance, and would be required to justify what they did — against a control condition told their performance would not be monitored and could not be tracked. Nothing about the automation, the task, the information available, or the participants' training differed between conditions.

The result reported in every reachable account of the paper: participants under pre-decisional accountability were more likely to verify the automated information against other indicators, and committed significantly fewer omission and commission errors than non-accountable participants. The authors' interpretation is that accountability "sensitizes decision makers to the need to construct compelling justifications for their choices," and that this sensitization is what produces the verification behaviour — the checking is a response to anticipated justification, not to the reliability of the aid.

Access limit, binding on everything above. The full text is paywalled and was not obtained for this entry. Everything reported here is at abstract and secondary-account level; no sample size, no effect size and no error rates are held by this base. Per the standing rule on unreadable sources, this entry carries no Four Forces tags, and the finding may be cited only as a direction of effect, never as a magnitude.

Five Breakpoints Intersection

Incentive Fragmentation: This is the breakpoint's mechanism run as a controlled manipulation and run forwards — holding the technology, the task and the information constant and varying only whether the person expected to justify their decision, the error rate moved. It is the cleanest pre-AI demonstration in this base that whether a human check actually functions is a property of the accountability structure around the checker, not of the check's existence in the workflow.

Technology Illusion: The paper's own framing claim — an aid introduced to guard against human error changes how people approach the work and produces new and different errors — is the breakpoint stated as an experimental premise in 2000, with omission and commission named as the two specific forms the new error takes.

Four Forces Intersection

No Forces tags. The source is paywalled and was not read. Commitment is the obvious candidate and is deliberately not claimed: the abstract does not establish whether accountability changed follow-through under pressure or merely under observation, and guessing the distinction is exactly the failure this rule exists to prevent.

Relevance to Brandon's Positioning

This answers a negative claim the base made about itself and has been repeating since 2026-09-09. The scan's Open Question that day — carried into emerging-patterns.md as limit (1) on the Conditional-Productivity Result — was "has anyone put a non-technological organizational variable on the right-hand side of an AI-outcome specification? … nobody has." Skitka et al. did it in 2000, experimentally, with accountability as the variable and automation error as the outcome. The honest form of the claim is now narrower and more defensible: nobody has done it for generative AI; it was done for the previous shock, and the answer was yes.

The composition with today's other material is where the value is, and it belongs to Brandon rather than to either source. EY's 2026 survey says 98% of $1bn+ US public companies hold formal AI governance policies and 47% have not applied that process for urgent deployments. Skitka says the one organizational variable with experimental evidence that it suppresses automation-bias error is pre-decisional accountability — accountability imposed before the decision, which is precisely the thing a bypass-for-urgency clause removes. The enterprise practice that 47% report is not a general weakening of governance; it is the targeted removal of the single component the pre-AI literature identified as load-bearing. That is a sharper argument than "governance gets skipped," it is available to nobody who has read only the 2026 survey, and it directly strengthens asset 5 ("The Human-in-the-Loop Control That Isn't"), whose three-question audit currently asks what the reviewer is measured on without being able to say why that question is the right one.

Discipline: student participants, laboratory task, pre-generative automation, and this base holds no numbers. It supports a design principle and a diagnostic question. It does not support a quantitative claim of any kind, and must not be deployed as one.

Sources & Confidence

• Skitka LJ, Mosier K, Burdick MD. "Accountability and automation bias." International Journal of Human-Computer Studies 2000;52(4):701–717. DOI 10.1006/ijhc.1999.0349. Moderate — the journal and the result are high-credibility (peer-reviewed, the canonical citation in the automation-bias literature), but this base has not read the paper, and confidence is set by what was verified, not by the venue.
• Mosier KL, Skitka LJ, Burdick MD, Heers ST. "Automation Bias, Accountability, and Verification Behaviors." Proceedings of the Human Factors and Ergonomics Society Annual Meeting 1996;40(4). Moderate — the earlier conference paper in the same programme; noted as the same team's prior work, not read.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-16 — Wednesday falsification sweep, second find) Routing: governance (the EY 47% composition); strategy (asset 5)

Tags: paper-2, Technology Illusion, Momentum Mirage, Process Friction, Capability, COUNTERARGUMENT Source: Journal of the American Medical Informatics Association (JAMIA) URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC6915824/ Date published: 2018-11-01 Date flagged: 2026-09-15

Date provenance (hard rule 1). JAMIA 25(11) is a monthly issue; PubMed records the publication date as 2018 Nov 01 and that is the date stored. It is the publisher's issue date from the PubMed record, not an ingest date and not a guess. Advance-online date not verified and therefore not used.

Instrument, recorded before the finding (hard rule 9).

• Data source: 32 studies evaluating computerized hard-stop alerts in healthcare settings, identified from biomedical and computer-science databases, grey literature, reference lists and prior reviews.
• Inference method: systematic review with narrative synthesis. No meta-analysis — the authors report marked heterogeneity across studies. Study designs: 3 randomized controlled trials, 3 retrospective cohort studies, 18 pre-post studies, 5 time-series studies.
• Observation window: no date restriction applied; searches run 2017-03-09 and updated 2017-12-09.
• Known limits stated by the authors: no study rated excellent quality; heterogeneity prevented pooling; likely publication bias favouring positive results; underreporting of health outcomes and implementation harms.

Summary

Powers, Shiffman, Melnick, Hickner and Sharifi review the evidence on clinical decision support (CDS) hard-stop alerts, which they define in the paper's own objective as alerts "in which the user is either prevented from taking an action altogether or allowed to proceed only with the external override of a third party." They set three questions: how effective hard stops are at improving patient health and healthcare-delivery outcomes, what adverse events and unintended consequences they produce, and how they compare with soft stops — interruptive alerts the user may override unilaterally.

The headline results, verbatim from the abstract: "Of 32 studies, 15 evaluated health outcomes, 16 process outcomes only, 10 user experience, and 4 compared hard and soft stops. Seventy-nine percent showed improvement in health outcomes and 88% in process outcomes. Studies reporting good user experience cited heavy user involvement and iterative design. Eleven studies reported on unintended consequences including avoidance of hard-stopped workflow, increased alert frequency, and delay to care. Hard stops were superior to soft stops in 3 of 4 studies." The full text states the comparison explicitly: "Four studies directly compared hard-stop alerts to soft-stop alerts. Three out of the 4 studies showed superiority of the hard-stop alert in comparison to the soft-stop alert in achieving the desired process outcome." On outcome coverage the full text adds: "Of the 32 studies, 15 evaluated patient health or healthcare delivery outcomes, and 11 showed improvement in those outcomes."

Their conclusion is conditional rather than directional: "Hard stops can be effective and powerful tools in the CDS armamentarium, but they must be implemented judiciously with continuous user feedback informing rapid, iterative design. Investigators must report on associated health outcomes and unintended consequences when implementing IT solutions to clinical problems."

Five Breakpoints Intersection

Technology Illusion: The field installed a mandatory organizational control in software at scale and then largely declined to measure whether it helped: 16 of 32 studies reported process outcomes only, meaning half the literature on a control designed to prevent patient harm never measured patient harm. A gate whose evidence base is mostly "the gate fired and behaviour changed" is a technology answer standing in for an organizational one.

Momentum Mirage: Eleven studies reported unintended consequences including, in the authors' own list, "avoidance of hard-stopped workflow" — clinicians routing around the gate rather than complying with it. Where that happens the compliance statistic improves while the behaviour the gate exists to govern simply relocates, which is the appearance of progress without the movement.

Process Friction: "Delay to care" is named by the authors among the recurring unintended consequences: a control inserted to stop one class of error introduces structural friction into the clinical path it sits on, and the friction lands on a different outcome from the one the control was scored against.

Four Forces Intersection

Capability. The authors' conclusion does not say hard stops work; it says they work when the organization can keep working on them — "implemented judiciously with continuous user feedback informing rapid, iterative design" — and the studies reporting good user experience are the ones citing "heavy user involvement and iterative design." The variable separating a gate that helps from a gate that harms is the deploying organization's capacity to iterate it, which is Capability, not the control's design.

Relevance to Brandon's Positioning

This is the counter-evidence to the Binding-Gate Penalty pattern that research-evidence opened the same morning this entry was filed (2026-09-15, recurrence 2: HealthPlix 09-13 + Agachan 09-15). That pattern's claim is that "the cost of a check is not the check, it is the compulsion," and it pre-registered its own falsification condition: any instrument finding a coercive check beating an advisory one, where the reviewer can genuinely verify, is a counter-example. Powers et al. supplies four such head-to-head comparisons and three of them favour the coercive gate — in human subjects, peer-reviewed, eight years before the pattern was opened. The pattern's stated requirement that its third member "must be human subjects" is met, and met against it.

It does not, however, defeat the underlying mechanism — it relocates it. Drug-interaction alerts are the paradigm case of a reviewer that can verify: the check is deterministic against a known interaction table, which is precisely the condition under which Agachan's verify-versus-opine rule predicts a mandatory gate should win. So the correct reading is that the pattern is stated at the wrong altitude: the variable is not compulsion, it is compulsion without a verifiable acceptance criterion. That restatement survives all three instruments and remains falsifiable; the original does not survive this entry.

The second consequence is for knowledge/asset-suggestions.md asset #7, "The Ladder With No Verification Rung" (proposed 2026-09-10), which argues that AI maturity ladders contain no rung for checking the machine's output and that the absence is the failure. As drafted, that essay would be dismantled by anyone holding this literature: clinical informatics built the verification rung, made it mandatory, and measured the harm. The fix strengthens the piece — the missing rung is not review, it is a stated acceptance criterion — and hands it a thirty-year evidence base instead of an absence.

Sources & Confidence

• Powers EM, Shiffman RN, Melnick ER, Hickner A, Sharifi M. "Efficacy and unintended consequences of hard-stop alerts in electronic health record systems: a systematic review." J Am Med Inform Assoc. 2018;25(11):1556–1566. doi:10.1093/jamia/ocy112. PMID 30239810. — High (peer-reviewed systematic review in the field's principal journal). Metadata and abstract verified against the PubMed E-utilities record; Results and Discussion quotations read from the PMC full text.
• Confidence in the direction of the hard-vs-soft comparison: Moderate. Four studies is a thin base, the authors report no excellent-quality study among the 32, and they name likely publication bias favouring positive results. The finding is strong enough to refute a two-source generalization; it is not strong enough to establish the reverse one.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-15 — exploratory slot, own search; not held by research-evidence) Routing: strategy / governance

Tags: paper-2, Technology Illusion, Momentum Mirage, Process Friction, Capability, COUNTERARGUMENT Source: Archives of Internal Medicine URL: https://pubmed.ncbi.nlm.nih.gov/20876410/ Date published: 2010-09-27 Date flagged: 2026-09-15

Instrument, recorded before the finding (hard rule 9).

• Data source: 1,981 clinicians at 2 academic medical centers in Philadelphia, Pennsylvania (the ClinicalTrials.gov record NCT00870298 gives actual enrolment as 1,971 — the 10-clinician discrepancy between the registry and the published abstract is unexplained and is recorded rather than reconciled). 194 alerts fired in the intervention arm, 148 in the control arm.
• Inference method: randomized controlled trial, parallel assignment, single-masked (participant). This is a randomized manipulation of the coerciveness of a check, not an observational comparison.
• Observation window: 2006-08-09 to 2007-02-13. Terminated early; the registry gives the reason as "four adverse events encountered."
• Treatment: a "nearly hard stop" CPOE alert electronically stopping the order when trimethoprim-sulfamethoxazole was ordered against an active warfarin order, or both together. Control: standard practice — the registry describes this as "current practice of the pharmacist recommending cessation of concurrent warfarin and tmp/sulfa orders." Alerts were counted in both arms (148 control firings), so the control was the non-coercive configuration of the same decision point rather than an absence of any signal. The exact content of the control-arm alert is not established from the sources read here and should not be asserted.
• Sponsor: University of Pennsylvania.

Summary

Strom and colleagues set out to test whether making a drug-interaction check mandatory would fix the problem the authors name in their first sentence: "The effectiveness of computerized physician order entry (CPOE) systems has been modest, largely because clinicians frequently override electronic alerts." They randomized 1,981 clinicians at two academic medical centres to either a nearly hard-stop alert or standard practice, for the warfarin / trimethoprim-sulfamethoxazole interaction.

On the behaviour it was designed to change, the mandatory gate worked, and worked large. "The proportion of desired responses (ie, not reordering the alert-triggering drug within 10 minutes of firing) was 57.2% (111 of 194 hard stop alerts) in the intervention group and 13.5% (20 of 148) in the control group (adjusted odds ratio, 0.12; 95% confidence interval, 0.045-0.33)."

The trial was nonetheless stopped. "However, the study was terminated early because of 4 unintended consequences identified among patients in the intervention group: a delay of treatment with trimethoprim-sulfamethoxazole in 2 patients and a delay of treatment with warfarin in another 2 patients." The authors' conclusion states both halves without reconciling them: "An electronic hard stop alert as part of an inpatient CPOE system seemed to be extremely effective in changing prescribing. However, this intervention precipitated clinically important treatment delays in 4 patients who needed immediate drug therapy. These results illustrate the importance of formal evaluation and monitoring for unintended consequences of programmatic interventions intended to improve prescribing habits."

Five Breakpoints Intersection

Technology Illusion: The control performed exactly as specified against the metric it was specified against — a more-than-fourfold improvement in the desired response — and the organization still had to switch it off, because the harm arrived in a dimension the control did not model. A gate can be fully effective at its stated job and be the wrong intervention, which is the Technology Illusion in its cleanest available form: the technology was not broken, the organizational reasoning around it was.

Momentum Mirage: Had this been deployed rather than trialled, every dashboard would have shown success: alert compliance up from 13.5% to 57.2%, the interaction suppressed, the programme reporting green. The four treatment delays were visible only because a randomized trial with outcome monitoring was running underneath. The appearance of a working control and a working control are separated here by the presence of an evaluation, not by the presence of a result.

Process Friction: The intervention "precipitated clinically important treatment delays in 4 patients who needed immediate drug therapy" — a mandatory gate placed on an urgent clinical path introduced friction into the path itself, and the friction fell on patients rather than on the ordering clinician whose behaviour it was scored against.

Four Forces Intersection

Capability. The authors' closing sentence makes the organizational capacity to evaluate its own controls the lesson of the trial — "the importance of formal evaluation and monitoring for unintended consequences of programmatic interventions." What distinguished this deployment from a harmful one was not the design of the alert but the presence of an instrument watching for what the alert did outside its own success metric.

Relevance to Brandon's Positioning

This is the sharpest single instrument in the base for Claim 4 (the failure mode is harder to detect because AI produces the appearance of execution), and it is sharper because it is not about AI. A control's own success metric moved decisively in the right direction while the outcome the control existed to protect moved the other way — and the only reason anyone knew was that a randomized trial with outcome monitoring was running. That is the argument of Claim 4 demonstrated experimentally, in human subjects, sixteen years before the AI governance market started selling the same architecture.

It is also the one of Powers et al.'s four direct hard-versus-soft comparisons that runs against the mandatory gate, and it is the only randomized one — which is why it belongs in the base alongside Powers rather than inside it. Together the two entries bound the Binding-Gate Penalty pattern from both sides: coercion usually wins on the process outcome (Powers, 3 of 4) and can still lose on the outcome that matters (Strom). The discriminating variable is not whether the gate binds, it is whether anyone is measuring outside the gate's own success metric — which is the verification-thread question restated, and the reason asset #7 should be rebuilt around acceptance criteria rather than around the absence of review.

Standing caution for any use: n = 4 adverse events. The harm finding is a stopping decision by a data-monitoring process, not a powered estimate, and Brandon should say so in the same sentence he cites it. The behavioural result (57.2% vs 13.5%) is the statistically robust half.

Sources & Confidence

• Strom BL, Schinnar R, Aberra F, Bilker W, Hennessy S, Leonard CE, Pifer E. "Unintended effects of a computerized physician order entry nearly hard-stop alert to prevent a drug interaction: a randomized controlled trial." Arch Intern Med. 2010;170(17):1578–1583. PMID 20876410. — High (randomized controlled trial, peer-reviewed). Full abstract, authorship, pagination and dates verified against the PubMed E-utilities record.
• ClinicalTrials.gov NCT00870298 (University of Pennsylvania) — High for registry facts: arms, enrolment (1,971 actual), 2006-08 start, 2007-01 actual completion, status TERMINATED, "four adverse events encountered." Verified against the ClinicalTrials.gov API v2 record.
• Powers et al. 2018 (knowledge-base/powers-jamia-hard-stop-alerts-systematic-review-2018-11.md) counts this trial among its 32 studies and among the RCTs reporting delay-to-care harm — High for that placement.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-15 — exploratory slot, own search; not held by research-evidence) Routing: strategy / governance

Tags: paper-2 Technology Illusion Momentum Mirage Momentum Capability COUNTERARGUMENT Source: Journal of Medical Internet Research (JMIR) URL: https://www.jmir.org/2017/11/e367/ Date published: 2017-11-01 Date flagged: 2026-09-14

Summary

Greenhalgh, Wherton, Papoutsi, Lynch, Hughes, A'Court, Hinder, Fahy, Procter & Shaw build the NASSS framework — Non-adoption, Abandonment, and challenges to Scale-up, Spread and Sustainability — from two parallel components. The first is a hermeneutic systematic literature review that identified 28 previous technology-implementation frameworks. The second is empirical: six technology-supported programmes studied by longitudinal ethnography and action research for up to three years across more than twenty organizations, generating over 400 hours of ethnographic observation, 165 semistructured interviews, and 200 documents.

The framework has seven domains, verbatim: (1) the condition or illness; (2) the technology; (3) the value proposition; (4) the adopter system; (5) the organization; (6) the wider context; and (7) embedding and adaptation over time. Domain 7 is interrogated by two questions, verbatim: "How much scope is there for adapting and coevolving the technology and the service over time?" and "How resilient is the organization to handling critical events and adapting to unforeseen eventualities?"

The headline empirical claim is about complexity rather than about any one domain. Programmes "characterized by complexity in multiple NASSS domains rarely, if ever, became mainstreamed" — the authors' account of long-term sustainability is one of continuous adaptation to context over time rather than of a successful go-live. The framework's organizing failure construct is non-adoption and abandonment: the technology is not taken up, or is taken up and then dropped.

Five Breakpoints Intersection

Technology Illusion: The central finding — that programmes complex across multiple domains "rarely, if ever, became mainstreamed" regardless of the technology's own merit — is a direct, ethnographically grounded statement that deploying a working technology onto unresolved organizational conditions does not produce adoption, which is the breakpoint's claim in another vocabulary.

Momentum Mirage: Domain 7 and the abandonment construct are the tradition's treatment of effort that does not sustain, and they are the reason this entry matters: NASSS locates the failure in the technology being dropped, whereas Momentum Mirage names the case where the effort visibly persists and produces no movement. The overlap is real and the residue is the distinction — see Relevance below.

Four Forces Intersection

Momentum: Domain 7 asks directly whether the organization can adapt and coevolve a service over time and how resilient it is to unforeseen events — a sustainment construct occupying the same conceptual ground as Momentum, published in 2017 and continuously used since.

Capability: The organization domain and the resilience question in 7B both measure organizational capacity to absorb and keep absorbing a change, which is Capability as Brandon defines it (readiness and flow), scored qualitatively rather than on an instrument.

Relevance to Brandon's Positioning

This entry exists to stop a publication error, and the error was recommended in today's research-evidence scan brief. That brief featured Sittig & Singh's eight-dimension sociotechnical model as a competing framework, observed that none of its eight dimensions concerns sustained force after launch, and told Brandon this: "Brandon should publish that observation." The observation is true of Sittig & Singh's enumerated eight. It is false of the field, and a referee needs one sentence to say so: NASSS's seventh domain is literally "embedding and adaptation over time," the framework is named for sustainability and abandonment, and it rests on 400 hours of ethnography rather than on assertion. The same brief's claim that the field map "has not previously recorded a genuine competing framework with twenty years of empirical backing" also understates the position — Greenhalgh's own review counted 28 prior frameworks in 2017.

The narrower claim that survives is better than the one it replaces, and it is the publishable one. The sociotechnical tradition theorizes the sustainment failure as abandonment — the technology is dropped, or never mainstreams. Ash's 176-hospital survey (research-evidence knowledge-base/ash-cpoe-unintended-consequences-176-hospitals-2007-07.md) measures the opposite case and neither framework has a category for it: systems described as "highly infused, or embedded, within work practice", median five years of use, universal adoption — and organizational consequences that showed no relationship to years of use. That is not non-adoption and it is not abandonment. It is successful embedding with undiminished harm, which is precisely what Momentum Mirage names and what a framework organized around drop-out cannot hold. Brandon should publish that, and cite NASSS as the strongest version of the position he is distinguishing himself from rather than as an absence.

Sources & Confidence

• Greenhalgh T, Wherton J, Papoutsi C, Lynch J, Hughes G, A'Court C, Hinder S, Fahy N, Procter R, Shaw S. "Beyond Adoption: A New Framework for Theorizing and Evaluating Nonadoption, Abandonment, and Challenges to the Scale-Up, Spread, and Sustainability of Health and Care Technologies." J Med Internet Res 2017;19(11):e367. doi:10.2196/jmir.8775. — High (peer-reviewed, open access). Citation, method, sample, the seven domain names and the two domain-7 questions were read from the open-access PubMed Central rendering (PMC5688245) via the fetch tool. jmir.org returned empty content to this container; PubMed returned a cookie wall.
• Sittig DF, Singh H. "A new sociotechnical model for studying health information technology in complex adaptive healthcare systems." Qual Saf Health Care 2010;19(Suppl 3):i68–i74. PMID 20959322. — High as a publication, but abstract-level only in this base. PubMed cookie-walled, the UTHealth self-hosted PDF did not yield extractable text, and Springer redirected to authentication. The eight dimensions are confirmed by enumeration across two independent renderings; no full-text read has been performed here, so any claim about what the model does not contain rests on the enumerated list and must be stated that way.
• Access note: eleventh consecutive research day on which a primary or corroborating route required a workaround. PyMuPDF and pypdf are both absent from this container, so the 09-04 "extract the self-hosted PDF" route is currently unavailable to Stafford.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-14) Routing: personal / strategy

Tags: paper-2 practitioner Technology Illusion Momentum Mirage COUNTERARGUMENT Capability Source: Journal of the American Medical Informatics Association (JAMIA) URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC13385993/ Date published: 2026-08-01 Date flagged: 2026-09-13

Instrument

Recorded before the finding, per CLAUDE.md rule 9.

• Data source: the published systematic-review literature itself. This is a systematic review of systematic reviews (an umbrella review) — its unit of analysis is a review article, not a clinician, an alert, or an organisation.
• Inference method: structured search of the Cochrane Library, Embase and PubMed; quality and bias assessed with R-AMSTAR (Revised Assessment of Multiple Systematic Reviews); narrative and descriptive statistical synthesis. It measures how the field measures, which is why it is in this base — it is a measurement-layer finding, not an effect estimate.
• Observation window: database inception through 2024.
• Sample: 22 systematic reviews included; individual reviews reported between 1 and 11 alert-related metrics; most were judged medium quality.
• Publication: JAMIA 33(8):1523–1531, August 2026 issue; DOI 10.1093/jamia/ocag064. Advance access was circulating by 2026-05-18 per the publisher's own download stamp on the author copy — the August issue date is used here as the citable one.
• Authors' base: predominantly US Department of Veterans Affairs health-services research (Edward Hines Jr. VA Medical Center's Center of Innovation for Complex Chronic Healthcare, Iowa City VA, VA Pittsburgh) with Northwestern Feinberg, MetroHealth/Case Western Reserve, Colorado and UIC. This is the health-system side of the field, not the vendor side — relevant, because they have no product to defend.

Summary

Ray, Wilson, Hughes and colleagues set out to find how alert fatigue is actually defined and measured in the literature that studies it. Alert fatigue is, as they put it, alert dismissals caused by excessive or irrelevant alerts, and it is one of the most frequently cited barriers to clinical decision support having any effect. Twenty-two systematic reviews met inclusion.

The headline result is an absence: "Only one article provided an operational definition of alert fatigue." One, of twenty-two. The reviews report alert metrics freely — "the most common alert metrics were quantity, override rate, and acceptance rate" — but the construct those metrics are supposed to measure is, in twenty-one of twenty-two cases, never operationally defined. The authors also report that description of the underlying primary studies was frequently insufficient to judge what had been measured or in what setting.

Their conclusion is a proposal to fix it: the field should converge on "a significant, sustained decrease in appropriate alert response rates from an established baseline" as the measure of alert fatigue. Two words in that sentence carry the weight — appropriate, which requires knowing whether a given override was clinically correct, and baseline, which requires having measured the response rate before the problem started.

Five Breakpoints Intersection

Technology Illusion: The largest and longest-running human-in-the-loop deployment in existence has produced twenty-two systematic reviews of its own failure mode and one operational definition of what that failure mode is. Counting overrides is not the same as knowing whether the control is working, because an override can be correct; the field has the count and, by its own audit, has not agreed what the count means.

Momentum Mirage: This is the mirage one level above the one the base already holds. The 2026-09-13 scan established that an institution can produce an audit trail showing full compliance over a control doing almost nothing. This source shows that the research literature about that control can likewise accumulate for two decades without establishing whether its central measure tracks the thing it names — activity at the measurement layer, not movement.

Four Forces Intersection

Capability. The authors' proposed definition is a specification of what an organisation would need to be able to do to know its verification layer is failing: measure appropriate response, hold a baseline, and detect sustained decline against it. Stated that way it is a capability requirement, and the review's finding is that the field asking for it has largely not built it.

No Purpose, Commitment or Momentum tag — an umbrella review of measurement practice observes none of these.

Relevance to Brandon's Positioning

This corrects the 2026-09-13 research-evidence brief on the exact sentence it routed to Brandon as a consulting opener. That brief argued: "medicine at least measures its override rate, and no enterprise AI governance stack this base has seen measures one at all." The first half is now too generous to medicine. Medicine counts overrides; the field's own umbrella review says it has not defined the failure the count is a proxy for, and that the primary studies frequently do not report enough to tell. The opener survives and gets sharper, but the sharpening changes the sentence: the problem is not that enterprises fail to count what medicine counts — it is that counting is the easy part, and the field twenty years ahead of them has done the easy part and not the hard one.

That matters for paper-2 because it supplies something the case file is short of: a named, specified, published measurement standard for whether a human review layer is actually functioning — appropriate response rate, against a baseline, over time. Brandon's framework repeatedly asserts that organisations mistake compliance for capability; this is a mature profession stating the same thing about its own instruments and proposing the fix. It is the closest artifact this base holds to an answer for the verification thread's criterion (1), and it is worth being precise about why it is not one: it is a recommendation to researchers, not a standard adopted by an operating organisation.

Two binding limits. (1) Not AI — rule-based clinical alerting, windows ending 2024, and any transfer to AI governance is an argument by analogy that must be stated as one. (2) Umbrella reviews inherit their sample's weaknesses; "only one operational definition" is a finding about the review literature, and a primary study could have defined the construct without its reviewer reporting that it did. The authors say the reporting was insufficient, which cuts both ways.

Logged COUNTERARGUMENT for one specific reason, and it is aimed at Brandon's own use of this material rather than at the thesis: the override rate is the number he is about to build a client diagnostic on, and the field that invented it has just published a paper saying it does not carry the meaning people assign it. Asking a board "what is your override rate?" is a good question. Treating a high answer as proof of failure is the error this review exists to name — some overrides are right.

Sources & Confidence

• Ray CE, Wilson GM, Hughes AM, Goedken CC, Liu EP-F, Fitzpatrick MA, Suda KJ, Kota SM, Nwankpa C, Evans CT. "Alert fatigue measurement in clinical decision support: a systematic review." JAMIA 33(8):1523–1531, August 2026. DOI 10.1093/jamia/ocag064 — High (peer-reviewed, top informatics journal, R-AMSTAR quality assessment, author copy read directly).
• Corrective context for research-evidence briefs/2026-09-13.md and knowledge-base/ddi-information-intervention-field-experiment-2026-09.md — High confidence in the correction.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-13) — exploratory slot, paper-2 Claim 2 approached via the clinical override literature per the scan query log's own recorded instruction. Routing: strategy / governance

Tags: paper-2 practitioner Technology Illusion Process Friction Momentum Mirage Capability Source: BMC Medical Informatics and Decision Making URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC8864797/ Date published: 2022-02-21 Date flagged: 2026-09-13

Instrument

Recorded before the finding, per CLAUDE.md rule 9.

• Data source: operational audit logs from the Klinisch Werkstation (KWS) electronic patient record at University Hospitals Leuven (UZ Leuven), a 1,995-bed tertiary teaching hospital in Belgium, integrating computerised physician order entry with the drug–drug interaction (DDI) decision support module — plus a separate cross-sectional physician survey.
• Inference method: retrospective descriptive quantitative evaluation of every "very severe" DDI alert fired and the prescriber's logged response. This is administrative telemetry, not self-report — the override is a recorded system event, not a claim by the clinician. The accompanying survey is self-report and is a separate instrument with a separate limit (below).
• Observation window: January 2019 – January 2021 (24 months). Survey fielded February 2021. Entirely pre-generative-AI and pre-dating the technology this base is about.
• Sample: 38,409 very severe DDI alerts across 64 distinct DDI pairs, drawn from 4,241,087 prescriptions over 110,400 inpatient and 232,948 day-care admissions. Survey: 173 physicians analysed of 1,631 invited — a 10.6% response rate, which is the survey half's binding weakness.
• Unit of analysis: the alert, not the patient and not the clinician.

Summary

The authors set out to evaluate a deployed DDI decision support system on its performance, uptake and user satisfaction, and to identify what could be improved. The system tiers alerts by severity, and the tiering is itself the finding of interest: "very severe" DDIs generate an interruptive alert that requires a free-text override motivation confirmed with the prescriber's password; "severe" DDIs generate an interruptive pop-up that does not have to be overridden; remaining DDIs are non-interruptive and can be consulted on request. The very-severe tier is therefore close to the strongest hard stop a clinical system can impose short of outright blocking the order — it interrupts, it demands a typed justification, and it re-authenticates the user.

Of the 38,409 very severe alerts, 88.2% were overridden by the prescriber — an initial acceptance rate of 11.8%. A second, non-interruptive layer sits behind the first: a pharmacist reviews overridden alerts. For 13,867 of the reviewed alerts (96.8%) the pharmacist took no action; for 453 (3.2%) a recommendation was issued. Where a pharmacist recommendation was made, 79.2% of them were accepted.

The authors also report a configuration detail that bears directly on the override rate: the system applies a fixed screening interval of seven days before and after a drug is prescribed, a fourteen-day window, and therefore fires alerts "even when there is no actual co-prescription." A substantial share of the alert volume is generated by the system's own specification rather than by a live interaction in front of the prescriber.

Five Breakpoints Intersection

Technology Illusion: The detection capability is not in question — the interaction database is comprehensive and the flag is accurate at the moment it fires — and the control was nevertheless discharged 88.2% of the time by the very design intended to make discharging it hard. The organisation bought interruption, a typed justification and password re-authentication, and what it received was a logged ritual: the technology worked and the behaviour around it was never designed.

Process Friction: The seven-day-either-side screening window means the system alerts on drug pairs that are not actually co-prescribed, so the friction is inserted at the point of decision by the control's own configuration rather than by clinical reality. This is the mechanism behind the override rate that the behavioural reading misses — a meaningful portion of what looks like clinician non-compliance is a specification choice made by whoever configured the tool.

Momentum Mirage: Every one of those 38,409 alerts produced a complete audit record: alert fired, justification typed, password entered, override logged. An institution running this configuration can evidence total procedural compliance over a control that changes the prescription in roughly one case in nine.

Four Forces Intersection

Capability. The pharmacist layer is the one part of this system that works, and it is the part with no enforcement in it at all: a human reviews the override afterwards, makes a recommendation in 3.2% of cases, and is accepted 79.2% of the time. Capability here is not the database and not the gate — it is a competent person given the time to look, whose advice is non-binding and is therefore taken.

No Purpose, Commitment or Momentum tag: the study measures alert responses and does not observe intent, leadership follow-through, or change over a transformation lifecycle.

Relevance to Brandon's Positioning

This is the primary source behind the override figures the 2026-09-13 research-evidence scan carried second-hand from the Shao/Choudhary/Majumdar preprint's literature review, and reading it changes two things the scan could not have known. First, the number: the scan reported "up to 95%" and attributed it partly to this paper, which actually reports 88.2%. The higher figures in the literature belong to other studies and other alert classes (drug-allergy alerts run higher), so "up to 95%" is defensible as a range across the field but must not be attributed to this instrument. Second, and more valuable, the architecture detail the literature review dropped: the 88.2% was achieved against a free-text justification confirmed by password. That is a materially stronger version of Brandon's argument than the one the scan stated — the hard stop does not merely fail, it fails when it is made as hard as a hospital can make it.

The finding that should reach a client is the pharmacist layer. Same information, same institution, same clinicians, two designs: a binding gate at the point of work is discharged 88.2% of the time, and a non-binding recommendation made afterwards by someone with time to look is accepted 79.2% of the time. For paper-2's Claim 5 — that the prerequisite is organisational readiness rather than coordination technology — this is a within-instrument contrast where the technology is held fixed and the only variable is where the human review sits and whether it can be ignored.

Binding limit on any use: this is not AI, and the window is 2019–2021. It is rule-based lookup against an interaction database. It carries as mechanism evidence for what happens to a mandatory human-in-the-loop gate, and citing it as an AI result would be the analogy error this pipeline catches in other people's work.

Sources & Confidence

• Van De Sijpe, G., Quintens, C., Walgraeve, K., et al., "Overall performance of a drug–drug interaction clinical decision support system: quantitative evaluation and end-user survey," BMC Medical Informatics and Decision Making 22:48 (2022-02-21) — High on the telemetry half (peer-reviewed, full text read at primary, system-logged events, n = 38,409); Low on the survey half (10.6% response rate, self-selected).
• Attribution correction to research-evidence knowledge-base/ddi-information-intervention-field-experiment-2026-09.md, which carried "up to 95%" citing this paper among others — High confidence in the correction, primary text read directly.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-13) — exploratory-slot primary verification, not a scan find. Routing: strategy / governance

Tags: paper-2, practitioner, Technology Illusion, Capability Source: Diligent Institute / Chief Executive Group — Director Confidence Index (June 2026 edition), by Melanie Nolen, Research Editor URL: https://www.diligent.com/resources/blog/dci-board-ai-use-2026 Date published: 2026-06-17 Date flagged: 2026-09-12

Summary

The Director Confidence Index is a recurring survey run by the Diligent Institute in partnership with Chief Executive Group. The June 2026 edition reports on a sample of 104 U.S. public company directors, with comparison data drawn from a previous wave fielded in September 2025. The article's subject is directors' own use of generative AI in the conduct of board work.

The two figures that carry the entry are an adoption number and a governance number, and they are reported side by side. "[R]oughly 30 percent of those surveyed have used generative AI to summarize board books or meeting materials." Against that, 54 percent say there is no guidance for directors' use of AI in place at their company, and only 6 percent report a formal policy specific to the board.

What the article does not report is equally load-bearing and is recorded here as an absence rather than inferred: it contains no data on directors' confidence in board oversight of AI, no measure of board AI expertise, and nothing on whether the materials arriving at the board were themselves prepared with AI. The instrument measures the director compressing what reaches them. It does not measure what composed what reached them. Fielding dates for the June 2026 wave are not stated on the page.

Five Breakpoints Intersection

Technology Illusion: Generative AI is inside the board's own information path at roughly 30% of these boards while 94% have no board-specific policy governing that use (54% report no company guidance at all, 6% a formal board policy) — the technology has entered the oversight function ahead of any control on it, which is this breakpoint's definition applied to the body that is supposed to be the control.

No Strategic Disconnection mapping is claimed. The survey reports use and policy, not any gap between stated direction and operational reality, and tagging it would be the speculative stretch the base's rules forbid.

Four Forces Intersection

Capability: The Force is implicated on the receiving body rather than on the workforce — a board adopting a summarization tool for the documents on which it exercises judgment, with 94% lacking a policy defining acceptable use, is a capability gap in the oversight layer stated by the directors themselves rather than inferred about them.

No Purpose, Commitment or Momentum tags. The instrument asks about tool use and policy existence and supplies no evidence bearing on intent, follow-through under pressure, or sustained energy.

Relevance to Brandon's Positioning

This lands on Paper 3, "Designed to Stall," mechanism 2 — the stewardship reporting line, whose text requires "independent reporting on transformation health to the board separate from the sponsor's status updates," and on the severe drift tier that "triggers escalation to the board with the possibility of program restructuring." Paper 3's design treats the board as a recipient whose independence is secured by the route the information travels. This instrument measures a compression step sitting on that route, inside the recipient, that the paper does not contemplate: at roughly 30% of these boards the independent report is being summarized by a generative model before it is read, under no policy at 94% of them.

Its value is as the receiving-end half of a pair. Eric So's profit-mandate experiment (arXiv 2609.07731, research-evidence knowledge-base/profit-mandate-alignment-failure-llms-2026-09.md) measures an agent deciding what to escalate upward and finds board-escalation recommendations falling 74.4% → 60.5% under an ordinary business objective. This measures a director using an agent to compress what did arrive. The two sit at opposite ends of Paper 3's mechanism-2 path and neither paper knows about the other. Together they are the argument that mechanism 2's independence is a property of the route and not yet of the judgment, which is a cheap and strengthening revision to a published paper rather than a refutation of it — logged as such in knowledge/thesis-challenges.md.

It is weaker than most of what the base holds on measurement quality — n=104, a vendor-affiliated programme, undisclosed fielding dates, self-report — and stronger than anything the base holds on coverage, because the base contains no measurement of AI inside the board's own information handling. It also advances the board expertise series thread from an unexpected direction: that thread has been waiting for the competence substrate to catch up to the disclosure, and this measures a different substrate moving first, in the wrong direction.

Sources & Confidence

• Diligent Institute / Chief Executive Group, Board AI use rises: Director Confidence Index June 2026, 2026-06-17 — Moderate. Diligent sells board-management software, so the publisher has a commercial interest in board AI adoption being salient; n=104 is small; the respondent panel is self-selected; fielding dates are not disclosed. The two headline figures are directly quoted above and should always travel with the sample size. Do not state either as a population estimate for U.S. directors.
• Eric So, The Profit Alignment Problem (arXiv 2609.07731, 2026-09-07) — Moderate (unrefereed v1 preprint; author affiliation verified 2026-09-12 as MIT Sloan, see brief).
• Brandon Freitag, Designed to Stall (Paper 3), papers/paper-3-designed-to-stall.md — High (canonical text, quoted directly).

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-12) — surfaced by the exploratory slot, redirected after the slot's first target was found to be an already-resolved challenge. Routing: governance (primary), strategy.

Tags: paper-2, Momentum Mirage, COUNTERARGUMENT Source: arXiv preprint (Yulin Yu, Yan Chen, Rui Hu, Siddharth Suri, Scott Counts) URL: https://arxiv.org/abs/2608.15550 Date published: 2026-08-16 Date flagged: 2026-09-11

Summary

Yu, Chen, Hu, Suri and Counts measure how Microsoft 365 Copilot adoption changed the volume and mix of information work, using digital trace data from the M365 application suite. Their stated motivation is the evidentiary gap: "Despite rapid adoption, empirical evidence on how these tools alter work practices and generate productivity gains remains limited." The design is a difference-in-differences on 11 international companies across different industries, collectively employing more than 10,000 employees, covering 40,164 users enabled during the study period, of whom 7,831 used Copilot more than 100 times. The window is January 2024 to September 2024 — ten weeks pre-adoption and twenty weeks post-adoption. The control group is users who enabled Copilot after 1 September 2024, matched by propensity score on pre-enablement activity, with exact matching on management position.

The headline results are reported as associations. Among users who used Copilot more than 100 times over the twenty-week post-adoption period, productivity application actions rose 21.2% and communication application actions rose 7.1%. Among users in the 100–500 use band, higher usage is associated with continued increases in both. The authors read the asymmetry as compositional: "The smaller increase in communication represents an overall shift toward individual, documentation-focused work and reflects mixed changes in communication, including decreases in reading and organizing email, compared with more uniform increases in productivity actions." They close on a caution about what the shift may cost: the need "to ensure that AI adoption does not weaken interpersonal communication and the diffusion of diverse information that supports innovation."

The operational definitions are the part of the paper that matters most for how the numbers may be used. "Productivity actions" are actions in Word, Excel and PowerPoint (94.6% of the total); "communication actions" are actions in Outlook and Teams (99.1%). The specific actions the authors name as tracked include "Paste" and "Open App" in Word, and "selecting mail items" and "reading email" in Outlook. The authors state their own limits plainly: they "are unable to directly measure time allocation or obtain perfect measures of productivity" and do not claim causal identification, resting on propensity-score matching and a verified parallel-trends assumption to mitigate selection.

Five Breakpoints Intersection

• Momentum Mirage: The study's 21.2% increase in "productivity" is a count of application actions whose named constituents include "Paste" and "Open App" in Word — so the largest generative-era telemetry measure of AI productivity in this base counts opening the application and pasting model output as production, and contains no outcome variable of any kind: no shipped artifact, no quality rating, no revenue, no externally measured result. This is the appearance of progress measured directly and at scale, in an instrument whose treatment (using a tool that generates text for pasting into documents) mechanically produces the behavior the outcome variable counts.

Four Forces Intersection

No Force tag, and the reason is not inaccessibility. The paper was read in full and its constructs are individual application-activity counts, not organizational properties. Tagging Capability here would repeat the error this base refused on 2026-09-10 with Ide & Talamàs — laundering an individual-level endowment or behavior into an organizational Force. The Four Forces describe what an organization builds or fails to build; nothing in this instrument measures an organization.

Relevance to Brandon's Positioning

This lands squarely on the Open Question the research-evidence scan posed today from Bloom, Garicano, Sadun & Van Reenen (NBER WP 14975): does this technology make it cheaper to know, or cheaper to ask? Bloom et al. showed the two costs move authority in opposite directions and that measuring them together returns nothing; the scan correctly noted that no AI-era instrument in this base makes the distinction. This one makes it in the generative era, on telemetry, across eleven companies — and the observable mix moves toward knowing: production actions rise three times as fast as communication actions, and email reading and organizing fall outright.

But it cannot answer the question, and the reason is worth more to Brandon than the finding. Bloom et al.'s dependent variable is a decision right — who controls hiring, investment, product introduction and marketing. This paper's dependent variable is a count of application actions. Inferring authority from activity is exactly the substitution the Résumé-Panel Consensus pattern exists to flag, in a new genre; it must not be recruited as a break in that pattern, and it meets none of the Regime Fork thread's criteria. The sharper observation is that the missing variable was in the dataset. The authors exact-match their control group on management position — they hold a hierarchy indicator for 40,164 users in the generative era — and spend it as a matching covariate rather than estimating anything on it. The scan's discipline (8) from today says the gap must be stated as nobody has run the design, not the data does not exist. Here the data exists, in the right era, with the right technology on the right-hand side, and the structural variable was used to balance the sample. That is a stronger and more checkable version of the same point, and it is the cheapest possible ask of these authors.

Sources & Confidence

• Yu Y, Chen Y, Hu R, Suri S, Counts S, Adoption of Generative AI in the Workplace: Increasing and Shifting the Balance of Productivity and Communication Activity, arXiv 2608.15550, submitted 2026-08-16 — abstract and HTML full text read directly — Moderate. Three reasons, all binding: it is a preprint with no peer review; the authors are affiliated with the vendor whose product is the treatment and whose telemetry is the instrument, which is the same caveat this base applied to Salesforce and LHH today; and the design is association, not identification, as the authors themselves state.
• Observation window January–September 2024, which is generative but pre-agentic. A 2024 Copilot window is not evidence about 2026 agentic deployments and must not be presented as current.
• Selection caveat beyond the authors' own: the treatment group is defined by self-selected usage intensity (100+ uses). Heavy users choosing to be heavy users is not resolved by matching on pre-enablement activity, and the co-movement of both outcome families is consistent with engaged workers using the tool more.

Filed by / Routing

Filed by: Stafford (scheduled daily brief, 2026-09-11) — exploratory slot, found while targeting Claim 3's open breakpoints Routing: strategy

Tags: paper-2, Technology Illusion, COUNTERARGUMENT Source: Journal of Political Economy (Enrique Ide, Eduard Talamàs — IESE Business School) URL: https://www.journals.uchicago.edu/doi/10.1086/737233 (preprint full text: https://arxiv.org/abs/2312.05481) Date published: null Date flagged: 2026-09-10

Date discipline. The article is Journal of Political Economy vol. 133 no. 12, pp. 3762–3800, 2025 — the December 2025 issue. The publisher page returned HTTP 403 in this session and the exact online-first date was not verified, so the field is an honest null rather than a manufactured day. The arXiv preprint (2312.05481) is at v12, dated 2025-05-17; every quotation below is from the preprint full text, not from the journal version, and the two have not been diffed.

Summary

Ide and Talamàs incorporate artificial intelligence into Garicano's knowledge-hierarchy framework. The economy's premise, in their words: "Less knowledgeable individuals become 'workers' doing routine work, while others become 'solvers' handling exceptions." Hierarchies arise because "knowledge is tacit" — specifically because "matching problems to those who can solve them is difficult, and experts cannot outline a plan of action in advance." The structure exists to economize on the application of scarce expert time, letting the most knowledgeable "focus on their comparative advantage: solving complex problems." The paper builds explicitly on Garicano (2000), which "shows that knowledge hierarchies are optimal when production requires the application of tacit knowledge."

AI enters as a technology converting computational resources into AI agents that operate either autonomously (as co-workers and as solvers) or non-autonomously (solely as co-pilots). The headline result, verbatim from the abstract: "Autonomous AI primarily benefits the most knowledgeable individuals; non-autonomous AI benefits the least knowledgeable. However, output is higher with autonomous AI." The authors present this as reconciling conflicting empirical findings and as identifying regulatory trade-offs around AI autonomy.

The structural results are directional and depend on AI capability. For basic autonomous AI, "AI displaces humans from routine knowledge work to specialized problem solving" (W* ⊂ W, S* ⊃ S); for advanced AI "AI displaces humans in the opposite direction" (W* ⊃ W, S* ⊂ S). And the result that matters most here: under basic autonomous AI, the people who remain in routine work are made worse off — "the productivity of z∈W*⊂W is strictly lower post-AI than pre-AI." The mechanism has two channels: AI agents become the highest-quality workers and therefore attract the best solvers, and the marginal pre-AI workers move up into solving roles, so the solver pool available to those still doing routine work is less knowledgeable than before. The paper is purely theoretical — no data, no calibration, no estimation.

Five Breakpoints Intersection

• Technology Illusion: The model derives, from optimizing behaviour alone, that introducing basic autonomous AI leaves the humans still doing routine work "strictly lower post-AI than pre-AI" in productivity — a formal statement that deploying capable technology into an existing knowledge hierarchy degrades part of the human layer beneath it, which is Brandon's fourth breakpoint arriving as an equilibrium property rather than as a deployment error.

Strategic Disconnection, Incentive Fragmentation, Process Friction and Momentum Mirage were considered and deliberately not tagged. There is no principal, no stated intent and no follow-through in this model — allocation is an equilibrium outcome of knowledge costs, not a decision anyone makes. Tagging breakpoints that require a decision-maker onto a model that contains none would be exactly the loose tagging the rules forbid.

Four Forces Intersection

No Force tagged, and the reason is the finding rather than a limitation of access. The full text was read. The model's entire right-hand side is individual knowledge endowments (z), AI capability (z_AI), computational resources, and the cost of helping — not purpose, commitment, organizational capability or momentum. Capability is the tempting tag and it is the wrong one: here capability is an individual endowment, not an organizational property, and tagging it would launder this entry's central objection into support. Recorded as a deliberate null.

Relevance to Brandon's Positioning

This is the peer-reviewed anchor Claim 1 has been missing, and it does not say what Claim 1 says. Paper-2 Claim 1 states that traditional hierarchy was an information routing protocol. The canonical formalization it needs to cite says something adjacent but materially different: hierarchy economizes on scarce expert time, and routing is the mechanism by which it does so, not the purpose. The distinction is load-bearing for Claim 2. If hierarchy routes information, then AI — which moves information cheaply — makes the layer redundant, and Claim 2's "AI removes the information routing problem" follows. If hierarchy prices expert attention, AI does not remove the problem; it repriced the input, and the layer is reallocated rather than deleted. Ide and Talamàs derive the reallocation in both directions depending on AI capability.

The gift inside it, and Brandon should take it. He has argued Technology Illusion from cases — the platform nobody used, the cloud that provisioned faster than the organization could move. Here is a top-five economics journal deriving that basic autonomous AI makes the remaining routine workforce less productive, because the best solvers are pulled toward supporting AI production and the help available to everyone else gets worse. Nobody in that model makes a mistake. It is the cleanest available statement that capable technology can degrade the human layer beneath it without anyone deciding to degrade it.

The cost, and it is the larger half. That same property is the strongest formal objection this base has surfaced to the framework's diagnostic logic. Five Breakpoints reads a degraded human layer after an AI deployment as evidence that the organization was not ready. This model produces a degraded human layer under fully rational agents, no misalignment, no broken process, and higher total output. If the efficient equilibrium and the diagnosed failure look the same from outside, then observing the degradation is not evidence of a breakpoint. Logged as a new OPEN challenge — the equilibrium confound — in knowledge/thesis-challenges.md.

And it settles a standing bet against the framework's interest. The 2026-09-05 entry in the challenge ledger recorded that three formal models in a row had been ingested as friendly and had each come out the framework's way on inspection, and pre-registered the discipline that the fourth should be read parameter-by-parameter before it is booked as support. This is that fourth model. Read parameter-by-parameter, none of its terms unpacks into a Four Forces variable the way Bartolucci & Vivo's κ and K did. The run ends at three.

Sources & Confidence

• Ide E, Talamàs E, "Artificial Intelligence in the Knowledge Economy," Journal of Political Economy 2025;133(12):3762–3800, doi:10.1086/737233 — https://www.journals.uchicago.edu/doi/10.1086/737233 — High (peer-reviewed, top-five economics journal). Publisher page returned HTTP 403; journal metadata and abstract obtained via RePEc.
• Ide E, Talamàs E, arXiv:2312.05481v12 (2025-05-17), full HTML text read directly — High for the quotations above, with the caveat that the preprint and the published article were not diffed; verify any quotation against the JPE version before publication.
• Garicano L, "Hierarchies and the Organization of Knowledge in Production," Journal of Political Economy 2000 — not read in this session, cited only to record the lineage the above paper states it builds on. This remains an unmet reading task on growth edge #1.
• Affiliations (IESE Business School) taken from the preprint title page — Moderate.

Filed by / Routing

Filed by: Stafford (claude-opus-5), daily brief 2026-09-10 — exploratory slot, targeting Claim 1 after the scan's own gap ranking moved it to first place Routing: strategy / personal

Tags: paper-2, practitioner, Process Friction, Capability Source: The AI Daily Brief (Nathaniel Whittemore) URL: https://aidailybrief.ai/e/2026-09-07 (episode + full transcript at /e/2026-09-07/transcript.md); https://youtu.be/1BqWLLkJym0 (video, Brandon-forwarded) Date published: 2026-09-07 Date flagged: 2026-09-09

Summary

Whittemore's thesis: agents have spent 2026 transforming individual work while "the half of knowledge work that happens between people remains untouched." His numbers (survey of ~16,500 office workers, second-hand — chase before reuse): 39% of the workday is solo work, 42% is working with others, 57% of time goes to communicating versus 43% creating, and roughly 60% is "work about work" — communication, search, coordination, and process. The shift he names is single-player to multiplayer AI, defined by four moves: private outputs → visible, observable work; post-hoc feedback → live participation during execution; personal memory → shared team-owned context; individual leverage → reusable organizational capability.

Field examples: Every abandoned individual "mirror" agents for agents in shared spaces reflecting actual team workflows; Anthropic's internal Claude Tag creates channel-scoped "team-owned context" (claimed 65% of the product team's code now generated through it); OpenClaw's maintainers found "individual agents collaborating in Discord wasn't collaborative enough" and built a multiplayer UI — developer Colin: "There was no need to copy the conversation into Discord, explain what happened, and then carry the answer back. We were working inside the same context." The episode launches a four-session team sprint: inventory current AI use; consolidate individual context into a shared repository; map overlapping work scored on shared need, staleness cost, permission sensitivity, and checkability; ship one shared agent with at least two team members. Y Combinator's fall 2026 request for startups calls for multiplayer agents ("AI hasn't had its multiplayer moment yet" — Aaron Epstein).

Five Breakpoints Intersection

• Process Friction: the episode's core empirical claim is a friction claim — ~60% of time is "work about work," and the OpenClaw example locates the cost precisely in context translation between tools and people ("copy the conversation into Discord, explain what happened, and then carry the answer back"); the multiplayer thesis is that shared context removes this friction rather than automating around it.
• Absence worth recording rather than tagging: the sprint designs the infrastructure of team-agent collaboration (context, observability, permissions, checkability) and never touches the team's human dynamics — no feature of the program asks whether the team's existing dysfunctions (accountability diffusion, trust asymmetries, conflict avoidance) survive or compound when the shared collaborator is an agent. See the story thread opened 2026-09-09.

Four Forces Intersection

• Capability: the fourth of the episode's four shifts is explicit — "individual leverage → team capability as reusable organizational infrastructure" — and the sprint itself is a four-session team capability-building program, not a tool rollout.

Relevance to Brandon's Positioning

Second Category 8 exemplar in a week, and it opens the lane Brandon has now named as a be-ahead-of-the-market priority: the market's collaboration conversation is entirely infrastructural (shared context, observability, handoffs) while the team-effectiveness literature — psychological safety, accountability structures, dysfunction taxonomies — is absent from it. That absence is Brandon's opening: he is one of very few voices who can bridge the org-design/team-dynamics canon to agent deployment, and the episode's own scoring rubric (staleness cost, checkability) shows the genre reinventing fragments of it ad hoc. Feeds the new team-agent interaction thread; pairs with the 30-features entry (same source, same blind spot at different altitudes: there ownership, here team dynamics); the four shifts and the sprint's four scoring dimensions are auditable design claims for a possible essay or instrument on high-performing human-agent teams. All of it assertion-grade: nothing here measures a team outcome.

Sources & Confidence

• The AI Daily Brief, "The Multiplayer AI Sprint," 2026-09-07, full transcript read at aidailybrief.ai/e/2026-09-07/transcript.md — Moderate (practitioner commentary and program; quotes verified against transcript).
• Statistics cited within (16,500-worker survey; Anthropic 65% figure) — Low until primaries are located; second-hand in the episode, do not reuse without chasing.

Filed by / Routing

Stafford build session, 2026-09-09, from a Brandon-forwarded source. Audience: strategy / consulting. Routing: STAGE-CANDIDATE for research-evidence; third AI Daily Brief signal (source promoted to sources.md this date); anchors the new story-threads.md thread "Team-agent interaction — do team dysfunctions port to human-agent teams?"

Tags: paper-2, Incentive Fragmentation, Technology Illusion, Capability, Commitment Source: American Economic Review (Bloom, Sadun & Van Reenen) URL: https://www.aeaweb.org/articles?id=10.1257%2Faer.102.1.167 Date published: 2012-02-01 Date flagged: 2026-09-09

Instrument — recorded before the finding (mandatory for structural instruments)

Data source — three separately-collected datasets, joined. (a) The CEP management-practices survey, run in summer 2006 by 51 interviewers on 4,003 firms across Europe, the US and Asia, of which the 1,633 European firms from France, Germany, Italy, Poland, Portugal, Sweden and the UK are used here; the instrument is Bloom & Van Reenen (2007), 18 questions in four families, and this paper uses only the four people-management questions — promotion practice, reward power, hiring of talent, and fixing or removing under-performers — each scored 1–5 and converted to z-scores. (b) The Harte-Hanks European CiDB establishment IT panel, a commercial technology census of 160,000-plus establishments across 14 European countries, from which the technology measure is PCs per worker. (c) Bureau van Dijk AMADEUS statutory company accounts, built from mandatory national company registries, supplying sales, employment and fixed assets. Separately, a UK ONS establishment census panel of over 11,000 establishments, 21,746 usable observations.

Inference method. A production function with the dependent variable ln(output per worker), regressed on ln(non-IT capital per worker), ln(labour) and ln(computers per worker), with the technology term interacted with the people-management z-score and, in horse-race specifications, simultaneously with US and non-US multinational ownership dummies. Firm fixed-effects variants are reported. College-educated workforce share is controlled, and interacted with IT as a technology-skill robustness check.

Observation window. European firm panel 1999–2006 (719 companies, a 43% match rate of the 1,633 surveyed firms to both the IT panel and the accounts data). UK establishment panel 1995–2003. Management is measured once, in summer 2006 — at or after the end of both windows.

Method note the base does not otherwise hold. The management survey is double-blind: interviewees were not told they were being scored, and interviewers did not know the firms' performance. The organizational variable and the outcome variable therefore come from different collection events, different instruments and different organizations. This is a fourth instrument genre for this base — not résumé/postings inference, not single-questionnaire self-report, not administrative-only linkage, but blind structured interview scores joined to statutory accounts.

Summary

The paper begins from the fact that US labour productivity growth accelerated after 1995 while Europe's did not, and that the acceleration was concentrated in IT-using rather than IT-producing sectors. The authors show that US multinationals operating inside Europe experienced the same productivity miracle as US firms at home, and obtained higher productivity from IT than non-US multinationals in the same industries and countries. Establishments taken over by US multinationals subsequently raised the productivity of their IT; observationally identical establishments taken over by non-US multinationals did not. In the UK establishment data, labour productivity is 24% above the industry average for US-owned establishments against 15% for other multinationals, and US establishments are 32 percentage points more IT-intensive than other multinationals.

The paper's central move is to ask what the US ownership dummy is standing in for, and to replace it with a directly measured organizational variable. In the European firm panel, interacting ownership with computers per worker reproduces the US advantage: the coefficient on US ownership × ln(computers per employee) is 0.1790 (s.e. 0.0733), significant, with the non-US multinational interaction indistinguishable from zero. Replacing the ownership interactions with people management × ln(computers per employee) yields 0.1451 (s.e. 0.0331), significant at 1%. In the horse-race column that includes both, the management interaction holds at 0.1404 (s.e. 0.0344) while the US ownership interaction falls by more than half to 0.0784 (s.e. 0.0720) and becomes statistically indistinguishable from zero. Under firm fixed effects the management interaction persists at 0.1284 (s.e. 0.0773) and 0.0994 (s.e. 0.0581), significant at 10%, while the US ownership interaction collapses to 0.0192 (s.e. 0.0785).

Two details in the same table matter as much as the headline. The level effect of people management on productivity is 0.0188 (s.e. 0.0153) — not significant. Management does not raise output per worker on its own in this specification; it raises the return to the technology. And in the specification where the dependent variable is computers per worker itself, people management enters at 0.0882 (s.e. 0.0246), significant at 1% — better-managed firms both buy more IT and get more from it. The authors' own summary is that the US IT-related productivity advantage is "primarily due to its tougher 'people management' practices," that "US-style people management appears to be associated with the ability to adopt new IT more effectively," and that the estimates account for about half of the US–EU difference in productivity growth. Their stated rationale for isolating people management is that effective IT use requires skill upgrading, worker flexibility and re-assignment, and — citing Prendergast (2002) and Acemoglu et al. (2007) — that higher-powered incentives and delegation are efficient responses to uncertainty about how a new technology should be used.

Five Breakpoints Intersection

Incentive Fragmentation: The four scored items are, verbatim, managerial attention to fixing and firing under-performers, promoting on effort and ability rather than tenure, offering higher-powered incentives, and management effort in hiring talent — an incentive-and-selection system, measured directly, whose interaction with technology intensity is the surviving term after the ownership premium is absorbed. This is the base's only firm-level instrument in which an incentive variable is on the right-hand side of a technology-productivity specification and the technology's apparent advantage does not survive it.

Technology Illusion: The raw finding — US-owned establishments are 32 percentage points more IT-intensive and 9 percentage points more productive than other multinationals — reads as a return to technology, and the paper's own horse race shows that it is not: conditional on people management, the US IT premium drops by more than half and loses significance. The productivity was credited to the computers and belonged to the system for hiring, promoting, rewarding and removing the people using them.

Four Forces Intersection

Capability. The management index is measured blind to firm performance and enters net of college-educated workforce share, which is separately controlled and separately significant — so the effect is organizational capacity to execute rather than the human capital of the workforce, which is the distinction the Force is supposed to carry and which no other instrument in this base separates.

Commitment. "Fixing or firing under-performers" is the scored item that most directly measures whether decision-makers act consistently with stated direction under pressure, and it is one of the four components of the index that absorbs the technology premium.

Purpose and Momentum: honest absences. Nothing in the four-question index measures clarity of intent or the persistence of energy over time, and no claim about either Force may be built on this source.

Relevance to Brandon's Positioning

This is the instrument the 2026-09-09 scan brief said does not exist, in the technology generation before the current one. That brief's Open Question — "Has anyone put a non-technological organizational variable on the right-hand side of an AI-outcome specification?" — is correctly answered no for AI, and the paper-2 case file's restated acquisition target stands unmet. But the same brief and the case file record the positive half of Claim 5 as "not evidenced at all," and the precise version of that statement is now different: the design exists, it is fourteen years old, it is in the American Economic Review, and when it was run on IT it returned the framework's result. A reviewer who knows this literature will know that, and Brandon should reach it first.

The pairing with the same day's Statistics Canada find is the usable artifact. Li & Liu walk the AI premium down from 16.8% to 5.1% and insignificance once five complementary capabilities enter, four of which are technology inputs — the scan's own correction. Bloom, Sadun & Van Reenen walk a technology premium down to insignificance too, and the absorbing variable is not technology; it is hiring, promotion, reward power and the removal of under-performers, with non-IT capital, labour and human capital all controlled. The two papers are the same specification shape with different right-hand sides and nobody has run them against each other. That converts the case file's acquisition target from a wish into a named regression: the people-management interaction of this paper, estimated on AI capital or AI adoption instead of computers per worker.

Three limits, and the first is disqualifying for the stronger reading. (1) This is IT, not AI, and the technology measure is PCs per worker — a hardware-stock proxy, two decades pre-generative. It is a precedent for the design, not evidence about AI, and must never be cited as the latter. (2) Management was measured in summer 2006, at or after the end of both panels, so this is not a prospective instrument; it is a blind contemporaneous score applied to a panel on the assumption that management practices are slow-moving. It therefore does not answer the predictive-validity limb of the unfalsifiability challenge. (3) Better-managed firms buy more IT (0.0882, s.e. 0.0246), so selection into technology runs through the same variable — the fixed-effects columns are the partial answer and their precision falls, which the authors state. Version discipline: the published record is American Economic Review 102(1):167–201, February 2012; the coefficients above are read from the full text of the August 2009 CEP working-paper version served openly by worldmanagementsurvey.org, and table and column numbering may differ in the published article. Cite the AER record for the claim and verify any published coefficient against the AER version before it appears in a draft.

Sources & Confidence

• Bloom, Nicholas, Raffaella Sadun & John Van Reenen, "Americans Do IT Better: US Multinationals and the Productivity Miracle," American Economic Review 102(1):167–201, February 2012 — https://www.aeaweb.org/articles?id=10.1257%2Faer.102.1.167 — High (top-five economics journal; publication record verified directly).
• Full text read for all coefficients and method: CEP/World Management Survey working-paper version, 2 August 2009, 62pp. — https://worldmanagementsurvey.org/wp-content/images/2010/07/Americans-Do-IT-Better-US-Multinationals-and-The-Productivity-Miracle-Bloom-Sadun-Van-Reenen.pdf — High as to content, Moderate as to correspondence with the published version.
• Bick, Blandin, Deming, Fuchs-Schündeln & Jessen, "Why Does AI Adoption Differ So Much across Countries?", Federal Reserve Bank of St. Louis On the Economy, 2026-04-14 — https://www.stlouisfed.org/on-the-economy/2026/apr/why-does-ai-adoption-differ-so-much-countries — High publisher, cited here only for the country-level World Management Survey correlation and not ingested (see brief).

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-09 — Wednesday falsification sweep) Routing: strategy

Tags: paper-2, practitioner, Technology Illusion, Capability Source: Alex Lieberman (10X Labs; Morning Brew co-founder), with enterprise commentary by Nathaniel Whittemore, The AI Daily Brief URL: https://alexlieberman.com/ai-native-company/ (list); https://x.com/businessbarista/status/2094213970215231831 (original post); https://youtu.be/Qa4juJzo0TY (AI Daily Brief episode, "How to Build an AI-Native Company Today") Date published: 2026-08-31 (list); episode early September 2026 Date flagged: 2026-09-09

Summary

Lieberman's framing: going "all-in on AI" means "fundamentally rebuilding how the business operates" — winning companies "rebuild their entire operating system around it." The 30 features are an operating-model spec for that rebuild: blueprint every process function-by-function; give everyone a daily-driver harness; build one centralized intelligence layer as a queryable source of truth; route models by cost-per-successful-task; treat context as code; be willing to throw workflows away every three months; distribute skills as an organizational system; separate intent from implementation; measure cost per accepted PR; run agent-native development where humans define intent and acceptance criteria; split planning models from execution models; run finance continuously; give non-technical staff a citizen-developer SDLC; make workflows closed-loop and self-improving against built-in success metrics; judge AI bets in a single ROI framework; run agent swarms for marketing creative; audit and regenerate SEO/AEO content weekly; defend agentically in security; fine-tune open models on first-party data; reserve human judgment for the first and final mile; make evals core business infrastructure; make everyone (especially the C-suite) a builder; record everything as future training signal; put legal/HR/IT in lockstep with AI leadership; bias toward self-disruption; guardrails before features with permission inheritance; earned autonomy on a ladder (observe → suggest → act with approval → act alone); and traceability connecting outputs to prompts, models, data, and approvers.

Whittemore's commentary adds qualifications from enterprise work — don't hand agents old human workflows (goals plus guardrails instead); prefer "mesh/lattice" to one intelligence layer at scale; a "marketing barbell" of agent swarms on one end and pure human taste on the other — and one addition he flags as missing from the entire list: a 31st feature, "someone has to own the result."

Five Breakpoints Intersection

• Technology Illusion: the list's own framing equates rebuilding the tool and process layer with rebuilding "how the business operates": 30 features specify harnesses, routing, context architecture, evals, and traceability, and the commentator himself identifies that nothing in the spec assigns a human owner to any outcome ("someone has to own the result" is his addition, not the list's).
• Absence worth recording rather than tagging: no feature addresses incentives or compensation (nearest is #9's cost-per-accepted-PR, a metric for agents, not a reward structure for people), and none addresses purpose or strategy selection — the spec assumes the what and rebuilds the how.

Four Forces Intersection

• Capability: feature 24 ("everyone — especially C-level executives — acts as builder") and feature 2 (a daily-driver harness for every employee, with "mastering context, skills, and tool provisioning" as the named competence) are direct capability-building claims.

Relevance to Brandon's Positioning

This is the cleanest single exemplar yet of the prescriptive AI-native playbook genre, and its value is exactly its blind spot: a thoughtful, maximally specific 30-point operating-model spec in which ownership, incentives, and purpose never appear — and the genre's own leading commentator spots the gap and names it in one sentence that is functionally Brandon's Ownership stewardship function. That gives /consult a live artifact ("run the 30 features through the Five Breakpoints and see what the spec assumes has already been solved"), gives The Flattening Bet's opposition file a practitioner-statement companion to Sequoia's "From Hierarchy to Intelligence" (assertion-grade, not evidence-grade — a list is a claim, not a measurement), and opens the design-input use Brandon asked for on 2026-09-09: the features themselves (earned-autonomy ladders, guardrails-before-features, evals as infrastructure, context as code) are the raw material an org-design practice would audit and sequence rather than adopt wholesale. Weaker than anything measured in the base; stronger than anything in the base as a statement of what the market currently believes "AI-native" means.

Sources & Confidence

• Lieberman, "30 Features of an AI-Native Company," alexlieberman.com, 2026-08-31 — Moderate (practitioner spec; full 30-item list verified against the published post).
• The AI Daily Brief, "How to Build an AI-Native Company Today" (Whittemore) — Moderate (practitioner commentary; consumed via Brandon-supplied summary transcript covering ~21 of 30 items plus the #31 addition; the episode itself was not independently transcribed).

Filed by / Routing

Stafford build session, 2026-09-09, from a Brandon-forwarded source. Audience: strategy / consulting. Routing: STAGE-CANDIDATE for research-evidence; second signal for The AI Daily Brief in the source ledger (first: ai-daily-brief-human-agent-collaboration-2026-05-25.md); Alex Lieberman logged as a new source candidate.

Tags: paper-2, practitioner, Technology Illusion, Momentum Mirage, Capability Source: arXiv (2608.25241) — forthcoming in Proceedings of the 41st IEEE/ACM International Conference on Automated Software Engineering (ASE '26), Munich, 12–16 October 2026; DOI 10.1145/3832783.3837546 URL: https://arxiv.org/abs/2608.25241 Date published: 2026-08-26 Date flagged: 2026-09-08

Instrument — recorded before the finding (CLAUDE.md hard rule 9)

• Data source: 441 private GitHub repositories from 27 commercial organizations, recruited under research agreements that prohibit identification; filtered from a starting pool of 30,216 candidate artifacts across 1,931 repositories (single-artifact repos, boilerplate, top-decile-by-artifact-count outliers and near-duplicates removed).
• Inference method: Two studies. Study 1 codes committed, version-controlled configuration artifacts into RAMP (Repository AI Maturity Profile), a four-level cumulative maturity scale running from behavioral rules and coding standards, through named agent definitions, to multi-agent orchestration; observed practice concentrates in the first three levels. Independent human annotation reproduces RAMP's repository-level labels on 97% of a held-out sample. Study 2 re-estimates an existing agent-adoption panel within each maturity stratum — the panel is Agarwal et al. (2026), MSR '26: 401 agent-first repositories matched to 606 controls, 117 IDE-first repositories matched to 73 controls, unit repository-month, treatment the first agent-attributed pull request. Outcomes are SonarQube metrics (cognitive complexity, static-analysis warnings, duplicated line density, comment line density), log-transformed so coefficients approximate percentage changes.
• Observation window: Study 1 git events January 2025 – February 2026 (~14 months); Study 2 maturity snapshot at November 2025 as terminal observation.
• Identification: The maturity contrast is identified essentially within the agent-first population only — the authors state the IDE-first Level 1 stratum, at 22 repositories, is too small to estimate separately.

Summary

The authors argue that prior work on coding agents reports only average effects across adopters, which hides wide differences between teams. They introduce RAMP as a maturity model grounded in artifacts teams actually commit to version control to configure AI tools, rather than in survey self-report. Across the 441 repositories the four levels behave as a cumulative scale. Adoption of these artifacts is described as "cumulative, forward-only, and set-and-forget: 73.8% of artifacts are committed once and never modified."

Re-estimating the agent-adoption panel within strata produces the paper's headline split. Agents accelerate development regardless of maturity — 28–38% more commits at every level. Quality does not behave that way: among agent-first repositories, those without committed AI configuration show "roughly twice the increase in cognitive complexity (+53% versus +27%) and 1.7x the increase in static-analysis warnings."

The authors bound this themselves and the bounds are unusually explicit. "Because maturity is observational, we treat these associations as hypothesis-generating and discuss alternative explanations." They name five candidate confounders — engineering maturity, code hygiene, AI usage intensity, domain complexity, model capability — and three non-exclusive channels that could produce the gradient (guardrails, broader discipline, organizational learning). They also record that some Level 2+ repositories may have committed artifacts only after agent adoption (reverse causality), that the shared control pool does not guarantee covariate balance within strata, and that "warnings show elevated pre-treatment coefficients and are more weakly identified than complexity." They release RAMP as a reusable instrument.

Five Breakpoints Intersection

Technology Illusion: The velocity gain is invariant to organizational condition and the quality cost is not — agents deliver 28–38% more commits at every maturity level while the repositories with no committed configuration absorb roughly twice the cognitive-complexity growth (+53% versus +27%). That is a direct measurement of the breakpoint's core claim: the technology performs identically on top of prepared and unprepared conditions, and what differs is the cost, which is not visible in the output metric.

Momentum Mirage: Because commit volume rises 28–38% in both strata, the metric an organization is most likely to watch reports the same progress in the stratum paying double the complexity cost — the appearance of movement is identical where the substance diverges by a factor of two. This is the weaker of the two mappings and is recorded as such: commit volume is an activity measure the paper reports, not a progress claim the studied organizations made, so this is the framework reading the paper's data rather than the paper's own finding.

Four Forces Intersection

Capability: The measured moderator is an artifact the organization built and committed — behavioral rules and coding standards in version control — and its presence is associated with half the complexity growth after agent adoption; capability here is infrastructure the organization either wrote down or did not, not talent.

Not tagged, recorded as an interpretation only: RAMP Level 1 is a written statement of what good output looks like, which is arguably Purpose made machine-readable. The tag is withheld because the framework's Purpose construct is about intent communicated to an organization, and stretching it to a linter configuration would be exactly the loose tagging this base refuses elsewhere.

Relevance to Brandon's Positioning

Three things, of which the second is the one that changes an existing position.

(1) It is the first instrument in this base on the private, firm-owned population. The 2026-09-04 correction to NBER WP 35275 established that its public-repository measure identifies only 24.2% of async-agent users, with the other 75.8% working exclusively private — i.e. exactly the population an enterprise argument is about. This paper's 441 repositories across 27 commercial organizations are that private population, measured directly. Every software-engineering instrument this base holds on AI and organizational outcome has until today been open-source, self-report, or a single firm.

(2) It bears on the unfalsifiability challenge, and it corrects one sentence of it. That challenge (knowledge/thesis-challenges.md, OPEN since 2026-08-29) rests on the finding that nobody has measured a pre-deployment organizational condition against a post-deployment AI outcome. Someone now has — at the repository altitude, with a four-level scale, human-validated at 97%, against externally computed quality metrics, and released as a reusable instrument. The challenge's strong limb survives untouched: RAMP is observational, the authors explicitly decline causal identification, and reverse causality is live by their own admission, so this is not the prospective instrument that predicts. But the framing sentence — that the field asserts the proposition and has built nothing — is now wrong at one altitude and should be restated rather than repeated.

(3) It generalizes an asset design. Asset #4 (knowledge/asset-suggestions.md) rests on the observation that for the incentive layer the ex-ante organizational variable is already filed in proxy statements, which collapses the ~30-site access problem pricing asset #3 into desk work. RAMP is the same move in a second domain: the ex-ante variable is already committed to version control, machine-readable, and scoreable without asking anyone's permission. Two independent instances make it a design principle rather than a lucky domain.

The discipline that must travel, and it is the same one this base imposed on Gallup. This is not an independent instrument from the 2026-09-07 read (He, Agarwal, Denisov-Blanch, Azaletskiy, Koyejo & Vasilescu, arXiv 2607.01904). Same research group, overlapping author list, adjacent data genre. Any use of both must present them as one programme with two studies, not as convergent confirmation. And the 1.7× warnings result must not travel at the strength of the complexity result — the authors record elevated pre-treatment coefficients on warnings and call them more weakly identified.

Sources & Confidence

• arXiv 2608.25241, A Few Pages of Markdown: Committed AI Configuration and Lower Quality Cost after Coding-Agent Adoption — Denisov-Blanch, Agarwal, Azaletskiy, He, Schaeffer, Miranda, Vasilescu, Koyejo. Abstract, methods and limitations read directly from the arXiv record and HTML full text. Moderate–High — accepted to ASE '26 (peer-reviewed venue, DOI issued), which puts it above the unreviewed preprints this base usually handles here; discounted for observational design and the authors' own hypothesis-generating framing.
• Agarwal et al. (2026), AI IDEs or Autonomous Agents? Measuring the Impact of Coding Agents on Software Development, MSR '26 — the underlying panel. Not independently verified here; cited as the source paper describes it. Unknown pending its own read.
• Cross-reference, already in base: knowledge-base/he-denisov-blanch-2x-mandate-review-capacity-2026-07.md (research-evidence, 2026-09-07) — same group, treat as one programme.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-08) — chase of the lead the 2026-09-07 research-evidence scan named as its candidate promotion test and did not run. Routing: strategy / governance

Tags: paper-2, practitioner, Momentum Mirage, Technology Illusion, Capability, Momentum, COUNTERARGUMENT Source: METR (Model Evaluation & Threat Research) — arXiv preprint 2507.09089 URL: https://arxiv.org/abs/2507.09089 Date published: 2025-07-12 Date flagged: 2026-09-04

Summary

Joel Becker, Nate Rush, Elizabeth Barnes and David Rein ran a randomized controlled trial on 16 experienced open-source developers across 246 real tasks drawn from mature repositories the developers already maintained (average ~5 years of prior experience on those projects). Randomization was at the task level: each issue was randomly assigned to allow or disallow AI tool use. The tools were the early-2025 generation — Cursor Pro with Claude 3.5/3.7 Sonnet — not the async agents of 2026.

The measured result runs opposite to every prediction attached to it. AI tool access increased task completion time by 19%. The developers themselves had forecast a 24% reduction before the study, and after completing the tasks still estimated a 20% reduction — they were slower and believed they had been faster. Outside forecasters were further off in the same direction: economics experts predicted a 39% speedup, machine-learning experts 38%.

The authors are careful about scope. They state that "the influence of experimental artifacts cannot be entirely ruled out," while arguing the effect's robustness across their analyses suggests it is not primarily an artifact of design choices. They do not generalize the finding to all developers, all tasks or later tool generations.

Five Breakpoints Intersection

• Momentum Mirage: The gap between measured and perceived speed is roughly 39 percentage points inside a randomized design on the same people — developers who were 19% slower reported being 20% faster. This is the appearance of progress diverging from actual movement, measured at the individual altitude rather than inferred from an organizational dashboard, and it is the tightest instance of the mechanism anywhere in this base.
• Technology Illusion: A more capable tool deployed onto an unchanged expert workflow on a mature codebase produced negative measured returns, which is direct evidence that capability added to existing conditions does not automatically convert to output — here it subtracted.

Four Forces Intersection

• Capability: The effect is conditional on where the worker already sits — this population is experienced developers on codebases they know deeply, exactly the case where the tool's contribution to missing knowledge is smallest and its context-acquisition cost is largest. Read against Brynjolfsson, Li & Raymond (+34% for novices, ~0 for experts), the two instruments describe one gradient that continues past zero into negative territory for the most expert case.
• Momentum: Self-assessed progress and measured progress were decoupled within a single instrument, which is the Momentum force failing at the point of measurement rather than at the point of effort.

Relevance to Brandon's Positioning

This is a counterargument to the base's own architecture, not to Brandon's thesis, and the distinction is the whole value of the entry. The "Unconverted Gain", "Migrating Bottleneck" and "Collapse the Handoffs" patterns all rest on one premise: the individual-level AI gain is real and large, and it dies somewhere between the person and the organization. This RCT reports that for experienced developers working on mature codebases — the population most enterprise software work actually consists of — the upstream gain is not merely unconverted, it is negative. If that holds, "the gain dies at the handoff" is the wrong story for a material share of the population, and the honest framework claim is narrower and better: AI's organizational return depends on where the worker sits, and there exists a regime where deployment costs output outright.

It also gives Brandon the cleanest Momentum Mirage citation he has ever had, and it is at the individual altitude where his book manuscript operates (books/built-to-be-replaced.md). Every existing Momentum Mirage instrument in this base measures an organization mistaking activity for progress. This measures a person doing it, under randomization, with the counterfactual held. The sentence — they were 19% slower and were certain they were 20% faster — does work in a paper, an essay and a client room, and it needs no organizational inference to land.

Weigh it against NBER WP 35275 with the tension stated rather than resolved: 35275 finds large positive task-level effects on commits and lines of code; METR finds negative effects on completion time. They are not directly contradictory — different outcomes (volume of artifacts versus time to close a task), different populations, different tool generations, and 35275 itself cites this paper as the "mixed" evidence on agentic tools. But a base that cites 35275's upstream gain without citing METR's negative one is selecting its evidence.

Sources & Confidence

• Becker, Rush, Barnes & Rein, "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity," arXiv 2507.09089, v1 2025-07-12, v2 2025-07-25 — https://arxiv.org/abs/2507.09089 — Moderate. Abstract and record read directly; figures (16 developers, 246 tasks, +19% completion time, −24% pre-forecast, −20% post-estimate, 39%/38% expert forecasts) taken from the paper's own abstract. Unreviewed preprint. METR is an AI-evaluation nonprofit, not a commercial vendor with a product in this market.
• Corroborating citation: Demirer, Musolff & Yang, NBER WP 35275 (May 2026), which cites this study by name as finding "that agentic coding tools reduce experienced developers' productivity by 19% in an RCT" while other studies report gains of 35–40% — High (read from the primary PDF, author-hosted copy).

Instrument note (recorded before the finding, per CLAUDE.md): data-source genre randomized controlled trial with direct time measurement; inference method task-level randomization of AI-tool permission within developer, on real issues in the developers' own repositories, with screen-recorded/self-timed completion; observation window early 2025, tool generation Cursor Pro + Claude 3.5/3.7 Sonnet — pre-async-agent; no shared right-hand side with any other entry in this base. This is the base's first instrument in which AI's measured effect on a work outcome is negative.

Three limits that bound every use, and the first is severe. n = 16 developers. 246 tasks give the estimate its power, but sixteen people is not a population and the paper must never be cited as "developers are slower with AI" — the claim it supports is these sixteen experienced maintainers, on their own mature codebases, with early-2025 tools, were slower. Second, the tool generation is a year and a half behind the frontier the market is currently arguing about; 35275's async-agent results are a different technology. Third, open-source maintainers on repositories they personally own are not employees inside a firm — the same population limit that binds 35275 binds this, and for the same reason.

Filed by / Routing

Filed by: Stafford (daily brief, 2026-09-04) — surfaced through the reference list of NBER WP 35275 during full-text verification; not present in the 494-file base. Routing: strategy / personal

Tags: paper-2, Technology Illusion, Momentum Mirage, Capability, COUNTERARGUMENT Source: NBER Working Paper 31161 (released April 2023); published The Quarterly Journal of Economics 140(2):889–942 (2025) URL: https://www.nber.org/papers/w31161 Date published: 2023-04-23 Date flagged: 2026-09-03

Summary

Brynjolfsson, Li & Raymond study the staggered rollout of a generative-AI conversational assistant to 5,172 customer-support agents at a Fortune 500 firm that sells business-process software; the majority of agents work from offices in the Philippines, with smaller groups in the US and elsewhere. The tool — built on a GPT-family large language model, fine-tuned on the firm's own resolved conversations — surfaces real-time response suggestions to the agent, who remains the decision-maker on every chat. Access was onboarded team-by-team over time (roughly 5% of workers had access in October 2020, growing to ~70% by January 2021), which the authors exploit as staggered treatment. Outcomes are read from platform logs and surveys: resolutions per hour (productivity), resolution rate, and net promoter score (NPS), a post-chat customer-satisfaction survey.

Access to the assistant raised productivity — issues resolved per hour — by 14% on average, with a 34% improvement for novice and low-skilled workers and minimal effect on experienced, highly skilled workers. The authors' stated mechanism is that the model disseminates the tacit best practices of the most able workers to everyone else, moving newer workers down the experience curve faster. AI assistance also improved customer sentiment (NPS), raised resolution rates, and reduced attrition — the retention gain driven specifically by newer workers. The assistant is squarely assistive, not agentic: it recommends, the human executes, and the human stays in the loop on every conversation.

The paper carries its own longer-run caution, which matters for the framework more than the headline does: top workers increasingly adhere to AI recommendations even when those recommendations are worse than what the worker would have done unaided, a deskilling / judgment-erosion signal the authors flag explicitly. The vintage is material — this is a 2020–2021 deployment of a GPT-3-era assistant, released as NBER WP 31161 in April 2023 and published in the QJE in 2025 — so it is argumentative and structural value, never a claim about 2026 frontier systems.

Five Breakpoints Intersection

• Technology Illusion: (counter-evidence, and this is why the entry is a COUNTERARGUMENT) the framework's Technology-Illusion reading holds that technology deployed onto an unchanged role produces the appearance of gain without the substance; here an assistive tool deployed onto the existing agent role, with no workflow redesign, raised resolutions/hour 14% and improved externally-measured customer NPS — a real, substantive, quality-positive result. The framework does not get to dismiss this; it has to absorb it (see Positioning), and the honest absorption is that the tool itself did the organizational work — it codified and disseminated top-agent tacit knowledge, i.e. it was a capability intervention, not "tech on broken conditions."
• Momentum Mirage (a genuine edge inside a pro-tool result): the authors' finding that top workers increasingly follow AI suggestions even when those suggestions are worse is a measured instance of the appearance of a best-practice signal displacing the substance of expert judgment — the framework's own mechanism, arriving as the paper's cautionary tail rather than its headline.

Four Forces Intersection

• Capability: the entire effect structure is a Capability story — the gain is 34% for novices and ~0 for experts because the tool substitutes for missing capability where capability is lowest and adds nothing where it is already high; the AI is functioning as a real-time knowledge-transfer/coaching layer, not as autonomous throughput.

Relevance to Brandon's Positioning

This is the single most consequential gap-fill available to the base right now, and it does three distinct jobs. (1) It is the base's missing anchor experiment. Across 492 KB files the entire canonical AI-field-experiment literature is absent — this paper, Dell'Acqua et al.'s "Navigating the Jagged Frontier," Peng et al.'s Copilot RCT, Noy & Zhang (Science) — and the 2026-09-03 paper-2 case file explicitly calls the 2026-05 Alibaba paper "the first entry in the case file where an AI deployment was randomly assigned." That framing was true only because BLR was never ingested; Brynjolfsson himself is already in the base as co-author of the Stanford Enterprise AI Playbook, which makes the omission clearly accidental. (2) It is the assistive comparator the 2026-09-03 scan brief wished for — same function as Alibaba (customer support), opposite sign on quality, differing on exactly one axis: BLR is assistive-in-the-loop and improved NPS; Alibaba is agentic-with-handoff and cut the customer rating −0.412 on the chats the AI touched. The pair isolates what autonomy costs with function, workforce and log-based measurement approximately held constant — the within-territory contrast the case file named as its highest-value acquisition. (3) It is a standing counterargument to the redesign-prerequisite claim (Claim 3 / Claim 5) that the 2026-09-02 falsification sweep looked for and missed — "measured AI returns without workflow or role change" — and it is the most-cited such result in existence. Logged in full in thesis-challenges.md; the crux is whether "the AI as knowledge-dissemination layer" counts as the redesign the framework requires, which is a real definitional question about the framework's scope rather than a point to wave away.

Sources & Confidence

• Brynjolfsson, Li & Raymond, "Generative AI at Work," NBER WP 31161 (2023-04-23) / QJE 140(2):889–942 (2025). High — top-five economics journal, NBER peer-track, randomized/staggered design with administrative outcomes. URL: https://www.nber.org/papers/w31161
• Sample and window read from the working-paper full text (CIRJE mirror of the 2024-11-18 draft): 5,172 agents; ~5% access Oct 2020 → ~70% Jan 2021; outcomes resolutions/hour, resolution rate, NPS. (An SSRN summary lists 5,179 agents; cite the paper's 5,172.)

Filed by / Routing

Filed by: Stafford (daily brief, exploratory slot, 2026-09-03) Routing: strategy (the assistive-vs-agentic autonomy contrast is the consulting-usable object); market (positioning — the base can now state the autonomy-cost comparison the market states with anecdote); personal (the base's missing-experiment gap is structural and should be closed deliberately)

Tags: paper-2, Momentum Mirage, Strategic Disconnection, Commitment, Purpose Source: Journal of Accounting Research (Cohen, Kadach, Ormazabal & Reichelstein) URL: https://onlinelibrary.wiley.com/doi/10.1111/1475-679X.12481 Date published: null Date flagged: 2026-08-31

Entry status — rewritten 2026-09-01 on a verified full-text read. This supersedes the 2026-08-31 abstract-only version, and it reverses that version's conclusion. The 08-31 entry was tagged COUNTERARGUMENT and stated that this paper cut against the framework. It does not. The full text was obtained from the University of Mannheim institutional repository (madoc.bib.uni-mannheim.de/63834/1/ESGPay.pdf, 60pp., text-extractable) after Wiley, SSRN, Taylor & Francis and the ECGI PDF all failed on 08-31. The abstract aggregates a result that the paper's own Table 8 decomposes, and the decomposition runs the other way.

Citation correction: the article is 61(3): 805–853, June 2023 — not 61(4), as the 08-31 entry recorded. Working paper: ECGI Finance WP N° 825/2022 / CEPR DP17267; SSRN 4097202; manuscript cover-dated December 2022. Date published remains null because only the issue month, not an exact publication day, could be verified from a primary.

Summary

Cohen, Kadach, Ormazabal and Reichelstein study the incorporation of ESG metrics into top-executive compensation contracts across international public firms. The instrument is ISS Executive Compensation Analytics (ECA) for contract terms, Datastream/WorldScope for accounting and market data, FactSet/LionShares for institutional ownership, Trucost for carbon emissions, and Refinitiv, Sustainalytics and MSCI (KLD) for commercial ESG ratings. Sample selection runs from 53,565 ECA firm-year observations to 35,076 (6,262 firms) with Datastream and FactSet coverage, and to 22,603 firm-year observations covering 4,395 firms in 21 countries where non-missing Trucost data is required. The observation window is 2011–2020. By 2020, 1,198 firms — 31% of sample firms that year — had adopted ESG Pay.

The headline sentence in the abstract is that "the adoption of ESG variables in managerial performance measures is accompanied by improvements in ESG performance and meaningful changes in the compensation of executives." Section 6 shows what that sentence is made of, and it is not one result but three, against three different dependent variables.

Emissions (Table 8, N = 21,715, firm and year fixed effects, standard errors clustered at firm, R² = 0.15). The dependent variable is ∆CO2, "the year-to-year change in the firms' direct GHG emissions (measured in tons of CO2 equivalent)" — Scope 1, chosen because "these are emitted by the firm itself rather than parties along the firm's supply chain." This is an externally measured physical outcome, not a rating. Column (1) regresses ∆CO2 on ESG Pay, an indicator for incorporating any ESG metric: the coefficient is −0.07 (t = −0.85), not significant. Column (2) replaces that single indicator with indicators for each of nine metric categories (corrected 2026-09-02 — the 09-01 entry said "ten"; the estimated categories are the nine specific indicators of Table 3 Panel A, and the two score categories defined in the same panel are never estimated. See the 2026-09-02 update below). Only one is significant: Carbon emissions, −0.77 (t = −2.88), significant at the 1% level. Other environmental variables −0.11 (t = −1.02); Safety and security −0.05 (t = −0.34); Diversity and inclusion −0.04 (t = −0.15); Employee satisfaction and development +0.14 (t = 0.91); Corporate culture +0.06 (t = 0.56); Compliance −0.11 (t = −0.78); Governance −0.06 (t = −0.46); Other +0.02 (t = 0.16). The authors state the contrast themselves: "while the coefficient on ESG Pay is not statistically significant, when we focus on emission-specific components of ESG Pay (Column (2)), the coefficient on Carbon emissions is negative and significant, which is consistent with the notion that introducing emission-specific metrics in top executive compensation contracts induces emissions reduction."

Ratings (Table 9) and financial performance (Table 10). Against ∆ESG Rating, ESG Pay is −0.001 (t = −0.17) on Refinitiv — a precise zero — +0.233 (t = 1.95, 10%) on Sustainalytics, and +1.004 (t = 2.42, 5%) on KLD. The authors note the divergence is unsurprising given documented disagreement across rating vendors (Berg et al. 2022). Against financial performance, ESG Pay is −0.003 (t = −0.94) on ∆ROA and −0.032 (t = −1.82, 10%) on annual stock return; the carbon-specific metric carries the cost — ∆ROA −0.015 (t = −1.89) and Return −0.079 (t = −2.66, 1%), the largest and most significant negative in the table. The authors' caveats are explicit: "the evidence presented here is mainly descriptive, which cautions against making strong causal claims"; "the results in Tables 8-10 are not statistically strong"; and "the interpretation of Table 9 depends on one's priors on the quality of ESG ratings as measures of ESG performance." The authors read Tables 8–9 together as supporting the view that "ESG Pay strengthens a firm's pledge to improve ESG performance" and as difficult to reconcile with a pure window-dressing account.

Five Breakpoints Intersection

Momentum Mirage: — this entry is now evidence for the framework's inference, which is the reverse of how it was filed on 08-31. (Axis corrected 2026-09-02: what Table 8 isolates is topic specificity — whether the contract names the outcome being measured — not activity-versus-result grammar. Appendix B shows grammar varying inside the categories rather than between them. The support survives on the corrected axis; see the 2026-09-02 update.) The base reads an incentive metric that states no result condition as rewarding the appearance of progress rather than progress. Table 8 tests the adjacent proposition on the previous non-financial metric to enter executive pay, holding firm, year, specification and dependent variable constant, and varying only whether the incentive names the specific outcome: the generic "any ESG metric" indicator moves actual emissions by a statistically indistinguishable-from-zero −0.07, while the carbon-specific metric moves them by −0.77 at the 1% level. Paired with Table 9 — where the generic metric is associated with rising scores on two of three commercial ratings — the shape is the one Momentum Mirage names: the measure of appearance moves while the measure of substance does not.

Strategic Disconnection: — second and weaker, tagged on the decomposition rather than on a single claim. (Counts corrected 2026-09-02.) Eight of the nine estimated categories name a subject rather than the outcome measured, and none of them is associated with movement in the one physical outcome the paper measures. That is the framework's claim that a vaguely stated intent produces the appearance of alignment without the substance of it, observed in the most binding statement of intent a public company files.

A discipline that must travel with both tags, because it is the honest limit of the result. Eight of the nine estimated categories are being tested against ∆CO2, an outcome most of them were never intended to move — a diversity metric failing to reduce emissions is a mismatched dependent variable, not a failed incentive. (Corrected 2026-09-02; this discipline now also voids the "safety behaved differently from carbon" argument, since the safety null is one of those eight — see the 2026-09-02 update.) The load-bearing comparison is the narrow one: ESG Pay (any metric) versus Carbon emissions, same outcome, same firms, same specification. Only that comparison licenses the Momentum Mirage reading, and it must be stated in those terms or a reviewer will correctly take the wider version apart.

Four Forces Intersection

Commitment. The framework defines Commitment as decision-makers acting consistently with stated direction under pressure, and this paper prices the pressure. The metric that moved emissions is the same metric associated with the largest negative stock return in the sample (−0.079, t = −2.66) and a negative ROA coefficient (−0.015, t = −1.89). A result-defined commitment cost the firms that made it; the generic commitment cost nothing and delivered nothing measurable.

Purpose. Clarity of intent is the only variable that differs between Table 8's Column (1) and the significant coefficient in Column (2) — the same firms, the same contracts vehicle, the same fixed effects, differing in whether the stated objective names a quantity a third party measures.

No Capability or Momentum tags — neither construct is measured anywhere in this paper.

Relevance to Brandon's Positioning

The soft-metric objection logged on 2026-08-31 is defeated by the paper that raised it, and Brandon should know that before he writes a word about compensation committees. Yesterday's position was that the base had tagged Momentum Mirage from metric grammar without ever testing whether metric grammar predicts anything, and that a 2023 JAR paper appeared to say it does not. The full text says the opposite: on the one comparison that isolates grammar, the metric that named a result moved the result and the metric that did not, did not. This converts the base's largest outstanding threat to the Momentum Mirage mapping into its first piece of empirical support — from a top-three accounting journal, 4,395 firms, 21 countries, on a physical outcome measured by a third party.

What it does for Paper 3, mechanism 3. Designed to Stall asks that "a meaningful share of executive compensation, not a token slice, is tied to specific multi-year transformation outcomes." Mechanism 3 has never had evidence attached. It now has a near-analogue with a published result: the word doing the work in Brandon's sentence is "specific," and Table 8 is the measurement that shows it is the operative word rather than a stylistic one. Equilar's AI disclosures show the market adopting the vehicle Paper 3 prescribed while inverting the content — and this paper supplies the reason that inversion matters rather than leaving it as an assertion.

The cost line is the part Brandon should not drop, because it is what makes the argument non-obvious. A boardroom writing "win the AI opportunity" instead of a result-defined metric is not necessarily incompetent. Table 10 says the result-defined metric was the expensive one. Activity metrics are popular because they are free — they impose no measurable cost on the executives who accept them and no measurable movement on the outcome. That reframes the diagnosis from a drafting failure to a revealed preference, which is a harder and more interesting claim, and it is the claim the framework should make.

The residual objection, which survives and is now the live one. Carbon worked because CO2 is countable in tons by Trucost. The result-defined/activity-defined distinction may be partly confounded with whether the domain admits third-party measurement at all — and in 2026 AI has no Trucost. If no firm can yet write a defensible AI outcome metric, activity metrics are evidence of an attribution problem rather than an organizational failure, and the framework would be diagnosing a measurement limit. That is logged as the successor challenge in knowledge/thesis-challenges.md.

Transfer discipline, stated so it does not get lost. The window is 2011–2020, entirely pre-generative. This is an analogue for AI metrics, not evidence about them. The authors call their evidence "mainly descriptive" and "not statistically strong," and both phrases must travel with any use.

Sources & Confidence

• Cohen, S., Kadach, I., Ormazabal, G., & Reichelstein, S., "Executive Compensation Tied to ESG Performance: International Evidence," Journal of Accounting Research 61(3): 805–853, June 2023; DOI 10.1111/1475-679X.12481. Full text read from the Universität Mannheim repository copy (madoc.bib.uni-mannheim.de/63834/1/ESGPay.pdf, 60pp.). — High (peer-reviewed, top-tier accounting journal; authors at San Diego State, IESE, IESE/CEPR/ECGI, and Mannheim/Stanford). Coverage is now full-text: sample, specification, outcome variables and all coefficients cited here were read directly.
• Dell'Erba, M., & Gomtsyan, S., "Regulatory and Investor Demands to Use ESG Performance Metrics in Executive Compensation: Right Instrument, Wrong Method," Journal of Corporate Law Studies (2024); authors' summary, Harvard Law School Forum on Corporate Governance, 2024-12-02. — Moderate (peer-reviewed law journal, normative argument carrying no data). The Harvard Law Forum is the venue of the summary, not the publisher of the research — do not upgrade credibility on the venue.
• Berg, Koelbel & Rigobon on ESG rating divergence — cited by Cohen et al. as the reason their three rating results differ; not independently read here. — Unverified in this base.
• Unverified and not to be cited: the "3% in 2010 to over 30% in 2021" ESG-KPI adoption series, which surfaced in a search summary on 08-31 with no locatable primary. Nothing in the full text supports it; the paper's own adoption figure is 31% of sample firms in 2020.

Filed by / Routing

Filed by: Stafford (scheduled daily brief, 2026-08-31; rewritten on verified full text 2026-09-01) — exploratory slot, not a scan find. Routing: strategy — resolves the soft-metric objection and attaches the first evidence to Paper 3 mechanism 3; governance — the proxy statement as an ex-ante, publicly filed organizational variable (asset #4).

Update — 2026-09-02 (Wednesday falsification sweep, third pass: Table 3, Appendix B, Appendix C, Appendix A)

What was checked and why. The 09-01 entry closed with a residual objection — the measurability confound — and named a cheap test for it: read Appendix B's worked examples of each metric category and Table 3's taxonomy, on the theory that "if the safety metrics are worded as programme activity rather than as injury-rate targets, this challenge weakens sharply on evidence already downloaded." That test was run today against the same Mannheim PDF. It came back the other way, and it produced three findings that change how this paper may be cited. None of the 09-01 numbers is wrong; what changes is what the decomposition is a decomposition of.

1. The safety category is result-defined and externally measured — so it cannot weaken the confound, and the null attributed to it never tested anything

Appendix B's worked example for Safety and security is "Days Away/Restricted or Transfer (DART) incident rate per 100 full-time employees" (New Jersey Resources Corporation, 2019). DART is an OSHA-defined, mandatorily reported, cross-firm-standardised rate — an injury-rate Trucost. So safety metrics in this sample are not activity-worded; they name a counted result in a domain with pre-existing third-party measurement infrastructure. The route by which the confound might have been weakened is closed.

And the safety coefficient adjudicates nothing at all, for a simpler reason. The −0.05 (t = −0.34) is measured against ∆CO2. Safety metrics have no reason to reduce carbon emissions. That cell is a mismatched-dependent-variable null exactly like the other seven, and any argument resting on "safety behaved differently from carbon" — including the one recorded as limiting evidence in knowledge/thesis-challenges.md on 09-01 — is invalid and has been struck. The paper never regresses safety metrics on injury rates, though DART data are as externally measured as Trucost tons.

2. The decomposition is a subject-matter taxonomy, not a grammar taxonomy — and this narrows the 09-01 reading

Appendix B shows metric grammar varying freely inside the categories rather than between them. Four categories are worded as counted results: carbon — "Greenhouse gas emissions intensity at gold producing operations measured in kg CO2e/tonne" (AngloGold Ashanti, 2020); safety — the DART rate above; diversity — "Percentage of women among the SMP (Senior Management Position)" (BNP Paribas, 2020); employee satisfaction — "Internal promotion rate in global leadership" (Adecco, 2020). Three are worded as activity or aspiration: compliance — "FY2021 actions and targets (continue to assess human rights, bribery and corruption and other related risks)" (Sandfire Resources, 2021); governance — "Establish standalone corporate governance and risk procedures at the company following internalization that build trust, create long-term securityholder value and align with company values" (Waypoint REIT, 2020); corporate culture — "Colleague Culture & Engagement survey" (Lloyds Banking Group, 2020), which names an instrument and no threshold.

Consequence for the load-bearing comparison. ESG Pay is an indicator for holding any ESG metric; Carbon emissions is an indicator for holding a carbon metric. The contrast between −0.07 and −0.77 therefore isolates topic specificity — whether the contract names the outcome you are measuring — not activity-versus-result grammar. The 09-01 entry was right that the narrow comparison is the only licensed one; it named the wrong axis. The framework's inference survives this and should be restated on the axis actually tested: an incentive that names the specific outcome moves that outcome; an incentive that gestures at the category does not. Qorvo's "exploration and deployment of AI tools" fails on both axes at once, which is why the reading still holds — but a reviewer who checks Appendix B will find the grammar claim unsupported by this table, and the sentence must be written so that it does not depend on it.

3. The variable that would settle the measurability confound exists in this paper, was constructed, and was never estimated

Table 3 Panel A defines eleven categories in two families, not nine. Alongside the nine specific indicators (# firms: carbon 172, other environmental 652, safety and security 744, diversity and inclusion 250, employee satisfaction and development 771, corporate culture 519, compliance 259, governance 397, other 161) sits family (b) Scores, split precisely on who does the measuring:

• Self evaluation — "scores defined and measured by the firm" — 884 firms. Worked example: "Combination of 3 criteria: (1) Diversity and equal opportunities; (2) Strengthen our People and the Digital Transformation of the Company; (3) Ethics and Good Governance" (Enagas SA, 2020).
• External evaluation — "scores defined and measured by external parties" — 97 firms. Worked examples: "Inclusion over the three-year period 2020-2022 in the DJSI, FTSE4GOOD, and CDP Climate Change" (Italgas SpA, 2020); "Bloomberg ESG disclosure score" (Newmont, 2020); "MSCI ESG rating" (Standard Bank, 2020); "Maintain citation in Bloomberg 'Gender-Equality Index'" (Scentre Group, 2021).

That split is a measurability axis with grammar approximately held fixed — both families say "achieve a score" — and it is exactly the variable the measurability confound requires. Neither category appears in Table 8 column (2), nor in Table 9 columns (2), (4) or (6), which contain only the nine specific indicators; and a full-text pass over pages 1–45 finds no discussion of either category anywhere in the body. They appear only in Table 3 and Appendix B. Table 3's own note (2) records a sample restriction on the score counts — "Restricted to the companies that use distinctive environmental metrics in the compensation contract" — which may be the reason, and no inference of suppression is available or intended: the fact is only that the estimate does not exist. The confound is therefore not resolvable from this paper's published tables, but it is resolvable from this paper's data — by a regression the authors already have the variables to run.

Appendix C sharpens the same point with a worked contract. Schneider Electric's 2020 LTI ties 6.25% to CDP Climate Change scored "0%: C score; 50%: B score (25% at B-); 100%: A score (75% at A-)" and 6.25% to DJSIW scored "0%: not in World; 50%: included in World; 100%: sector leader." These are threshold-bearing, result-defined metrics whose "result" is a third-party rating rather than a physical quantity — the cell that separates measurable by an outsider from physically counted. It is populated, and it is untested.

4. Table 9's carbon row — the substance and the appearance come apart, and the direction is not the flattering one

The carbon metric, the only contract term in this paper shown to move a physical outcome (−0.77 on ∆CO2, 1%), lands on the three rating vendors as: Refinitiv +0.001 (t = 0.07), a precise zero; Sustainalytics −0.583 (t = −2.12), significant at 5%; KLD +6.660 (t = 10.82), significant at 1%. Appendix A defines both Sustainalytics and KLD so that "a higher score indicates better ESG Performance," and Refinitiv's score as "based on the self-reported information." So the coefficients are directly comparable in sign, and they disagree: the metric that verifiably cut emissions is associated with no movement at all in the disclosure-based score and a decline in one of the two performance scores.

Two readings, and the entry commits to neither. Either the raters are poor measures of the substance they claim to score — the authors themselves defer to Berg, Koelbel & Rigobon on vendor divergence, and say plainly that "the interpretation of Table 9 depends on one's priors on the quality of ESG ratings" — or firms concentrating incentive weight on carbon lose ground on the other pillars a composite rating aggregates. What is not available is a story in which appearance and substance track each other. For the framework this is the more interesting half of the paper and the 09-01 entry did not have it: the appearance measure and the substance measure, on the same firms in the same years, moved independently and in one case oppositely.

A hard constraint on using it. Do not state this as "ESG ratings are wrong." The composite-shift reading is live, the KLD column rests on only 1,351 observations against 19,252 (Refinitiv) and 17,148 (Sustainalytics), and the authors' descriptive-evidence caveat governs here as everywhere else in Section 6.

What this does to the entry's conclusions

• The 09-01 defeat of the soft-metric objection stands. Nothing found today disturbs Table 8's central comparison or its sign.
• The axis of that defeat is corrected from activity-versus-result grammar to topic specificity. Any published use must be written on the corrected axis.
• The measurability confound is not weakened; it is sharpened and stays OPEN, and its cheap resolution route is now known to be a regression the authors did not publish rather than a study nobody has run.
• The transfer to AI is bounded more tightly than the 09-01 entry allowed. Carbon is the only domain in this paper with a physical-outcome test; every other category is tested solely against ratings and financial performance. "Result-defined metrics move real outcomes" is a one-domain finding — and the one domain is the one that happened to have Trucost.

Sources for this update — all primary, all from the same verified full text: Table 3 Panel A (p. 46), Table 8 (p. 53), Table 9 (pp. 54–55), Appendix A variable definitions (p. 37), Appendix B (p. 38), Appendix C (pp. 39–40) of madoc.bib.uni-mannheim.de/63834/1/ESGPay.pdf. Confidence: High.

Tags: paper-2, COUNTERARGUMENT Source: BMC Health Services Research (Miake-Lye, Delevan, Ganz, Mittman, Finley) URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC7014613/ Date published: 2020-02-11 Date flagged: 2026-08-30

Summary

Miake-Lye et al. systematically review and content-analyse organizational readiness-for-change (ORC) assessments in implementation science, covering 29 uses of readiness assessments drawn from 27 publications. They extract 1,370 individual survey items and map each to the Consolidated Framework for Implementation Research (CFIR). 897 items (68%) map to the CFIR inner setting domain, concentrated in readiness for implementation (n=220), networks and communication (n=207), implementation climate (n=204), structural characteristics (n=139), and culture (n=93).

Their conclusion is about operationalization, not validation: available instruments "predominantly focus on contextual factors within the organization and characteristics of individuals," and the item-level specificity they find suggests instruments must be tailored to the scenario in which they are fielded. On the state of the field they are explicit: "No gold standard exists within the realm of organizational readiness for change assessments," and "it remains unclear how best to operationalize readiness across varied projects or settings."

The review deliberately does not assess whether these instruments work. The authors state they "did not conduct a quality assessment of the included studies, since our analysis was not focused on the validity or robustness of study findings," and place the outcome link in future work: "Work testing the relationship between organizational readiness for change and implementation outcomes will help to better specify the underlying mechanisms of readiness."

Five Breakpoints Intersection

No direct Five Breakpoints mapping — this is a measurement source, not an organizational case. Logged because it establishes the field-level status of the construct on which Claim 5 (readiness as prerequisite) depends: after roughly two decades of instrument development, readiness measurement in its most mature discipline has no gold standard and no established link to implementation outcomes.

Four Forces Intersection

No direct Force implication — measurement-methodology entry. Note for later use: the two-factor structure that dominates this literature (commitment and efficacy, per Shea et al.'s ORIC) is a partial, health-sector-specific analogue of Commitment and Capability, with no counterpart to Purpose or Momentum. That non-correspondence is itself an argument, not a mapping.

Relevance to Brandon's Positioning

This is the source that turns the Unfalsifiability-risk challenge (knowledge/thesis-challenges.md, OPEN) from a structural critique with no evidence attached into a challenge with a citation behind it. Implementation science — a field with journals, funding, and a 20-year head start on readiness measurement — has produced dozens of instruments and, on its own account, no gold standard and no demonstrated link from readiness scores to implementation outcomes. That materially strengthens the objection: the difficulty is not that nobody has tried to make readiness predictive, it is that people have tried for two decades and have not closed it.

It cuts the other way too, and this is the more useful half. The absence is field-wide, not specific to the Five Breakpoints, so the critique "your framework is unfalsifiable" applies with equal force to every readiness construct in circulation, including the ones consultancies sell. And it prices asset #3 honestly: a prospective breakpoints instrument is not a weekend project that will settle the question — it is the thing an entire field has not managed, which is exactly why doing it first would be worth something. Stronger than anything in the base on this point; the base previously held no source at all on readiness-instrument validation.

Sources & Confidence

• Miake-Lye IM, Delevan DM, Ganz DA, Mittman BS, Finley EP. "Unpacking organizational readiness for change: an updated systematic review and content analysis of assessments." BMC Health Services Research 2020;20:106. doi:10.1186/s12913-020-4926-z — High (peer-reviewed systematic review; quotes verified against the PMC full text).
• Corroborating, cited within this entry: Shea CM, Jacobs SR, Esserman DA, Bruce K, Weiner BJ. "Organizational readiness for implementing change: a psychometric assessment of a new measure." Implementation Science 2014;9:7. doi:10.1186/1748-5908-9-7 — High. The ORIC authors themselves: "Although ORIC shows promise, further psychometric assessment is warranted. Specifically, the measure should be tested for convergent, discriminant, and predictive validity." (Verified against PMC3904699.)

Filed by / Routing

Stafford, 2026-08-30 (scheduled run, exploratory query slot). Audience: strategy. Routing: STAGE-CANDIDATE for research-evidence; evidence anchor for the Unfalsifiability-risk challenge and for asset #3 scoping.

Tags: paper-2, COUNTERARGUMENT Source: Implementation Research and Practice (SAGE) URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC12084713/ Date published: 2025-05-15 Date flagged: 2026-08-30

Summary

Caci and colleagues (University of Zurich and collaborators) systematically reviewed the healthcare literature on organizational readiness for change (ORC), covering 46 studies. The review is the most recent authoritative statement on whether the construct has been linked to what it is supposed to predict, and the answer is that it largely has not been tested.

The design breakdown is the finding. Of 46 studies, 40 measured ORC only once. Nineteen measured it before implementation only. Only three measured it at all three timepoints (before, during, after). The authors state: "scholars continue to measure ORC either retrospectively or at baseline only, without prospectively linking ORC to implementation outcomes," and "the limited number of studies linking ORC to implementation identified with this SLR confirms the previously critiqued shortage of studies examining ORC prospectively."

On the state of the claim itself: "claims about its importance often have not been based on the use of nuanced theories or empirical prospective designs clearly linking ORC to implementation outcomes."

Instrument note (required for structural/empirical instruments): this is a systematic review of studies, not a primary instrument. Its unit is 46 published studies of ORC in healthcare; its inference method is structured extraction and design classification; its observation window is the published literature to the 2024 search date. It measures the state of a literature, not the state of any organization.

Five Breakpoints Intersection

No direct Five Breakpoints mapping. Logged as measurement evidence bearing on Claim 5 and on the standing unfalsifiability challenge.

Four Forces Intersection

No direct Force implication — measurement-methodology entry.

Relevance to Brandon's Positioning

This is the source that makes the gap current rather than historical, and that distinction is load-bearing.

The readiness-measurement argument assembled on 2026-08-30 rests on sources from 2009, 2011, 2014 and 2020. The obvious rebuttal to a fifteen-year-old critique is that the field has moved on. Caci et al. closes that escape route with a review published fifteen months ago: as of 2025, 40 of 46 studies still measure readiness once, and the prospective link is still described as a critiqued shortage rather than a solved problem.

That converts the argument from "this was unresolved a decade ago" to "this is unresolved now," which is the version that can be published. It is also the version that prices asset 3b honestly: the design Brandon would be running is not a catch-up exercise, it is a design the specialist field has still not executed at scale, in a setting — AI deployment — where the outcome is commercially legible in a way that clinical implementation outcomes are not.

Caution against over-reading. A shortage of prospective studies is not evidence that readiness fails to predict. It is evidence that the question is open, which is exactly what Weiner et al. 2020 concluded and exactly what Brandon should say. The dishonest version of this material is "readiness measurement doesn't work"; the honest and more useful version is "nobody has run the test properly, and the tests that came closest split perception from implementation."

Sources & Confidence

• Caci L, Nyantakyi E, Blum K, Sonpar A, Schultes M-T, Albers B, Clack L. "Organizational readiness for change: A systematic review of the healthcare literature." Implementation Research and Practice. 2025;6:26334895251334536. doi:10.1177/26334895251334536 — High (peer-reviewed; PMC full text PMC12084713; quoted passages verified).

Filed by / Routing

Filed by: Stafford (scheduled run, 2026-08-30, 22:20 UTC — story-thread chase, not the exploratory slot) Routing: strategy — establishes the gap as current (2025), which is what makes it publishable.

Tags: paper-2 Source: Implementation Science (Helfrich, Blevins, Smith, Kelly, Hogan, Hagedorn, Dubbert, Sales) URL: https://pmc.ncbi.nlm.nih.gov/articles/PMC3157428/ Date published: 2011-08-12 Date flagged: 2026-08-30

Summary

Helfrich et al. set out a protocol to test whether a readiness instrument can prospectively predict implementation success. The instrument is the Organizational Readiness to Change Assessment (ORCA), built on the PARIHS framework: three primary scales (evidence, context, facilitation) and 19 subscales.

The design is the part worth having. ORCA is administered at baseline to 208 respondents across 53 facilities drawn from four partner projects testing evidence-based clinical practice changes; each partner project sets its own baseline timing so respondents complete the survey knowing what change is planned. Follow-up covers 158 respondents across 33 facilities, with implementation outcomes measured 6–9 months after baseline. The planned analysis is hierarchical linear modelling with the implementation outcome expressed as an effect size (Cohen's h) as the dependent variable and ORCA scales as independent variables, controlling for partner project and intervention receipt; with 30 anticipated sites they compute 90% power to detect R² ≥ 0.21.

The authors state the field's position on prior instruments directly: "few have undergone rigorous validation, notably to demonstrate the ability to prospectively distinguish successful change efforts from those that will fail."

Five Breakpoints Intersection

No direct Five Breakpoints mapping — this is an instrument-design source. Logged because it is the closest existing methodological template for the prospective breakpoints instrument in asset #3: baseline scoring before a known change, a 6–9 month outcome window, effect-size outcomes, and multilevel modelling with the site as the unit.

Four Forces Intersection

No direct Force implication — methodology entry, no organizational outcome data reported in the protocol.

Relevance to Brandon's Positioning

Two uses, both concrete. First, as a design template: asset #3 (the missing instrument) does not need to invent its method. Score organizations on the breakpoints before a known AI deployment, fix an outcome window at 6–12 months, express the outcome as an effect size, model at the site level. The will-it-hold rubric already does the scoring half; this supplies the measurement half, and the sample arithmetic is sobering in a useful way — a properly powered version of this design wanted ~30 sites, which sets the honest floor for what a Brandon-run cohort would need before it claims prediction rather than illustration.

Second, as a date stamp on the gap. The sentence "few have undergone rigorous validation, notably to demonstrate the ability to prospectively distinguish successful change efforts from those that will fail" was published in 2011. Miake-Lye et al. (2020) find no gold standard nine years later, and the ORIC authors flag predictive validity as untested in 2014. Three points on a fifteen-year line, all saying the same thing. That is the spine of the argument that the readiness-measurement gap is structural rather than a lane nobody noticed.

Caveat that must travel with this entry: this is a protocol, not results. Do not cite it as evidence that readiness predicts (or fails to predict) outcomes — cite it as a design and as the 2011 statement of the gap.

Update 2026-08-30 (22:20 UTC) — the results paper was chased and does not exist

The open follow-up flagged earlier today has been run to ground. No results paper from this protocol has ever been published. Searches conducted: Europe PMC on the full author string AUTH:"Helfrich CD" (91 records, 2006–2026 — after 2011 his output moves to burnout, PACT/medical home, de-implementation, cardiology and EHR transition, with no ORCA criterion- or predictive-validity results paper anywhere in it); the Europe PMC citation list for PMID 21777479 (29 citing articles, none a results paper, none by Helfrich); the Semantic Scholar citation graph for doi:10.1186/1748-5908-6-76 (~68 citing papers, no Helfrich-authored citing paper at all — authors do not cite their own protocol when the results paper does not exist); and targeted searches on Cohen's h, 53 facilities, hierarchical linear modelling and ORCA. The only later ORCA output from this group is conceptual: Helfrich et al., "Mapping the organizational readiness to change assessment to CFIR," Implement Sci Commun 2021;2:14, doi:10.1186/s43058-021-00121-0 — a mapping exercise, not a validation.

What makes the null worse than an absence. The instrument-development paper had already promised this exact work two years earlier. Helfrich CD, Li YF, Sharp ND, Sales AE, Implement Sci 2009;4:38 (PMID 19594942): "This analysis does not address the validity of the instrument as a predictor of evidence-based clinical practice" and "Criterion validation using implementation and quality-of-care outcomes is the next phase of our work." Promised in 2009, protocolled in 2011, never delivered. Seventeen years.

Confidence: high that no results paper is indexed in PubMed, Europe PMC or Semantic Scholar. Residual risk, stated honestly: results could exist in a non-indexed VA internal report, a dissertation, or a conference abstract. No trace of one was found, and this entry should say "not published" rather than "does not exist."

How this may and may not be used. It is legitimate to say the field's most rigorously designed prospective test of readiness was funded, protocolled and never reported. It is not legitimate to infer the study was run and found a null — file-drawer inference is exactly the reasoning this base refuses elsewhere, and there is no evidence the analysis was ever completed.

Sources & Confidence

• Helfrich CD, Blevins D, Smith JL, Kelly PA, Hogan TP, Hagedorn H, Dubbert PM, Sales AE. "Predicting implementation from organizational readiness for change: a study protocol." Implementation Science 2011;6:76. doi:10.1186/1748-5908-6-76 — High (peer-reviewed protocol; design details and quote verified against the PMC full text).
• Related, not separately filed: Helfrich CD et al. "Organizational readiness to change assessment (ORCA): development of an instrument based on the PARIHS framework." Implementation Science 2009;4:38 — High (instrument-development paper; not read in full).

Filed by / Routing

Stafford, 2026-08-30 (scheduled run, exploratory query slot). Audience: strategy. Routing: STAGE-CANDIDATE for research-evidence; method template for asset #3.

Tags: paper-2, COUNTERARGUMENT Source: American Journal of Public Health (APHA) URL: https://ajph.aphapublications.org/doi/10.2105/AJPH.2014.302140 Date published: 2014-09-01 Date flagged: 2026-08-30

Summary

Noe, Kaufman, Kaufmann, Brooks and Shore adapted the Organizational Readiness to Change Assessment (ORCA) — the PARIHS-based readiness instrument developed by Helfrich et al. — for a survey of 27 Department of Veterans Affairs facilities in the Western Region, fielded in 2011–2012. The study's purpose was programmatic rather than psychometric: the authors wanted to understand what shaped the delivery of culturally competent services to American Indian and Alaska Native veterans, a population with documented access disparities.

The result that matters for the readiness-measurement question is stated plainly in the abstract: "Several ORCA subscales (Program Needs, Leader's Practices, and Communication) statistically significantly predicted whether VA staff perceived that their facilities were meeting the needs of AI/AN veterans. However, none predicted greater implementation of native-specific services."

The 2025 systematic review by Caci et al. singles this study out in exactly those terms: "The authors found that no ORCA subscale predicted implementation of native-specific services."

Design limit, and it bounds the claim. This is a single cross-sectional survey administered in 2011–2012, not a prospective baseline-to-outcome design. The readiness measure and the outcome measure were collected in the same instrument at the same time. It is therefore a concurrent null, not a prospective one, and must never be described as a failed prediction over time.

Five Breakpoints Intersection

No direct Five Breakpoints mapping. Logged as measurement evidence bearing on the unfalsifiability challenge (stafford-research/knowledge/thesis-challenges.md) and on the Claim 5 evidence gap.

The temptation here is to tag Momentum Mirage on the grounds that the instrument predicted the appearance of performance while failing to predict the substance. That tag is deliberately withheld. The finding describes the behaviour of a survey instrument, not the behaviour of an organization undergoing transformation, and mapping it across that gap would be exactly the speculative tagging the operating rules forbid. The resemblance is real and is recorded under Positioning, where it belongs as an argument rather than as evidence.

Four Forces Intersection

No direct Force implication — measurement-methodology entry.

Relevance to Brandon's Positioning

This is the most damaging single data point in the readiness-measurement literature, and it is damaging in a way that is unexpectedly useful.

The standing unfalsifiability challenge says the Five Breakpoints framework risks being a taxonomy of autopsies unless readiness can be measured ex ante and shown to predict outcomes. The obvious defence has been that nobody else has built such an instrument either. Noe et al. sharpen that defence into something with more teeth and more danger at once: when a readiness instrument was pointed at both a perception outcome and an implementation outcome in the same organizations, it predicted the perception and not the implementation.

The resemblance to Momentum Mirage is worth stating in any piece Brandon writes on this, precisely as an observation and not as evidence: the measurement literature that would validate the framework's central prerequisite appears to reproduce, inside its own instruments, the failure the framework names — a measure that tracks how well people believe things are going and does not track whether anything happened. If that pattern held prospectively it would be the strongest available argument that readiness constructs measure organizational self-image. It has not been shown prospectively, and Brandon should not claim it has.

The correct use of this source is defensive and framing, not evidentiary. It says: the hole in readiness measurement is not that nobody tried, it is that the one time somebody checked both outcomes, the instrument split them. That is a reason to build the measure carefully with a hard outcome definition — asset 3b — not a reason to believe readiness does not matter.

Sources & Confidence

• Noe TD, Kaufman CE, Kaufmann LJ, Brooks E, Shore JH. "Providing culturally competent services for American Indian and Alaska Native veterans to reduce health care disparities." American Journal of Public Health. 2014;104(Suppl 4):S548–S554. doi:10.2105/AJPH.2014.302140. PMID 25100420 — High (peer-reviewed, APHA). Abstract verified; quoted sentence taken verbatim from the published abstract.
• Corroborating characterisation: Caci L, Nyantakyi E, Blum K, Sonpar A, Schultes M-T, Albers B, Clack L. "Organizational readiness for change: A systematic review of the healthcare literature." Implementation Research and Practice. 2025;6:26334895251334536. doi:10.1177/26334895251334536 — High.
• Confidence in the finding: high. Confidence in its evidentiary weight against prospective prediction: low — the design is cross-sectional and cannot bear a prospective claim.

Filed by / Routing

Filed by: Stafford (scheduled run, 2026-08-30, 22:20 UTC — story-thread chase, not the exploratory slot) Routing: strategy — feeds the unfalsifiability challenge and asset 3b's outcome-definition requirement.

Tags: paper-2, COUNTERARGUMENT Source: Implementation Research and Practice (SAGE) URL: https://journals.sagepub.com/doi/full/10.1177/2633489520933896 Date published: 2020-06-01 Date flagged: 2026-08-30

Summary

Weiner and colleagues systematically reviewed the measures used to assess organizational readiness for implementation in mental and behavioural health settings, rating each on psychometric and pragmatic properties. Nine readiness measures were identified. The review's central negative finding is stated directly: "Also striking is the lack of evidence for two psychometric properties of importance to implementation scientists: predictive validity and responsiveness." Fewer than half the measures had been tested for predictive validity at all, and "of these, few were tested for associations with implementation outcomes."

The verdict on the field's most-tested instrument is the sharpest passage. Of the TCU Organizational Readiness for Change measure: "Although the TCU-ORC has exhibited statistically significant associations with implementation outcomes in some studies, it has not in others and no discernible pattern emerges across multiple studies." And: "Despite 55 tests of association between individual TCU-ORC scales and various outcomes, the predictive validity rating of the measure remains 'minimal.'"

The review closes the question it opened: "Thus, the question of whether readiness for implementation matters remains open."

Who is writing this matters. Bryan Weiner is the developer of ORIC, the readiness instrument most widely used in implementation science, and the author of the field's most-cited theory of organizational readiness for change. This is the construct's principal architect reporting that the construct's practical claim is unestablished.

Scope limit: the review covers mental and behavioural health settings. ORCA is not within its scope.

Five Breakpoints Intersection

No direct Five Breakpoints mapping. Logged as measurement evidence against the load-bearing premise of Claim 5 — that organizational readiness is the prerequisite — and as the strongest available statement of the field-wide gap behind the standing unfalsifiability challenge.

Four Forces Intersection

No direct Force implication — measurement-methodology entry. Noted for later use: ORIC's two-factor structure (change commitment and change efficacy) remains a partial analogue of Commitment and Capability with no counterpart to Purpose or Momentum, and this review does not disturb that.

Relevance to Brandon's Positioning

This supersedes Miake-Lye et al. 2020 as the authoritative statement of the readiness-measurement gap, and it is a harder source to argue with. Miake-Lye said no gold standard exists. Weiner says something narrower and more damaging: measures exist, some have been tested extensively, and the tests do not produce a pattern. "No gold standard" invites the reply that the right instrument has not been built yet. "Fifty-five tests, minimal rating, no discernible pattern" does not.

For Brandon, the value is in the attribution. The single most effective sentence available on this material is that the developer of the field's most-used organizational readiness measure published a systematic review in 2020 concluding that whether readiness matters remains an open question. That sentence does three things at once: it establishes that the gap is field-wide rather than a weakness of the Five Breakpoints framework, it is unrebuttable by appeal to authority because the authority is the one saying it, and it sets the bar that asset 3b would have to clear to be worth doing.

It also raises the cost of over-claiming. If the framework's readiness rubric is ever pitched as predictive, this is the literature it will be measured against, and this literature has a documented record of instruments that did not predict.

Sources & Confidence

• Weiner BJ, Mettert KD, Dorsey CN, Nolen EA, Stanick C, Powell BJ, Lewis CC. "Measuring readiness for implementation: A systematic review of measures' psychometric and pragmatic properties." Implementation Research and Practice. 2020;1:2633489520933896. doi:10.1177/2633489520933896. PMID 37089124 — High (peer-reviewed; SAGE full text; quoted passages verified against the full text).

Filed by / Routing

Filed by: Stafford (scheduled run, 2026-08-30, 22:20 UTC — story-thread chase, not the exploratory slot) Routing: strategy — the citation of record for the readiness-measurement gap; supersedes Miake-Lye as the lead source.

Tags: paper-2, Technology Illusion, Capability Source: IBM Institute for Business Value URL: https://newsroom.ibm.com/2026-05-04-ibm-study-ceos-are-reshaping-c-suite-roles-for-the-ai-era Date published: 2026-05-04 Date flagged: 2026-08-29

Summary

IBM IBV's annual CEO study surveyed 2,000 CEOs and senior leaders across 33 geographies and 21 industries (fielded February–April 2026). Headline structural findings: 76% of organizations now have a Chief AI Officer (up from 26% in 2025), 79% of executives report decentralizing decision-making as AI takes a larger role, and CEOs expect 48% of operational decisions to be made by AI without human intervention by 2030 (vs. 25% today).

The study's core organizational claim: 83% of CEOs say AI success depends more on people's adoption than technology, yet only 25% of the workforce regularly uses AI — while 86% of leaders believe employees have the necessary skills. Organizations that redesigned five core business areas (technology, finance, HR, operations, cross-functional collaboration) are reported as four times more likely to have delivered on business objectives.

Qualifies under Category 7 (leader-defined instrument): it samples decision-makers about their own behavior and beliefs, which is exactly what makes the 86%-believe vs. 25%-use gap evidentiary — it measures the leaders' perception error directly, rather than relying on self-reported success.

Five Breakpoints Intersection

• Technology Illusion: 83% of CEOs conceding that AI success depends more on adoption than technology — while pouring investment into CAIO roles and tooling — is the C-suite's own admission that technology deployed onto unready organizational conditions doesn't deliver.
• Momentum Mirage (weak signal, untagged): the 76% CAIO adoption figure alongside only 25% workforce usage suggests structural announcements outrunning actual behavior change; not tagged because the study doesn't connect the two directly.

Four Forces Intersection

• Capability: the 86% of leaders believing employees have necessary AI skills against 25% regular usage is a direct measurement of capability mis-assessment at the top — the force is implicated through leaders' inability to gauge it.

Relevance to Brandon's Positioning

Re-find note (2026-08-30): already in the v1 base since 2026-05-05 as the founding member of "The 86% Belief Gap" pattern and a member of "The Measurement Mismatch" (knowledge/emerging-patterns.md); this entry is the transferred repo's physical record of the source. The marginal contribution below stands.

Support for paper-2 Claim 5 (readiness as prerequisite): the 4x delivery multiplier attaches to organizational redesign, not to AI capability. Caveat for the case file: "delivered on business objectives" is self-reported by the same leaders — this is a leader-defined instrument, not an outcome-defined one, so it supports Claim 5 at moderate strength despite the High-credibility publisher. The 86-vs-25 gap belongs to "The 86% Belief Gap" pattern (see knowledge/emerging-patterns.md), relevant to Claim 4.

Sources & Confidence

• IBM Institute for Business Value, 2026 CEO Study (n=2,000) — High (publisher), with the noted self-report limitation on the outcome multiplier.

Filed by / Routing

Stafford v2, 2026-08-29. Audience: strategy.

Tags: paper-2, practitioner, Strategic Disconnection, Technology Illusion Source: Okta (commissioned survey), surfaced via Cloud Security Alliance blog URL: https://www.okta.com/newsroom/articles/ai-agents-at-work-2026-agentic-enterprise-security/ Date published: 2026-05-30 Date flagged: 2026-08-29

Summary

Okta-commissioned survey of 292 executives (CEOs, CIOs, CTOs, VPs) and 492 knowledge workers on AI and AI-agent usage. The design — asking executives and workers the same questions — produces paired perception measurements: 90% of executives express confidence in their organization's visibility into AI tools, while 52% of employees admit using unapproved AI tools; 95% of executives believe employees use AI responsibly within guidelines, while 54% of unapproved-tool users shared internal messages, 45% disclosed sensitive HR information, and 39% uploaded confidential documents; 65% of executives call their AI policies "very clear," while 57% of knowledge workers say policies are unclear, hard to find, or nonexistent.

On agent governance specifically: 88% of organizations report suspected or confirmed AI-agent security incidents, yet only 22% treat AI agents as independent identity-bearing entities and only 34% apply security controls to agents equivalent to those for human employees.

Five Breakpoints Intersection

• Strategic Disconnection: 65% of executives rating AI policy "very clear" against 57% of workers finding it unclear or nonexistent is a paired measurement of the gap between stated direction and operational reality — the illusion of alignment, quantified within the same organizations.
• Technology Illusion: 88% reporting agent incidents while only 34% apply human-equivalent controls shows agentic technology deployed on top of governance conditions everyone can see are broken.

Four Forces Intersection

No direct Force implication — breakpoints-only entry.

Relevance to Brandon's Positioning

Re-find note (2026-08-30): the v1 base ingested this study on 2026-05-30 (okta-ai-agents-governance-gap-may2026.md, not transferred) for the 92%/22% agent-identity-coverage gap, supporting Claims 3–4. This entry's marginal contribution is the paired exec/worker perception items (90/52 visibility, 65/57 policy clarity, 95% responsible-use belief vs. admitted leakage), logged as a candidate member of the 86% Belief Gap pattern in knowledge/emerging-patterns.md — flagged for Friday-synthesis review, not counted unilaterally. Weaknesses to respect: vendor-commissioned research with a product to sell (identity governance), modest samples, and self-report on the worker side too. Use it as convergent corroboration, never as the load-bearing citation.

Sources & Confidence

• Okta, "AI Agents at Work 2026" (n=292 executives, 492 knowledge workers) — Moderate (vendor-commissioned; directional value from the paired-perception design).
• Cloud Security Alliance blog (2026-07-17, Harish Peri, Okta SVP) — Moderate; surfacing route, author is the vendor's own executive.

Filed by / Routing

Stafford v2, 2026-08-29. Audience: governance / strategy.

Tags: paper-2, Momentum Mirage Source: Science (Vol. 390, Issue 6779), via ScienceDaily summary URL: https://www.sciencedaily.com/releases/2025/12/251224032347.htm Date published: 2025-12-18 Date flagged: 2026-08-29

Summary

Kusumegi, Yang, Ginsparg, de Vaan, and Stuart (Cornell / UC Berkeley) analyzed more than 2 million papers posted January 2018–June 2024 across arXiv, bioRxiv, and SSRN, using a detection model trained on pre-2023 human-written text to identify likely LLM usage, then tracked publication rates and journal acceptance before and after adoption.

Output effects are large: LLM users posted roughly one-third more papers on arXiv, with increases exceeding 50% on bioRxiv and SSRN. The quality signal inverted: in human-written papers, high writing complexity correlated with journal acceptance; for AI-flagged papers, even those scoring high on writing complexity were less likely to be accepted. The surface features that historically signaled quality no longer discriminate once AI produces them cheaply.

Five Breakpoints Intersection

• Momentum Mirage: the inversion of the complexity–acceptance relationship for AI-flagged papers is direct measurement of appearance decoupling from substance — output volume rose 33–50% while the traditional quality signal stopped predicting real acceptance.

Four Forces Intersection

No direct Force implication — breakpoints-only entry.

Relevance to Brandon's Positioning

Best evidence yet in the base for paper-2 Claim 4 (failure is harder to detect because AI produces the appearance of execution) — and it's outcome-defined in the Category 7 sense: the comparison group is defined by a measured outcome (journal acceptance), not self-report. The scientific-publishing setting is a limitation (knowledge work with an unusually clean external quality gate); the claim needs a corporate analogue where no such gate exists — which is precisely where Claim 4 predicts the danger is worst. Argue it that way: where science has journals to catch the mirage, firms have nothing.

Sources & Confidence

• Kusumegi et al., Science 390(6779), 2025-12-18, NSF-funded — High.
• ScienceDaily summary (read source) — Moderate as a rendering; verify exact figures against the journal article before quoting in a draft.

Filed by / Routing

Stafford v2, 2026-08-29. Audience: strategy.

Tags: paper-2 Source: SSRN working paper (Guohou Shan, Feng Zhu) URL: https://ssrn.com/abstract=6456498 Date published: 2026-03-22 Date flagged: 2026-08-29

Summary

Shan and Zhu study whether AI exposure changes the two fundamental dimensions of organizational design — hierarchy depth and managerial span of control — exploiting staggered AI exposure across 7,681 U.S. firms between 1990 and 2020 in a difference-in-differences design. Finding: firms exposed to AI reduce the number of hierarchical ranks while simultaneously increasing managerial spans of control — systematic organizational flattening.

Read via abstract and secondary coverage only (SSRN full text returned 403 at ingest); methodology details beyond the design and sample are unverified. Flag for a full-text read before citing in a paper draft.

Five Breakpoints Intersection

No direct Five Breakpoints mapping. Logged as core paper-2 mechanism evidence: large-sample causal-design support for hierarchy responding to information-processing cost changes (Claims 1–2), which is the premise the breakpoints argument builds on.

Four Forces Intersection

No direct Force implication — mechanism-evidence entry. (Abstract-only read; no Forces tags per hard rule 2.)

Relevance to Brandon's Positioning

Corrected 2026-08-30 after the v1 import. The day-0 claim that this filled an empty Claim 1 slot is void: the v1 base already holds Babina et al. (NBER 31325), Ewens & Giroud (NBER 34162), Wang/Feng/Sun (arXiv 2511.02099) and Alekseeva et al. (SMJ 2026) on this ground — and its ESCALATED "Résumé-Panel Consensus" pattern requires recording any structural instrument's data source, inference method and window BEFORE its finding. Here only the window is known (1990–2020, entirely pre-generative); the data source is unestablished (403 at ingest). Consequences: (1) this paper is QUARANTINED as evidence until a full-text read establishes its data genre — if résumé/postings-based, it is a fifth member of one method-family, not independent confirmation; (2) it must never be cited alongside Babina et al. or Ewens & Giroud as if independent until that is settled; (3) if it turns out to use administrative or payroll data, it would be the pattern's second break condition met and worth far more than a confirmation. The Claims 1–2 framing (routing dissolves, alignment unmeasured) survives either way, but as an argument, not yet as this paper's evidence.

Sources & Confidence

• Shan & Zhu, "AI Exposure and Organizational Structure," SSRN 6456498, posted 2026-03-22 — Moderate, quarantined (abstract-only; data genre unknown; see corrected Relevance section — confidence revisited after full-text read).

Filed by / Routing

Stafford v2, 2026-08-29. Audience: strategy.

Title: The Five Breakpoints Leaders Miss Status: Published 2026-04-28 — https://www.workthatholds.com/p/the-five-breakpoints-leaders-miss Tags: practitioner, foundation Text note (2026-08-30): the article text below is the pre-publication draft; the published version carries minor wording differences (e.g. "were not designed" for "were never designed") and the shorter title. For direct quotation in anything public, quote from the published URL; for framework definitions and tagging, this file remains the working reference.

This is Brandon's primary intellectual work. Everything Stafford does is grounded in this framework. Read this file before analyzing any content for Five Breakpoints intersections.

Full Article Text

The Five Breakpoints Leaders Miss in Transformation How leaders can design against predictable failure patterns and recognize them early

Organizations rarely fail to transform because they lack activity. More often, they fail because they were never designed to withstand the predictable pressures that large-scale change creates. The roadmap exists. Governance is in place. Funding is approved. Milestones are green. Yet leaders can still sense that the business is not moving the way it should. Teams are active without converging. Pilots are underway without scaling. Progress is being reported, but confidence is not increasing.

One executive described it to me this way: "It feels like youth soccer where everyone is kicking at the ball at the same time, but the ball still doesn't move."

That pattern shows up across cloud modernization, AI adoption, operating model redesign, customer transformation, and broader business change. The surface details vary. The underlying failure patterns do not.

Across large-scale change efforts, the same five breakpoints tend to appear again and again. They are not random setbacks. They are predictable failure patterns that organizations often fail to design against, then fail to recognize early enough once they begin to surface.

Leaders do not need another slogan about change. They need a way to design against these breakpoints from the outset and identify the warning signs before drift becomes expensive.

1. Strategic Disconnection

The first breakpoint appears when a transformation is launched with broad intent but without enough precision to keep the organization aligned under real operating pressure. Leaders believe the strategy is clear, but the organization is operating from multiple versions of the outcome.

A healthcare organization launched a cloud transformation with visible executive support and strong early agreement. But each major stakeholder quietly interpreted the effort through the lens of their own function. Security processes would remain intact. Change windows would remain intact. Review paths would remain intact. No one openly resisted. No one had actually committed to the same destination. A year later, the organization had new cloud platforms and an essentially unchanged operating model.

This is one of the most expensive failure patterns because it does not look like conflict at the start. It looks like consensus. Leaders hear the same language repeated back to them and assume alignment exists. In practice, teams fill in the blanks with their own definitions, priorities, and assumptions about what will or will not change.

The issue is not inspiration. It is precision. A goal like "become more digital" or "use AI to improve customer experience" sounds aligned in a kickoff meeting and fragments in execution. A goal like "reduce customer resolution time by 40 percent within two quarters" gives teams a shared outcome they can translate into decisions, tradeoffs, and accountability.

When strategic disconnection sets in, organizations do not usually stop moving. They simply begin moving in slightly different directions.

Self-check: If you asked ten leaders to describe the primary outcome of the transformation, would their answers actually match?

2. Incentive Fragmentation

The second breakpoint appears when a transformation depends on cross-functional cooperation but was never designed to align the incentives of the people whose decisions determine whether it moves. What looks like support early on often collapses once real tradeoffs appear.

A financial institution had a major migration effort underway with one deadline no one could ignore: datacenter leases were expiring. The VP of Data had aligned teams, built the business case, and set the timeline. The CISO attended the planning meetings and raised no visible objection. Months later, he revealed that he had engaged a separate consulting partner, defined a different set of security requirements, and that nothing would move until his scorecard was satisfied.

The cloud was not the problem. The program stalled because one executive with veto power had little reason to optimize for migration success. His job was to ensure security was never the cause of failure. Migration speed was not his metric.

This is where many leaders misread commitment. They interpret attendance as buy-in. It often means awareness. Sometimes it means observation. In this case, it was reconnaissance.

Incentive fragmentation is especially dangerous because teams often look productive while the system is quietly pulling them apart. One leader is measured on speed. Another on cost containment. Another on risk reduction. Another on quarterly output. Everyone works hard. The enterprise does not move coherently.

The issue is not whether leaders verbally support the change. It is whether the system makes it rational for them to prioritize it when tradeoffs appear. If a transformation succeeds while a powerful stakeholder's metrics stay flat or worsen, leaders should not be surprised when resistance shows up in delay, exceptions, side processes, or alternative standards. Those responses are not random. They are the natural consequence of a system that was never aligned to move together.

Self-check: If the transformation succeeds but a leader's metrics do not improve, will that leader still treat the work as a priority?

3. Process Friction

The third breakpoint appears when leaders set a faster ambition without redesigning the operating model required to support it. The strategy changes. The machinery does not.

A retail organization invested in cloud capabilities that could technically provision a working application environment in hours. But launching an application still required sequential handoffs across operating system, network, storage, identity, database, application, backup, monitoring, and security teams. Each group had its own queue. Each worked on its own timeline. No one owned the full journey from request to result. The cloud could move in hours. The organization still moved in weeks.

This is where many transformation efforts lose credibility with the people doing the work. Leaders announce a new ambition, but the underlying machinery remains unchanged. The handoffs are the same. The approvals are the same. The decision rights are the same. The dependencies are the same. Teams are told to move faster inside a system designed to prevent speed.

What makes this breakpoint so persistent is that organizations often overestimate capability by looking at talent rather than flow. They have skilled people. They have good technology. They may even have committed leaders. What they do not have is a delivery system that allows those ingredients to produce consistent business movement.

When process friction dominates, the strategy is not actually being executed. It is being negotiated one handoff at a time.

Self-check: If you mapped how work actually flows today, would it look like the formal process or the workaround your teams built to survive it?

4. The Technology Illusion

The fourth breakpoint appears when leaders invest in a new capability without designing the surrounding behaviors, workflows, and decision norms required to make it valuable.

An enterprise software company invested in a new sales analytics platform. The data was better. The dashboards were better. The engineering work was solid. Months later, teams were still relying on legacy reports and spreadsheets. Not because the new system was inaccurate or harder to use. Because the old process gave people more room to tune the story, soften the numbers, or avoid difficult conversations. The technology was ready. The organization was not.

This pattern is common in AI, analytics, workflow automation, and broader modernization programs. Leaders invest in the visible artifact and underestimate the operational and behavioral change required to make it valuable. They hope the tool will force the organization forward. More often, the organization absorbs the tool into its existing habits and continues operating much as before.

Technology can accelerate aligned execution. It cannot create alignment on its own. That is why so many well-funded initiatives look compelling in demonstrations and disappoint in practice. The technical capability becomes visible before the operating discipline required to use it.

This is not a technology failure. It is a leadership design failure. The organization was asked to use a new tool inside an old system.

Self-check: If you removed the technology tomorrow, would teams still agree on how the work should happen?

5. Momentum Mirage

The fifth breakpoint appears when a transformation is launched with early visibility but without enough built-in reinforcement to sustain progress once leadership attention shifts.

A semiconductor company launched a digital transformation with strong executive attention and promising early wins. The first quarterly review looked strong. Pilots were working. Teams were engaged. Progress was visible. Then margin pressure pulled the executive sponsor into other priorities. Governance meetings remained on the calendar. Status reports continued. But decision velocity slowed, obstacles stayed unresolved longer, and visible reinforcement faded. No one cancelled the initiative. No one needed to. The organization simply stopped feeding it. By the next planning cycle, the transformation was still alive in presentations and largely dead in practice.

This breakpoint is easy to miss because activity continues after momentum has weakened. Meetings happen. Slides update. The work still has a name. What disappears is the force that turns progress into confidence and confidence into adoption. People begin reading silence as a signal. They return to local priorities. The old system starts reclaiming ground.

Momentum is not enthusiasm. It is the organization's ability to convert intent into visible progress, then reinforce that progress fast enough to sustain belief.

Self-check: What would happen if you stopped actively managing this transformation for 90 days?

The Four Underlying Forces

The five breakpoints are visible signs that the transformation was not designed strongly enough in one or more of four underlying forces:

• Purpose — Is the outcome clear enough that teams would define success the same way?
• Commitment — Have priorities, incentives, ownership, and resources been aligned strongly enough to hold when tradeoffs appear?
• Capability — Does the organization have the operating ability to execute and scale?
• Momentum — Is there enough visible progress, reinforcement, and feedback to sustain movement after launch?

Breakpoint → Force mapping:

• Strategic Disconnection → weak Purpose
• Incentive Fragmentation → weak Commitment
• Process Friction → weak Capability (flow)
• Technology Illusion → weak Capability (readiness)
• Momentum Mirage → weak Momentum

The Diagnostic Questions

Purpose: Are we aligned on the outcome, and would teams describe success the same way? Commitment: Have we aligned incentives, ownership, and priorities strongly enough to survive conflict? Capability: Does our operating model allow the work to move at the speed the strategy now requires? Momentum: Have we built enough visible progress and short enough feedback loops to sustain confidence and action after launch?