Thesis challenges

• Logged: 2026-08-29 · Updated 2026-08-30 (v1 import attached the evidence) · Targets: thesis
• The objection, at full strength: The measurable gains from AI in organizations so far come overwhelmingly from coordination and information-processing improvements; "readiness" and alignment constructs are post-hoc explanations that don't predict outcomes better than plain execution capacity. Coordination primacy is now a small formal literature, not a vibe: Farach, "AI as Coordination-Compressing Capital" (arXiv 2602.16078) predicts spans expand, management share falls and wage gaps widen as AI compresses coordination cost; Klein & Wieczorek, "The Headless Firm" (arXiv 2602.21401) derive an hourglass equilibrium as integration cost falls from O(n²) to O(n).
• Evidence for the objection: Ewens & Giroud (NBER WP 34162) — firms flatten after adopting AI (all four coefficients negative), a directional confirmation of Farach. Carried caveats from the v1 thread log: significance at the 10% level only, no span-of-control or wage-premium measurement, résumé-inferred structure, pre-generative window.
• Evidence against / limiting the objection: the flattening result is compatible with the Fahn/Li/Sun bad-job equilibrium (structure thins and discretion is designed out), so confirmed flattening does not by itself defeat the thesis; the Conditional-Productivity pattern (Wang/Feng/Sun + BIS + BCG) puts organizational variables, not coordination tooling, on the right-hand side of measured productivity outcomes; Klein & Wieczorek's own model relocates the constraint to verification, which is an organizational property.
• What it would take for this objection to win: firm-level measurement showing spans/management-share moving as the compression models predict AND transformation outcomes tracking coordination-tooling quality better than any readiness measure (Category 7 standard).
• Current best response: coordination compression is real and the thesis should absorb it rather than deny it — the fork is whether what remains binding after compression is organizational (verification, incentives, purpose) or residual coordination. The live tracker is the Regime Fork thread (knowledge/story-threads.md), next check 2026-11-09.

Update 2026-09-05 (Stafford) — a third formal model arrived, it looked like it took the base's answer away, and the paper's own algebra says otherwise. Logged including the reading that failed.

The reading I opened with, at full strength. This challenge's current best response leans on Klein & Wieczorek: coordination compression is real, and what remains binding afterwards is verification, which is an organizational property. Bartolucci & Vivo, Queue & AI (arXiv 2605.27202, 26 May 2026) — ingested by the 09-05 scan as a friendly mechanism — appears to take that answer away. They derive verification collapse from a queue: arrival rate, service-time second moment, reviewer capacity. If oversight failure is fully derivable from load, the residual after coordination compression is capacity, not organizational design. That would make this the third formal model in a row — after Farach (coordination-compressing capital) and Klein & Wieczorek (the headless firm) — explaining this base's observed patterns with no organizational variable on the right-hand side, all three of which the base has ingested as confirmations.

Why the reading fails, and the refutation is the paper's Eq. 8 read in the body rather than the abstract. The review threshold is π⋆(θ) = θ/(κK). Verbatim: "the review threshold rises with the congestion cost of reviewer time, θ, and falls with both the cost of an escaped error, K, and the effectiveness of review, κ." And on κ: "The parameter κ > 0 represents the reviewer's verification skill: a higher κ means each hour of review catches a larger fraction of errors." So the right-hand side is not load. It is load relative to how skilled the reviewer is (κ — Capability) and how much the organization loses when the check fails (K — accountability architecture, i.e. Incentive Fragmentation). Only the arrival rate λ and the AI's service-time variance c²_A are genuinely exogenous. The rival reading requires κ and K to be constants of nature; they are line items, and the authors treat them as fixed parameters without argument.

Net effect: the challenge stays CONTESTED and got weaker today, on the strongest ground it has been contested on. A rival model, unpacked, contains the framework's variables in its own denominator — which is a better outcome for the thesis than another confirming survey, because it is a hostile-witness result.

Two disciplines, both binding. (1) The paper has no data — its "Empirical calibration" section is an agenda ("a useful empirical agenda would therefore combine time stamps, reviewer effort, error flags, and rework links at the task level"), and its figure parameters are described in the text as "a representative calibration" and "illustrative levels." A model that makes the framework's claim precise is not evidence that the claim is true, and no term in θ/(κK) has been estimated anywhere, by anyone. (2) The genre exposure survives the individual refutation and should be watched: three formal models, three non-organizational left-hand sides, three ingested as friendly. Each has come out the framework's way on inspection. That is a run, not a law, and the fourth should be read parameter-by-parameter before it is booked as support.

• Logged: 2026-08-29 · Targets: thesis, Claim 5

• The objection, at full strength: "Readiness is the prerequisite" risks explaining every failure after the fact (they failed, therefore they weren't ready). Without an ex-ante, measurable definition of readiness that predicts outcomes, the Five Breakpoints framework is a taxonomy of autopsies, not a theory.

• Evidence for the objection (attached 2026-08-30, scheduled run): implementation science has been building organizational readiness instruments for two decades and has not established that any of them predicts anything. Helfrich et al. 2011 (knowledge-base/implementation-science-helfrich-orca-predictive-validity-2011-08.md): "few have undergone rigorous validation, notably to demonstrate the ability to prospectively distinguish successful change efforts from those that will fail." Shea et al. 2014, the authors of ORIC — the field's most-used readiness measure — on their own instrument: "the measure should be tested for convergent, discriminant, and predictive validity." Miake-Lye et al. 2020, systematic review of 29 assessment uses across 27 publications (knowledge-base/bmc-hsr-miake-lye-readiness-assessments-review-2020-02.md): "No gold standard exists within the realm of organizational readiness for change assessments," with the readiness→outcome link placed in future work. Three points on a fifteen-year line, same finding each time.

• What it would take for this objection to win: continued absence of any prospective instrument built on the framework's constructs. Sharpened 2026-08-30: the objection's real force is no longer "you haven't built one" — it is "the discipline that specializes in this has not built one either, which suggests the readiness→outcome link may not be measurable at useful effect sizes." Answering it now requires a prospective instrument that predicts, not merely one that exists.

• Current best response (updated 2026-08-30 after v1 import): still OPEN, but the base holds raw material for an answer — the Conditional-Productivity pattern (organizational context as a measured moderator in firm-level production functions), Zeng Ming's "60-point baseline" as a candidate readiness threshold, and the Measurement-Mismatch diagnostic (measure what employees optimize for, not tool access). None of these is yet a prospective Five Breakpoints instrument; building one remains growth edge #2 in brandon-development.md.

• Update 2026-08-30 (evening): the scan's first logged exploratory query established that NOBODY has measured pre-deployment organizational condition against post-deployment AI outcome — the field asserts the proposition and has no instrument. The challenge is therefore symmetric: the critics have no disconfirming instrument either. First mover on the instrument settles it; see asset-suggestions.md entry 3.

• Update 2026-08-30 (scheduled run, 22:20 UTC) — the thread chase came back, and the challenge is now the sharpest OPEN item in this file. Read this update before the 21:55 one below it; it supersedes that update's evidence base without contradicting its conclusion.

  Three things were established against primary sources.

  (1) The definitive test was designed, funded, protocolled — and never reported. Helfrich's 2011 prospective ORCA study (53 VA facilities, baseline-to-6–9-month outcome as Cohen's h, hierarchical linear model, 90% power at R² ≥ 0.21) has no published results paper. Europe PMC's full author list for Helfrich CD (91 records to 2026), the 29 articles citing the protocol, and ~68 Semantic Scholar citations contain no results paper and no Helfrich-authored citing paper at all. The 2009 development paper had already deferred it: "Criterion validation using implementation and quality-of-care outcomes is the next phase of our work." Promised 2009, protocolled 2011, never delivered. File: knowledge-base/implementation-science-helfrich-orca-predictive-validity-2011-08.md (Update section). Do not infer a file-drawer null — there is no evidence the analysis was completed.

  (2) The construct's own architect says the question is open. Weiner BJ et al., Implementation Research and Practice 2020;1 (doi:10.1177/2633489520933896) — Weiner is the developer of ORIC and the author of the field's most-cited theory of organizational readiness: "Also striking is the lack of evidence for two psychometric properties of importance to implementation scientists: predictive validity and responsiveness"; on the most-tested instrument, "Despite 55 tests of association between individual TCU-ORC scales and various outcomes, the predictive validity rating of the measure remains 'minimal'"; and the conclusion, "the question of whether readiness for implementation matters remains open." File: knowledge-base/weiner-readiness-measures-psychometric-pragmatic-review-2020-06.md. This supersedes Miake-Lye 2020 as the lead citation — "no gold standard" invites the reply that the right instrument is coming; "55 tests, no discernible pattern" does not.

  (3) The gap is current, not historical. Caci L et al., Implementation Research and Practice 2025;6 (doi:10.1177/26334895251334536), published 2025-05-15: of 46 studies, 40 measured readiness only once; "scholars continue to measure ORC either retrospectively or at baseline only, without prospectively linking ORC to implementation outcomes." File: knowledge-base/caci-organizational-readiness-change-healthcare-review-2025-05.md. This closes the "the field has moved on since 2011" escape route.

  And one finding that cuts differently from all of the above — the most interesting thing in tonight's yield. Noe TD et al., American Journal of Public Health 2014;104(Suppl 4):S548–S554 (doi:10.2105/AJPH.2014.302140) adapted ORCA across 27 VA facilities and reported: *"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 instrument tracked the belief that things were going well and did not track whether anything was implemented. File: knowledge-base/noe-orca-cultural-competence-va-predictive-null-2014-09.md. Two disciplines of interpretation are mandatory here. The design is a single cross-sectional survey, so this is a concurrent null and cannot be described as a failed prediction over time. And the resemblance to Momentum Mirage — a measure of the appearance of progress that does not track the substance — is an argument, not a mapping; the KB entry deliberately withholds the breakpoint tag, and any published use must do the same.

  Net effect on the challenge: it stays OPEN and gets harder, and the symmetry defence gets stronger. Harder, because the objection is no longer "Brandon hasn't built a predictive readiness instrument" but "a specialist discipline has spent two decades on this, its own architects report the question open, its definitive test was never published, and the closest thing to a result split perception from implementation." Stronger, because every one of those facts applies to every critic deploying the objection. The response plan is unchanged and now better priced: asset 3b, with a hard outcome definition fixed before scoring begins, or the challenge stays open honestly. What must not happen is asset 3a being described in 3b's language — see the 21:55 update below, which stands.

• Update 2026-08-30 (scheduled run, 21:55 UTC): the symmetry argument above survives but gets more expensive. Stafford's exploratory query widened the search out of AI and into implementation science, where readiness measurement is a mature 20-year discipline — and found the same hole (three sources above). Two consequences. (1) The symmetry defence is stronger than it looked: no readiness construct anywhere has cleared this bar, so the critique is a critique of a whole measurement class, not of Brandon's framework in particular, and anyone deploying it against him is standing on the same ground. (2) The offensive move gets harder and more valuable at once. Helfrich's design wanted ~30 sites for 90% power at R² ≥ 0.21; a dozen scored cases is an illustration, not a prediction, and should be described as such. Do not let asset #3 be pitched as "we'll settle this" until the sample supports it — over-claiming a prospective instrument that then fails to predict is the one outcome that would convert this OPEN challenge into a DEFEAT of the thesis rather than of the objection.

• Update 2026-09-03 (Stafford, on the Alibaba field experiment the scan ingested 09-03). The challenge stays OPEN and its core is untouched — no prospective readiness instrument that predicts was built — but one limb of it, the "taxonomy of autopsies, not a theory" charge, is now weaker than it was 48 hours ago, and the honest ledger should say so. Yesterday's scan brief named, ex ante, the observation that would count against the framework's prerequisite claim: a function with instrumented output that deployed a tool, changed nothing else, and captured an externally measured result. A randomized case matching that specification exists (Wang, Zhu, Feng, Lu & Jia, arXiv 2605.14830; research-evidence KB), and it came back the framework's way — unchanged role + agentic AI → the externally measured customer rating fell −0.412 on the chats the AI touched. Two disciplines keep this from being oversold. (1) This falsifies-and-survives the directional prerequisite/mechanism claim (deploying onto an unchanged role degrades the touched outcome), which is the "autopsy" charge; it does not touch the predictive-instrument half (an ex-ante readiness score that forecasts outcomes), which is the charge's stronger form and remains entirely unmet. (2) The experiment observes a behaviour under an unchanged design, not a scored organization that later failed — so it is a mechanism confirmation, not a readiness prediction. Net: the objection's weaker limb (unfalsifiable-in-principle) is answered — the framework specified a disconfirmer and searched for it; the stronger limb (no instrument predicts) is not. Split the two in any published response and do not let the mechanism win be described as the instrument win.

---

Resolved in one day, by the paper that raised it, and the resolution reverses the sign. The entry below is preserved verbatim as it was logged, because a challenge ledger that quietly edits its own losing entries is worthless. Read this resolution block first.

• Resolved: 2026-09-01 · Targets: Momentum Mirage as a diagnostic; the "Paid for Deployment" pattern; Paper 3 mechanism 3
• How it was resolved. The 08-31 entry rested entirely on Cohen et al.'s abstract, because Wiley, SSRN, Taylor & Francis and the ECGI PDF were all inaccessible, and it named the decisive unknown precisely: the outcome variable. The full text was obtained on 2026-09-01 from the Universität Mannheim institutional repository (madoc.bib.uni-mannheim.de/63834/1/ESGPay.pdf, 60pp., text-extractable). The abstract aggregates three separate results against three different dependent variables, and the paper's own Table 8 decomposes them the other way.
• The evidence that defeated it. Table 8 (N = 21,715 firm-years, 4,395 firms, 21 countries, 2011–2020, firm and year fixed effects, SEs clustered at firm) regresses ∆CO2 — Scope 1 emissions in tons, measured by Trucost, an externally measured physical outcome and not a rating — on ESG Pay. Column (1), the indicator for incorporating any ESG metric: −0.07, t = −0.85, not significant. Column (2), decomposed into ten metric categories: only Carbon emissions is significant, at −0.77, t = −2.88, p < 0.01. The authors state it 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." Table 9 supplies the other half: against ∆ESG Rating, the generic ESG Pay indicator is +0.233 (10%) on Sustainalytics and +1.004 (5%) on KLD — and a precise −0.001 (t = −0.17) on Refinitiv.
• Why this is a defeat and not merely a wash. The objection's empirical premise was that a soft, aggregate, result-agnostic metric moves the outcome it names. On the one comparison that isolates metric grammar — same firms, same contract vehicle, same specification, same dependent variable, varying only whether the incentive names a quantity a third party counts — the soft metric is a null and the result-defined metric works. That is the framework's inference, tested and confirmed, in a top-three accounting journal. The base's largest outstanding threat to the Momentum Mirage mapping is now its first piece of empirical support.
• The disciplines that must travel with the win, none of them optional. (1) The load-bearing comparison is the narrow one — ESG Pay versus Carbon emissions, both against ∆CO2. Nine of the ten categories are tested against an outcome most were never meant to move; a diversity metric failing to cut emissions is a mismatched dependent variable, not a failed incentive. Stating the wide version invites a correct demolition. (2) The authors call the evidence "mainly descriptive, which cautions against making strong causal claims" and say plainly that "the results in Tables 8-10 are not statistically strong." Both phrases travel. (3) The window is 2011–2020, entirely pre-generative. This is an analogue for AI metrics, not evidence about them. (4) The authors do not endorse this reading. They read Tables 8–9 as supporting a "strengthened pledge" account and as hard to reconcile with window-dressing. The decomposition is theirs; the contrast drawn from it is Brandon's. Say so. (5) Do not run the window-dressing story on Refinitiv. Refinitiv is the self-reported-disclosure-based score and it is exactly the one that did not move. The rating results are vendor-divergent and Table 9 will not carry a clean "generic metrics buy disclosure scores" claim.
• Correction appended 2026-09-02 — the defeat stands, the axis was misnamed. Discipline (1) above correctly identified the narrow comparison (ESG Pay versus Carbon emissions, both against ∆CO2) as the only load-bearing one. It then described what that comparison isolates as metric grammar — activity-defined versus result-defined. Appendix B, read today, shows that is wrong: grammar varies inside Cohen et al.'s categories rather than between them (carbon "kg CO2e/tonne", safety "DART incident rate per 100 full-time employees", diversity "percentage of women among the SMP" are all result-worded; compliance "continue to assess human rights..." and governance "establish standalone corporate governance and risk procedures... that build trust" are activity-worded). The decomposition is by subject matter, so the −0.07-versus-−0.77 contrast isolates topic specificity — whether the contract names the outcome you are measuring. The result, the sign and the defeat are unaffected; the sentence Brandon may publish is. Write it as "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 Momentum Mirage reading of it survives — but it must not be argued from the grammar axis using this table, because a reviewer who opens Appendix B will find that axis untested. Recorded here rather than by editing discipline (1), per the rule at the top of this entry.
• What this does not resolve: see the successor challenge immediately below, which is the live objection now.
• File: knowledge-base/cohen-kadach-ormazabal-reichelstein-esg-pay-international-2023.md (rewritten 2026-09-01 on the verified full text; the 08-31 COUNTERARGUMENT tag is withdrawn, and the citation corrected to 61(3), June 2023 — the 08-31 entry recorded 61(4)).

---

• Logged: 2026-09-01 · Targets: Momentum Mirage as a diagnostic; the "Paid for Deployment" pattern; Paper 3 mechanism 3
• The objection, at full strength. Grant that Cohen et al. Table 8 shows result-defined metrics move outcomes and result-agnostic ones do not. That result may not be about metric grammar at all. Carbon worked because CO2 is countable in tons by Trucost — an independent, standardized, third-party measurement that existed before the metric was written. The distinction the framework calls activity-defined versus result-defined may be substantially confounded with whether the domain admits third-party measurement in the first place. Note that every one of the nine null categories — culture, employee satisfaction, compliance, governance, "other" — is a domain with no Trucost. In 2026, neither does AI. There is no independent, standardized, cross-firm measure of an AI transformation's result. If no company can yet write a defensible AI outcome metric, then Qorvo's "exploration and deployment of AI tools" is evidence of an attribution problem, not of an organizational failure — and the framework is diagnosing a measurement limit as a breakpoint. That is a category error, and it is the same error the base has spent two weeks accusing others of.
• Evidence for the objection: Cohen et al. Table 8's own category structure — the single significant coefficient sits on the single category with an established external measurement infrastructure. Also the base's standing finding (unfalsifiability challenge, updates 2026-08-30) that no field has built an instrument linking an ex-ante organizational measure to a measured implementation outcome, which is the same absence one layer down.
• Evidence against / limiting the objection: (1) Measurability and grammar are separable in principle and the separation is testable — a metric can name a result in a measurable domain and still be written without a threshold, and that cell is where the two accounts diverge. Nobody has looked at it. (2) Safety and security is a partial counterexample worth pressing: injury and incident rates are externally reported in many jurisdictions, and that category is still a null (−0.05, t = −0.34). If the confound were the whole story, safety should have behaved like carbon. STRUCK 2026-09-02 — this was bad reasoning and it was mine. The safety coefficient is measured against ∆CO2. Safety metrics have no reason to reduce carbon emissions, so that cell is a mismatched-dependent-variable null exactly like the other seven non-carbon categories, and it cannot distinguish anything. The paper never regresses safety metrics on injury rates. See the 2026-09-02 update below. (3) The objection, if it wins, relocates the framework rather than refuting it — from the wording of the metric to the measurement infrastructure the organization has or has not built, which is a Capability claim and still squarely inside the Four Forces.
• What it would take for this objection to win: evidence that within externally measurable domains, result-defined and activity-defined metrics perform alike — i.e. that grammar adds nothing once measurability is held fixed.
• What it would take for it to be defeated: the safety-and-security category in Cohen et al. turning out to be measurable-but-activity-worded, which would show grammar operating independently of measurability inside their own data; or any instrument comparing threshold-bearing and threshold-free metrics within one measurable domain.
• Current best response, and it is cheap. Read Cohen et al.'s Appendix B, which gives worked examples of each metric category, and Table 3, which gives the taxonomy and per-category firm counts. If the safety metrics are worded as programme activity rather than as injury-rate targets, this challenge weakens sharply on evidence already downloaded. Next action: Appendix B and Table 3 of the Mannheim PDF. RUN 2026-09-02 (Wednesday falsification sweep). It did not weaken the challenge. See below.

Update 2026-09-02 — the cheap test was run, the challenge survives it, and the route to settling it turns out to be shorter than anyone thought

(1) The named escape route is closed. 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) — an OSHA-defined, mandatorily reported, cross-firm-standardised rate. Safety metrics in this sample are result-defined targets in a domain with pre-existing third-party measurement, not programme activity. The condition under which this challenge would have weakened does not hold.

(2) The safety test never existed, and limiting evidence (2) above is struck. The −0.05 was always against ∆CO2. Safety metrics have no reason to move carbon. The paper does not regress safety metrics on injury rates, although DART data are as externally measured as Trucost tons — so the one within-paper comparison that could have separated measurability from grammar in a physical domain was simply not run.

(3) The objection gets a new and stronger limb, and it is aimed at the defence rather than at the framework. Appendix B shows metric grammar varying inside Cohen et al.'s categories, not between them — carbon ("kg CO2e/tonne"), safety (DART rate), diversity ("percentage of women among the SMP") and employee satisfaction ("internal promotion rate") are all crisply result-worded, while compliance ("continue to assess human rights, bribery and corruption and other related risks"), governance ("establish standalone corporate governance and risk procedures... that build trust... and align with company values") and corporate culture ("Colleague Culture & Engagement survey") are activity-worded. So Cohen et al.'s decomposition is a subject-matter taxonomy, and the −0.07-versus-−0.77 contrast isolates topic specificity, not activity-versus-result grammar. The 09-01 defeat of the soft-metric objection stands — the sign and the comparison are unaffected — but it was won on a different axis than it was recorded on, and any published use must say "naming the specific outcome" rather than "naming a result." Recording this here rather than only in the KB file because it is the kind of correction a ledger exists to hold.

(4) The thing that actually moves this challenge: the variable exists, was constructed, and was never estimated. Table 3 Panel A defines eleven categories, not nine. Alongside the nine specific indicators sits family (b) Scores, split exactly on measurability: Self evaluation — "scores defined and measured by the firm", 884 firms — versus External evaluation — "scores defined and measured by external parties", 97 firms. Grammar is approximately fixed across that split (both say "achieve a score"); who measures is what varies. Neither category appears in Table 8 or in any column of Table 9, and neither is discussed anywhere in the body (full-text pass, pp. 1–45); they appear only in Table 3 and Appendix B. Table 3's note (2) records a sample restriction on the score counts, which may be the reason — no inference of suppression is available or intended. The consequence is what matters: this challenge is not resolvable from the paper's published tables, but it is resolvable from the paper's data, by a regression whose variables already exist. That is a far cheaper resolution than "nobody has looked at it" implied when this challenge was logged.

(5) One finding that cuts toward the framework and should not be buried in a challenge entry. Table 9's carbon row: the one contract term shown to move a physical outcome lands at Refinitiv +0.001 (t = 0.07), a precise zero on the disclosure-based score; Sustainalytics −0.583 (t = −2.12), significant at 5%; KLD +6.660 (t = 10.82). Appendix A defines Sustainalytics and KLD so higher = better performance, so these are sign-comparable and they disagree. Whatever the explanation — poor raters, or composite shift away from other pillars — no reading is available in which the appearance measure and the substance measure track each other. Caveats that must travel: KLD rests on 1,351 observations against 17,148 and 19,252; the authors defer to Berg, Koelbel & Rigobon on vendor divergence and call all of Section 6 "mainly descriptive." Do not state this as "ESG ratings are wrong."

Revised status: OPEN, unchanged in force, cheaper to settle. What it would take to win and what it would take to defeat are both unchanged. What changed is the next action.

Next action (revised). Not a new study and not another read of this paper. Either (a) ask the authors, or a co-author's replication package, for the Table 8/9 specification with the two score dummies included — the single most efficient step available and the only one that separates measurability from grammar inside an existing dataset; or (b) accept that the framework's metric claim is currently supported on the topic-specificity axis only, and write it that way. (b) costs nothing and should be done regardless of (a).

Update 2026-09-03 (Stafford) — one piece of limiting evidence, adjacent not decisive

The Alibaba field experiment the scan ingested 09-03 supplies a general-purpose rebuttal to the deeper worry underneath this challenge — that appearance/substance divergence might be a measurement-infrastructure artifact rather than a real organizational phenomenon. In that deployment the external outcome measure existed (a 1–5 customer star rating, read off platform logs) and the organization's aggregate reporting still concealed the substance: the touched-work rating fell −0.412 (p<0.001) while the all-chats coefficient was +0.055, indistinguishable from zero. So concealment there was a reporting/aggregation choice, not an absence of measurement. The disciplines matter and bound the point. This challenge is specifically about exec-pay metric grammar in domains lacking third-party measurement (does Momentum Mirage reduce to "AI has no Trucost"); the Alibaba case is about dashboard aggregation where measurement was present. So it is adjacent limiting evidence, not a resolution — it shows appearance/substance divergence is a real design artifact in at least one measured domain, which weakens the "it's only a measurement limit" reading, but it says nothing about whether activity-vs-result grammar predicts in unmeasured domains. Logged as an "evidence against / limiting the objection" data point; the next action above is unchanged.

---

• Logged: 2026-09-03 (Stafford exploratory slot) · Targets: thesis; Claim 3; Claim 5 (the redesign-prerequisite claim)

• The objection, at full strength. The framework's load-bearing prerequisite claim is that AI value requires organizational redesign — deploy a tool onto an unchanged role and you get the appearance of gain, not the substance. The most-cited AI field experiment in existence says otherwise, on an externally measured quality outcome, in the same function the base's new flagship result comes from. Brynjolfsson, Li & Raymond, Generative AI at Work (NBER WP 31161, 2023; QJE 140(2):889–942, 2025) staggered a generative-AI assistive chat tool across 5,172 customer-support agents at a Fortune 500 firm and measured, off platform logs: +14% resolutions per hour on average (+34% for novices, ~0 for experts), improved customer NPS, and lower attrition. No role was redefined; the tool was laid over the existing agent workflow. That is precisely the falsifier the base's own 2026-09-02 falsification sweep went looking for — "measured AI returns without workflow or role change" — declared not to exist, and it not only exists, it is famous and its co-author (Brynjolfsson) is already in the base via the Stanford Enterprise AI Playbook. A framework that predicts tech-on-unchanged-conditions fails has to explain the single largest, cleanest case of it succeeding.

• Evidence for the objection: knowledge-base/brynjolfsson-li-raymond-generative-ai-at-work-2023-04.md — QJE, staggered/administrative design, external outcomes (resolutions/hour, resolution rate, NPS), quality-positive on every axis.

• Evidence against / limiting the objection (this is why it is CONTESTED, not OPEN, and none of it is capitulation): (1) The tool was the organizational intervention. BLR's own mechanism is that the AI codified the tacit best practices of top agents and disseminated them to novices — a capability / knowledge-flow intervention, not "pure tool on unchanged conditions." The 34%-for-novices / ~0-for-experts structure is the signature of a capability substitute, and Capability is inside the Four Forces. Whether "AI as knowledge-dissemination layer" counts as the redesign the framework requires is the crux, and it is a genuine question about the framework's scope rather than a dodge. (2) Assistive, not agentic — the human never left the loop. BLR's tool recommends and the human executes every chat; yesterday's Alibaba result is the agentic-with-handoff case in the same function and it came back the other way (rating −0.412 on AI-touched chats). The two do not contradict — together they isolate what autonomy costs, which is the more valuable reading than treating either as decisive. (3) BLR contains its own Momentum-Mirage tail: top workers increasingly follow AI suggestions even when those suggestions are worse — measured judgment erosion, the framework's own mechanism, inside the pro-tool paper. (4) Vintage: a 2020–2021 GPT-3-era deployment; argumentative value, not a 2026 frontier claim.

• What it would take for this objection to win: a redesign-free deployment producing durable, externally-measured gains where the mechanism is demonstrably not a capability/knowledge intervention the org failed to do itself — i.e. genuinely "tool alone," with the organizational conditions left untouched and not substituted-for by the tool.

• What it would take for it to be defeated (absorbed): establishing that BLR's knowledge-dissemination effect is itself an instance of the readiness/capability the framework names — which reframes it as confirming the thesis (the tool succeeded because it supplied the missing capability) rather than refuting it. This is the likely resolution and it turns on the scope definition in limiting-evidence (1).

• Current best response: the counterexample is real, correctly aimed, and must be met on the merits, not waved off — but the honest framework reading is that BLR succeeded because the tool did organizational work (codifying and moving capability), which relocates the win inside the Four Forces rather than outside the thesis. Pair it publicly with Alibaba as the autonomy contrast; do not cite the redesign-prerequisite claim as if BLR did not exist. Next action: ingest the rest of the field-experiment canon (Dell'Acqua et al.; Peng et al.; Noy & Zhang) so the base argues this from the whole evidence class, not one paper — see the PROPOSED-UPDATE / structural-gap note in today's brief.

• STRENGTHENED 2026-09-06 (Stafford exploratory slot) — a second, independent instrument, and it closes the escape route that made this CONTESTED rather than OPEN. The 09-03 entry above sets the win condition as "a redesign-free deployment producing durable, externally-measured gains where the mechanism is demonstrably not a capability/knowledge intervention the org failed to do itself." Rotenstein LS, Holmgren A, Thombley R, … Adler-Milstein J, Mishuris RG, JAMA 2026;335(16):1408–1417 (already in the base since 2026-08-27 as knowledge-base/rotenstein-jama-ai-scribes-clinician-time-multisite-2026-04.md, research-evidence) meets that description on all four clauses where BLR meets it on two. (a) The mechanism is not knowledge dissemination. An ambient scribe transcribes an encounter and drafts a note; it substitutes for a documentation task and disseminates nobody's tacit expertise. Limiting-evidence (1) above — "the tool was the organizational intervention" — has no purchase here. (b) The organizational conditions were verifiably untouched, in the paper's own words: "none of the organizations in the study required that clinicians book additional patients to qualify for AI scribe use." Stated, not inferred. (c) Externally measured (EHR-system-derived, not self-report) and durable (26 months, five academic health systems, 1,809 adopters against a 6,772-clinician comparison group). (d) The gain is real and bounded away from zero: +0.49 weekly visits (95% CI, 0.17–0.81) overall, +1.0 (0.5–1.6) among clinicians using the scribe in ≥50% of visits.

• What this changes, and it is a change to the claim rather than to the response. The strong form of Claim 5 — organizational readiness is the prerequisite — predicts approximately zero output from deployment onto an unchanged workflow. It is not zero. The defensible restatement is a magnitude-and-distribution claim: without redesign the gain is real but small, captured by the minority who changed their own practice, while the outcome the investment was justified by does not move at all (off-hours EHR time, the burnout-proximate measure, showed no significant change across the whole cohort). Every number in the paper supports that version, and it is falsifiable where "prerequisite" is not.

• The dose-response is the part that should unsettle the framework, and it was in the base unread. Against the all-adopter estimates, heavy users show 1.59× the total-EHR-time saving, 1.71× the documentation saving, and 2.04× the visit gain. The conversion from freed time to output does not degrade with dose — it holds or improves. The framework's instinct is that organizations leak the gain in conversion; on this instrument they do not. The population-level shortfall is produced by breadth (21% adoption; ~32% of adopters at ≥50% intensity), not by leakage. That is consistent with threshold-shaped returns — below some intensity the saved minutes never aggregate into a bookable slot — which is a more precise and more useful meaning for "readiness" than the framework currently gives it. Discipline: these are separate subgroup estimates on a self-selected subgroup in a cohort that was opt-in at four of five sites; consistent with a threshold, not distinguishable from selection. Do not cite the 2.04× as a dose-response.

• Convergent on the wording, not on the evidence. The research-evidence scan reached the same restatement the same day from Applause's rollback finding — "not 'AI without redesign produces no gain' but 'AI without redesign produces a gain too small to be worth running.'" Different constructs, different populations, and Applause will not disclose its denominator: these are not convergent evidence and must never be cited as such. They are two independent routes to the conclusion that Claim 5's wording is the thing that needs fixing.

• Why this is not logged as a separate challenge. It is the same objection with a second instrument. A second ledger entry would double-count one argument, which is the failure this file disciplines elsewhere.

• Status: remains CONTESTED, and closer to winning than it was on 09-03. Resolution now turns on whether Brandon restates Claim 5 (in which case the objection is ABSORBED — the thesis revised to account for it) or defends the prerequisite form (in which case it stays live and he should expect a referee to find Rotenstein). Recommended: absorb it before publication. Retiring an overreaching claim of his own is the same move as Paper 3's retirement of the 70% figure, and cheaper while the paper is a draft.

---

• Logged: 2026-09-04 (Stafford exploratory slot) · Targets: the base's own architecture — the "Unconverted Gain", "Migrating Bottleneck" and "Collapse the Handoffs" patterns — and, through them, the machine-speed engine of paper-2
• The objection, at full strength. Every pattern the base has built around AI's organizational failure shares one unexamined premise: the individual-level gain is real and large, and it dies between the person and the organization. That premise has never been tested on the population enterprise engineering actually consists of — experienced practitioners working on mature systems they know deeply. It has now been tested, under randomization, and the gain is not merely unconverted. It is negative. Becker, Rush, Barnes & Rein, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (METR, arXiv 2507.09089, 2025-07-12): 16 experienced open-source developers, 246 real tasks in their own mature repositories, task-level randomization of AI-tool permission. AI access increased completion time by 19%. If that generalizes, "the gain evaporates at the organizational handoff" is the wrong story. The organization is being charged for a deficit the tool created, and Brandon's diagnostic would be attributing to organizational design a shortfall that exists at the keyboard.
• Evidence for the objection: knowledge-base/becker-rush-barnes-rein-metr-early-2025-ai-developer-productivity-2025-07.md. Randomized, direct time measurement, real tasks, no vendor interest — METR is an AI-evaluation nonprofit with no product in the enterprise-AI market. Cited by NBER WP 35275 itself as the "mixed" evidence on agentic tools.
• Evidence against / limiting the objection, and none of it is capitulation: (1) n = 16. The 246 tasks carry the statistical power; sixteen people do not carry a population claim. The finding supports these sixteen maintainers, on their own codebases, with early-2025 tools, and nothing wider. (2) Tool vintage. Cursor Pro with Claude 3.5/3.7 Sonnet, early 2025 — a generation and a half behind the async agents WP 35275 measures. The market's current argument is about a technology this study did not test. (3) Different outcome, not a contradiction. METR measures completion time per task; 35275 measures volume of artifacts produced. A tool can raise output volume and lengthen per-task time simultaneously. The two instruments are in tension about what "productivity" names, not about a fact. (4) 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. Neither instrument has reached the enterprise.
• What it would take for this objection to win: a replication at scale, or any instrument showing negative measured returns for experienced practitioners on the current tool generation. Either would force the framework to state a regime condition it currently does not state — that its diagnostic applies where the individual gain is real, and that there exists a regime where it is not.
• What it would take for it to be defeated (absorbed): evidence that the negative effect is specific to early-2025 tooling or to the open-source maintenance task genre, and does not appear where AI is deployed on the work enterprises actually do. Note that absorption here is not free: even a defeated version leaves the framework owing a sentence it has never written — the diagnostic assumes a real upstream gain, and that assumption is an empirical condition, not a background fact.
• The half that helps. The same instrument contains the tightest Momentum Mirage measurement in the base and it is at the individual altitude where books/built-to-be-replaced.md operates: participants who were 19% slower estimated afterwards that they had been 20% faster — a ~39-point gap between measured and perceived progress, under randomization, on the same people, with the counterfactual held. Expert forecasters were wrong in the same direction (economists 39% predicted speedup, ML researchers 38%). Every other Momentum Mirage instrument here measures an organization mistaking activity for progress; this measures a person doing it.
• Standing instruction, effective immediately: do not cite WP 35275's upstream gain without citing METR's negative one. The base's brand is that it checks the numbers everyone repeats; citing one of a paired result and not the other is the failure mode it exists to avoid.
• Next action: chase the rest of the negative-result class — whether any replication of METR exists, and whether Dell'Acqua et al. (Organization Science, in advance) reports a comparable below-the-frontier decrement. Both are cheap and neither has been run.

---

• Logged: 2026-08-31 · Targets: Momentum Mirage as a diagnostic; the "Paid for Deployment" pattern; Paper 3 mechanism 3
• The objection, at full strength: The framework infers Momentum Mirage from the form of a metric — a goal with no stated result condition is read as rewarding the appearance of progress. That inference has never been tested. It has now been tested on the previous non-financial metric to enter executive pay, and it failed. Cohen, Kadach, Ormazabal & Reichelstein, Journal of Accounting Research 61(4):805–853 (2023), document ESG metrics spreading through executive compensation contracts internationally and report that "the adoption of ESG variables in managerial performance measures is accompanied by improvements in ESG performance." ESG metrics in pay plans are notoriously soft, aggregate and discretionary — Dell'Erba & Gomtsyan (JCLS 2024) build a whole critique on exactly that — and the outcome still moved. If a soft metric moves the outcome, then reading Qorvo's "exploration and deployment of AI tools" at 20% of an LTIP as Momentum Mirage is a claim about metric grammar that the one available piece of evidence does not support. Worse for the framework: it is the same evidence pattern the base would cite approvingly if the sign went the other way.
• Evidence for the objection: knowledge-base/cohen-kadach-ormazabal-reichelstein-esg-pay-international-2023.md — peer-reviewed, top-tier accounting journal, international sample, and the only test of this proposition the base holds on any metric.
• Evidence against / limiting the objection, and it is not yet strong enough to rely on: (1) The authors' verb is "accompanied by." That is association, and the abstract claims nothing more. The result is compatible with firms that were already improving being the firms that adopted the metric. (2) The outcome variable is unknown and it decides the objection. If ESG performance is measured by ratings that are themselves largely disclosure- and activity-scored, the finding partly reduces to "paying executives to score better on a disclosure index makes them score better on a disclosure index" — which would leave Momentum Mirage untouched and, read carefully, would be an instance of it. This cannot be asserted, only queued: Wiley, SSRN and T&F all returned 403 and only the abstract was read. (3) Dell'Erba & Gomtsyan, from the same practice, reach the opposite normative conclusion — the short-term goal's attainment "does not necessarily translate into better overall financial performance or more responsible corporate behaviour in the long-term." Legal scholarship looking at the same instrument sees what the framework sees.
• What it would take for this objection to win: Cohen et al.'s outcome variable turning out to be a substantive, externally measured outcome (verified emissions, injury rates, audited diversity data) rather than a rating — plus any AI-specific replication showing activity-defined AI metrics associated with subsequent conversion. That combination would force Momentum Mirage to be restated as a claim about accountability architecture rather than about metric wording.
• What it would take for it to be defeated: the outcome variable turning out to be rating-based, which converts the headline finding into support for the framework rather than against it.
• Current best response: the objection is live and correctly aimed, and the honest position today is that the base tagged a breakpoint from metric grammar without ever having tested whether metric grammar predicts anything. Do not soften the "Paid for Deployment" pattern yet — it is logged at two sources and makes no prevalence claim, so it is not over-extended. But the pattern's Momentum Mirage mapping should carry this challenge by name until the outcome variable is read. Next action: obtain the JAR article and read its outcome construction. That single step resolves the challenge in one direction or the other and is the cheapest open item in this file.