Stafford Brief — 2026-09-20
Delta layer over today's 12:00 research-evidence scan (3 staged, 5 skipped; Sunday — no weekly synthesis). Read that brief first; this one does not repeat it.
Operations: clean. Yesterday's run merged to main (8b21e1c), today's started from main, nothing outstanding.
Today's scan is the best day this pipeline has had in a month, and I checked its two big finds against the primary documents. The facts hold. Three of the judgments built on them do not. I read the IFS working paper in full, and its mechanism section says something the scan's breakpoint mapping does not survive intact. I re-read the Reuters report, and the reason it gives for Meta's reversal is no reason — the numbers the scan calls the cause are dated two weeks after the decision. And the Migrating Bottleneck pattern was incremented with a case that violates the condition in the pattern's own title. All three are fixable today, and the fixed versions are stronger than the originals.
📖 TODAY'S READ
The same paper the scan led with — Merritt, Dominski, Hoy & Lee, IFS WP 26/23 — read end to end from the PDF, because the part that matters is in §4.3 and no summary carries it.
The point in one sentence: the paper measures why managers pulled back, and the answer is that they stopped believing in the tool — not that they were protecting themselves.
Every number the scan reported is correct. I verified each one: 1,995 managers reached randomisation (995 UK, 1,000 US), over 70% positive toward AI beforehand, 0.4–0.5 standard deviations off adoption and advocacy, 0.2 off staffing, and the bounding sentence on the productivity arm is verbatim. The abstract's own summary of direction — "Rather than inducing substitution away from labor and toward AI, information about AI's labor-displacing potential leads managers to scale back both planned AI adoption and their staffing intentions" — is quoted accurately. Nothing here is a factual correction.
The correction is to what the finding is evidence of. The scan filed this as the strongest Incentive Fragmentation evidence the base has ever held, and described it as Paper 1's CISO self-check run as an experiment — the leader whose own metrics won't improve, declining to prioritise the work. That is a story about a manager's stake. The paper's mechanism section is a story about a manager's beliefs, and the authors take it apart line by line.
Roughly 70% of the drop in AI adoption, and over 80% of the drop in advocacy, is accounted for by two measured channels: the manager's rating of AI's benefits went down, and his rating of AI's risks went up. In plain terms: he watched a two-minute video about AI displacing workers and revised his estimate of how good the technology is. That is not an incentive failure. It is a person updating on information — and the framework has a different name for what follows from it.
The part that would be Brandon's mechanism is the part the authors say they cannot explain. Staffing intentions fell 0.2 SD, and the belief channels account for almost none of it — under 16% through benefits, under 6% through risks. The authors' own words: "the reduction in staffing intentions cannot be explained by any of our measured belief channels, suggesting that the labor treatment may trigger a more general retrenchment that extends beyond attitudes toward AI itself." Their speculation is "an overall sense of uncertainty."
So the flagship Incentive Fragmentation claim currently rests on the authors' admitted residual. A hostile reader with the PDF open finds that in five minutes, and the base has spent six weeks cataloguing exactly this failure in other people's citations.
There is a second sentence in the paper that cuts the same way and that nobody has quoted. The authors state directly: "employment and AI adoption are not intrinsically in tension within a firm or business unit." That is the authors declining the framing the scan's Routing note builds on.
What survives, and it is still the thing worth having. The asymmetry is untouched and is not a belief story: both videos were information, and only one moved anybody. The organisation's evidence moved this layer by an amount you can bound near zero; information the organisation does not control moved it half a standard deviation. That claim is fully supported, it is Brandon's argument, and it does not require anyone to have protected their scorecard. The version that must not ship is "managers protected their own headcount" — unmeasured, and contradicted by the authors' own reading.
So what: the paper-2 case file entry written today needs one bounding paragraph before Claim 3 leans on it. PROPOSED-UPDATE in Routing. New challenge logged at knowledge/thesis-challenges.md.
🔭 THINKER TO WATCH
Nicole Forsgren — and she is here because of an absence this system asserted two days ago and never checked.
The 2026-09-18 Google challenge states that this base holds zero files on SRE, error budgets, SLOs, Forsgren, or the DORA/Accelerate research programme, and that the objection "wins by default if Brandon publishes the two-variable diagnostic without acquiring any of that literature." I ran the grep across both repositories. The assertion was true — every hit is Brandon's own practitioner account or the challenge ledger quoting itself. Neither query log has ever searched for it. A named, load-bearing, self-identified gap has sat unchased for two days.
Why she matters specifically. Forsgren, with Jez Humble and Gene Kim, built Accelerate (2018) and the DORA research programme on a claim structurally identical to Brandon's: how an engineering organisation delivers predicts how the business performs — delivery performance and organisational performance are one subject, not two. That is the interface-accountability argument, made in a different vocabulary, backed by multi-year survey research across thousands of organisations, and already the default vocabulary of every engineering leader Brandon would be talking to.
Where he is exposed, and it is narrow and real. He is preparing to argue about DevOps — error budgets, shared consequence, the sponsor who installs the measure — in front of an audience for whom this programme is the literature. DORA added Reliability as a fifth outcome in 2022, assessed through SLOs and error budgets — the exact mechanism Brandon's Google case turns on. The first reader who knows this will not disagree with him; they will notice he did not cite it, which is worse.
Where he is genuinely ahead, and this is the useful half. DORA measures capabilities and outcomes — does the organisation do the practice, and does it perform. It does not measure what Brandon's two-variable diagnostic asks: whether a shared measure was installed between two parties who had separate leadership and separate metrics, and what happened to it. Today's search returned nothing measuring that anywhere. So acquiring this literature answers the "you're arguing without citing the field" charge and leaves Stafford's timing conjecture — that the window for installing a shared interface measure closes once both sides have their own leadership and their own number — still unchecked by anything published. That is a gap Brandon could own rather than a gap he is caught in.
So what: this is a reading assignment, not a search. Two documents — Accelerate and the most recent DORA report — convert a CONTESTED challenge into an argument he wins. Recorded against the challenge; no KB file, because nothing was read end to end today and this base does not file what it has not read.
🏢 CASE IN THE WILD
Meta, Project OT — and the correction is one sentence long: Reuters does not say why the second wave was cancelled.
The scan's brief states it flatly — "The internal numbers are why" — and then cites Bosworth's 220%-against-36% post. I re-read the report for the chronology. There are two problems and they compound.
First, the dates run backwards. Wave 2 was cancelled on the night of 19 May. Bosworth's post carrying the 220%/36% figures is dated early June — the report's own words are "according to a post by Bosworth in early June." The incident and firefighting figures carry no date at all. The evidence being offered as the cause of a decision was written roughly two weeks after the decision was made.
Second, the report declines to name a cause. Its sentence is "on the night of May 19, just hours before the first layoff wave, Zuckerberg blinked" — no reason attached. It names two conditions, in this order: "By then, Meta employees were in open revolt, convinced that the company's AI transformation initiatives were partly aimed at replacing them. Internal data was also suggesting that autonomous AI 'agent' technology ... was failing to deliver hoped-for productivity gains." Reuters puts the workforce first and the data second, as a supporting clause. The scan kept the second and dropped the first.
Why this is worth a correction rather than a footnote. The scan withheld a Commitment tag from Meta deliberately, as an act of framework integrity — cancelling on the strength of internal metrics is a leader acting on evidence under pressure, which is what the third Force asks for, and tagging it "would invert the framework to make a company look worse." That reasoning is right in principle and unsupported in fact. Giving a company credit on an unevidenced causal chain is the same error as tagging one; it just runs in the flattering direction, which makes it harder to catch and more embarrassing to be caught at.
And the two candidate causes are not interchangeable — they are opposite readings of the same event. If the metrics stopped it, Meta is a company that measured well and corrected. If the revolt stopped it, Meta is a company whose workforce supplied the brake its governance layer did not — which is a considerably more interesting finding, because it is the organisation functioning as the control the design left out. Neither is established. The base should hold both and claim neither.
The open question the thread should carry: the October 2026–January 2027 window will produce commentary on whether the capability arrived. It is unlikely to produce anyone saying why Zuckerberg blinked on 19 May — and that, not the three-to-six-month claim, is the fact that decides which case study this is.
So what: two PROPOSED-UPDATEs in Routing — bound the causal sentence in the KB file and the case file, and add the cause question to the Project OT thread's check criteria.
⚙️ FOUR FORCES CONCLUSION
Commitment, and the movement is subtraction: the exculpation Meta was granted today is withdrawn pending evidence.
The non-tag stands as a decision not to claim a failure — that part was correct and stays. What does not stand is the reason given for it. "Meta acted on evidence under pressure" is a Commitment finding, and it requires knowing what Meta acted on. Nobody does. The honest state is unresolved, and unresolved is a perfectly respectable entry.
This is the discipline the base already applies in the other direction. Incentive Fragmentation was considered and deliberately not tagged against Meta today because "nothing in the reporting measures or even names it." The same test applied to the Commitment reasoning fails it. A framework that refuses speculative tags and accepts speculative exemptions is not being rigorous; it is being polite.
💭 OPEN QUESTION
The managers who collapsed hardest were the ones who were most enthusiastic. Does that mean the Momentum Force is tracking something that is a liability rather than an asset?
This is in the IFS paper's heterogeneity section and it is the most uncomfortable thing in it. Over 70% of the sample was positive toward AI beforehand. Among the minority who were not positive, the displacement video did almost nothing — the effect is near zero. The entire pullback is carried by the enthusiasts. The authors state it plainly: managers with a more positive AI posture had higher baseline staffing plans and showed "exaggerated retrenchment in staffing intentions in the labor arm." The higher they were, the further they fell.
Why this is a real problem and not a paradox to admire. The Four Forces define Momentum as sustained energy — transformation not losing force as novelty fades. The implicit model is that enthusiasm is fuel and the risk is that it runs out. This says something different: enthusiasm was the fragile state. The sceptics had nothing to revise. The believers had a position to lose, and two minutes of contrary information took half a standard deviation of it.
Three readings, and they lead to different instruments. (1) Ceiling. The enthusiasts simply had further to fall; the finding is arithmetic, not psychological, and Momentum is fine. (2) Unearned momentum is brittle. Energy built on general optimism rather than on the manager's own verified experience has nothing under it — which would mean Momentum should be measured by what the manager has personally confirmed, not by how positive he sounds, and every readiness survey in this base measures the wrong one. (3) Momentum is downstream of Purpose and Capability and has no independent existence — in which case the Force is a symptom the framework has promoted to a cause.
Brandon can discriminate these with an instrument he could field. Ask managers two things: how positive they are about AI, and what they have personally verified about it in their own unit. Reading (1) predicts the pullback tracks the first. Reading (2) predicts it tracks the gap between them. Nobody in this base has separated the two, and the paper cannot — it measured sentiment, not verification.
So what: this is the second time in four days that a Force has come back as possibly-not-independent. On 09-18 the concern was that monitorability was explaining too much with one term. This is the mirror image, one layer down. Worth watching as a pair, not as two incidents.
🥊 CHALLENGE
"Your best causal evidence on the manager layer is a belief-updating result, and you have filed it as an incentive result."
Stated at full strength: Merritt et al. randomised information and measured what mediated the response. Seventy to eighty percent of the effect on AI adoption and advocacy runs through the manager's revised beliefs about the technology — benefits down, risks up. Nothing in that chain involves his headcount, his scorecard, his span, or his bonus. A manager who watches evidence that a tool displaces workers, concludes the tool is riskier than he thought, and advocates for it less is behaving exactly as a well-functioning manager should. That is not a breakpoint. That is a person reading the room correctly.
The objection's sharp edge: this is the same class of problem as the 09-19 specification objection, and the two now rhyme. In both cases the framework's incentive layer was assigned a finding whose measured mechanism sits somewhere cheaper and more ordinary — a badly written clause in one case, an updated belief in the other. Twice in two days the base has reached for Incentive Fragmentation where the source offers something duller and better evidenced. That is a pattern in the analysis, not in the world, and it is the kind that gets a framework accused of explaining everything.
What it would take for the objection to win: it wins by default if the case file's Claim 3 entry ships as written. It is answered — not defeated — by the asymmetry: both arms delivered information, and only the one touching the manager's own unit moved anything. Pure belief updating does not predict that. The honest position is that the asymmetry is incentive-flavoured and the mediation is belief-flavoured, and the paper cannot separate them because it measured beliefs about AI and never measured beliefs about the manager's own position.
What would settle it: the Defensive Gate pattern's pre-registered second-member bar already asks for the right thing — behaviour, against what the deployment implies for the manager's own unit. One clause should be added to it: the instrument must also measure the manager's beliefs about the technology, or it will reproduce this ambiguity at higher cost. Logged OPEN at knowledge/thesis-challenges.md.
📈 PATTERN BUILDING
The Migrating Bottleneck went from 4 sources to 5 today, and it should go back to 4 — for a reason the scan itself supplied.
The pattern is titled "AI Productivity Without Organizational Redesign" and its statement is that AI boosted individual productivity "without redesigning the management layer that translates that output into organizational outcomes." The absence of redesign is the pattern's scope condition, not decoration.
Meta was added as member five explicitly because it redesigned. The scan's own words: "Meta redesigned — two-to-three-person pods, middle layers eliminated, Org Leads at 30–50 spans — and the bottleneck migrated anyway." You cannot count a case as an instance of "X happens without Y" on the grounds that Y happened. An ESCALATED pattern's count is load-bearing, and this one now includes a member that does not instantiate its claim.
The base already knows how to handle this, and it did so twice — once today. NBER WP 35275 was immune to every charge against the pattern and was held at 4 because it measured no structure. Bartolucci & Vivo was the pattern's best explanation and was held at 4 because it measured nothing. And this morning the Defensive Gate declined Wadi & Ma as a second member on a level-of-analysis argument, in the same brief. Three refusals of exactly this kind, and then the scope condition was waived for the best story of the week. Good stories are when the discipline is load-bearing.
Nothing here argues for discarding Meta — it is the most valuable thing in the base this week, and it is being filed in a way that wastes it. Counted as member five, it is one more tally on a pattern that already had four. Filed correctly, it is something the pattern has never had: the first case where the remedy was attempted and the bottleneck migrated anyway. That is not more evidence for the existing claim. It is the first evidence for a stronger and separate one — redesign aimed at the generation step relocates the constraint rather than removing it, because the design was specified against the tool's capability instead of against the constraint. That proposition is Brandon's Flattening Bet, and it now has its first named instance with internal numbers attached.
So what: PROPOSED-UPDATE in Routing — revert the count to 4, keep the entire Meta write-up as a boundary case under a named scope amendment, and record the stronger claim as the thing the case supports. The one-primary-many-echoes discipline the scan attached to the member travels with it unchanged.
🪟 WINDOW FLAG
None. The scan argued the Meta window down at length this morning and the argument was right; nothing in today's verification changes it. One addition to the pre-registered trigger: the scan says a window opens if the second wave returns or the three-to-six-month claim is restated or dropped. Add — if any account emerges of why 19 May was reversed. That is the fact the case study turns on and the only one nobody currently has.
🧱 ASSET SUGGESTION
Amendment to asset 2, "The Flattening Bet" — the essay has been a prediction without a completed instance. It now has one, and the instance also tests the prediction the essay gained on 09-18.
Asset 2 argues the live delayering wave is the next transformation failure the series describes. Project OT is that wave's first instance to run start to finish with internal numbers and a reversal attached: layers eliminated, spans widened to 30–50, a new role title, executed in May and abandoned in November before November arrived. An essay predicting a class of failure gets materially stronger when it can open on a named company that already completed the cycle.
And it discriminates the essay's two rival mechanisms, which is the thing the 09-18 amendment said could not be done without new data. That amendment set up a contest: the coordination-cost story predicts hierarchy shrinks; the monitorability story predicts authority re-sorts while layer counts barely move. Meta did both at once, and the second one on purpose. It cut layers — coordination story — and specified Pod Leads running day-to-day work without formal authority, which is decision rights deliberately detached from reporting lines. That is the re-sorting signature, written into a target operating model rather than inferred from one. The essay can now say: here is what it looks like when both mechanisms run in the same company in the same quarter, and here is which one the design was conscious of.
The bound that must travel with it, in Brandon's own voice: every Meta figure is a journalist's reading of internal documents nobody can audit, on Meta's own definitions; Meta disputes the framing on the record; and — per today's Case section — the reason for the reversal is not reported, so the essay must not claim Meta stopped because the numbers told it to. Full spec at knowledge/asset-suggestions.md. Status: asset 2 stays DRAFTED v4, third consecutive amendment without movement, which is itself the finding the ledger has been recording since 09-11.
🎓 SPARRING NOTE
Not issued — the cap binds, and the reason is logged rather than left as a gap. Two notes ran this week (09-14, the sociotechnical concession; 09-17, the Kotter paragraph), both unanswered. A third would be noise stacked on two live prompts.
SOURCES & THREADS
Threads: none due today for Stafford. The scan advanced The middle-manager load-bearing problem and opened Project OT; nearest remaining check is 2026-09-23 (team-dynamics), then 2026-09-30 (agents in production). One proposed amendment to the Project OT thread — add "any account of why 19 May was reversed" to the check criteria, ranked at the top, because it decides the case's classification and the existing criteria would all be satisfied without it.
Candidates: none added. DORA/Google Cloud was considered as a source candidate and deliberately not logged — it is a research programme to be read, not an outlet to be monitored, and this ledger is for things worth checking repeatedly. Recorded so a later run does not re-decide it.
Exploratory query — the gap this repo named itself and never chased. Target: the 2026-09-18 Google-special-case challenge, which asserts the base holds zero SRE/DORA/error-budget/Forsgren literature and that the objection wins by default if Brandon publishes without it. First move was to test the asserted absence rather than trust it — grep across both repositories confirmed it, and confirmed neither query log has ever searched the terms. Queries: DORA Accelerate State of DevOps research four key metrics organizational performance Forsgren Humble Kim empirical evidence, then error budget SLO adoption empirical study organizations why error budgets fail adopted shared reliability target development operations separate incentives.
Yield: honest partial, one line of it load-bearing, and nothing ingested. Confirmed the programme's existence, scale and central claim, and that DORA added Reliability as a fifth outcome measure in 2022, assessed through SLOs and error budgets — the exact mechanism Brandon's Google case rests on, which raises the cost of not citing it. Null on the thing that would decide the challenge: nothing found measures the installation of a shared interface measure between two parties with separate leadership and separate metrics, against an outcome — so Stafford's timing conjecture is unchecked by anything published, which is an opportunity rather than a hole. One figure was surfaced and deliberately not used: a secondary source attributes "only 20% of SREs regularly use error budgets" to a 2021 SRE report; the primary's summary does not carry it and the 2025 edition's PDF is not machine-readable here. If it verifies it is the strongest single piece of evidence for the objection — the mechanism does not spread even inside the discipline that invented it. Recorded as a named chase, not a finding. Do not re-run this as a general web search; the next move is reading Accelerate and the current DORA report.
Verification pass (Stafford's lane, and the substantive work today): IFS WP 26/23 read in full from the primary PDF — all scan figures confirmed, mechanism section contradicts the scan's mapping. Reuters/BNN report re-read for chronology — causal claim not supported. Both corrections are in Routing.
Routing
Strategy — the claim to make, and the one to drop. The supported version of today's flagship is that the organisation's own evidence does not move the layer that has to execute, and information the organisation does not control does. That is fully carried by the experiment's asymmetry and is Brandon's argument. The version that must not ship is "managers protected their own headcount" — the paper's mediation points at revised beliefs about the technology, the staffing effect is unexplained by anything measured, and the authors state that "employment and AI adoption are not intrinsically in tension within a firm or business unit."
Market — asset 2 gains its first completed instance. Project OT is the delayering wave running start to finish with internal numbers and a reversal, and it distinguishes the essay's two rival mechanisms in one company. Amendment filed. Third consecutive amendment without the draft moving.
Governance — DORA is a two-document reading assignment, not a search. Accelerate and the current DORA report convert the Google-special-case challenge from CONTESTED-and-exposed to an argument Brandon wins, and the gap they leave — nobody measures installing a shared interface measure against an outcome — is his to own.
Personal — the flattering error is the one this system will keep making. Today it withheld a speculative Incentive tag from Meta correctly, and granted a speculative Commitment exemption to Meta in the same file. Both were judgment calls about unmeasured things; only the unflattering one got caught. The check that would have caught the second is the same one that caught the first, applied in the other direction.
PROPOSED-UPDATE 1 — knowledge/paper-2-case-file.md (research-evidence)
Anchor: the 2026-09-20 Merritt/Dominski/Hoy/Lee entry, appended to its Limits paragraph.
Mechanism bound, added 2026-09-20 (Stafford, from the primary PDF §4.3). The paper decomposes its own effect and the decomposition does not support an incentive reading of the adoption result. Roughly 40% of the −0.403 SD effect on AI adoption runs through the AI-benefits index and ~30% through the AI-risks index (≈70% mediated; ~30% unexplained); for AI advocacy (−0.435 SD) the figures are ~40% and ~42% (≈82% mediated). The measured mechanism is belief revision about the technology, not the manager's stake. The staffing effect is the reverse case and is the one an incentive account would need: under 16% via benefits and under 6% via risks, with the authors writing that "the reduction in staffing intentions cannot be explained by any of our measured belief channels, suggesting that the labor treatment may trigger a more general retrenchment that extends beyond attitudes toward AI itself." The authors also state directly that "employment and AI adoption are not intrinsically in tension within a firm or business unit." What supports the Incentive Fragmentation tag is the asymmetry — both arms delivered information and only the one touching the manager's own unit moved anything — and not the mechanism, which the paper measured and attributed elsewhere. Claim 3 may cite the asymmetry. It may not claim managers protected their own headcount, scorecard or span: that is unmeasured here and is the reading the authors decline.
PROPOSED-UPDATE 2 — knowledge/emerging-patterns.md (research-evidence)
Anchor: The Migrating Bottleneck, the 2026-09-20 update block.
Scope correction, 2026-09-20 (Stafford). Recurrence count reverts to 4. This pattern's scope condition is in its title — AI Productivity Without Organizational Redesign — and its statement is that productivity rose "without redesigning the management layer." Meta was added as member five on the explicit grounds that it redesigned, which cannot make it an instance of the claim. Precedent is this file's own, three times: NBER WP 35275 held at 4 for measuring no structure; Bartolucci & Vivo held at 4 for measuring nothing; Wadi & Ma declined as the Defensive Gate's second member on the same morning. The Meta write-up is retained in full as a named boundary case, not deleted — it is the pattern's most valuable attachment. What it evidences is a stronger and distinct proposition that should be tracked separately: redesign specified against the tool's capability rather than against the constraint relocates the bottleneck rather than removing it. Single source; not a pattern; second-member bar — a named organisation that restructured around AI capability and reported a downstream measure (incidents, rework, cycle time, escaped defects) moving against it. All sourcing discipline recorded with the original member travels unchanged.
PROPOSED-UPDATE 3 — knowledge-base/reuters-meta-project-ot-ai-native-restructuring-2026-08.md (research-evidence)
Anchor: replace the Summary's sentence "The reversal came as Meta's own internal data raised questions about whether the technology was working."
The report gives no reason for the reversal. Its sentence is "on the night of May 19, just hours before the first layoff wave, Zuckerberg blinked." It names two contemporaneous conditions, in this order: "By then, Meta employees were in open revolt, convinced that the company's AI transformation initiatives were partly aimed at replacing them. Internal data was also suggesting that autonomous AI 'agent' technology at the heart of the strategy was failing to deliver hoped-for productivity gains." The quantified internal figures post-date the decision: the 220%/36% split comes from a Bosworth post the report dates to "early June", roughly two weeks after 19 May; the incident and firefighting figures carry no date. Cite the reversal as unexplained. Two candidate causes — workforce revolt and internal data — imply opposite readings (a workforce supplying the brake the governance layer did not, versus a company measuring well and correcting), and neither is established.
PROPOSED-UPDATE 4 — same KB file, Four Forces Intersection
Anchor: the paragraph beginning "Commitment was considered and not tagged."
Amended 2026-09-20 (Stafford). The decision not to tag Commitment stands. The reason given does not — "cancelling on the strength of internal metrics" asserts a cause the report does not supply and cites figures written two weeks after the decision. Record as unresolved: no Commitment finding is available either way until someone reports why 19 May was reversed. A framework that refuses speculative tags must also refuse speculative exemptions; Incentive Fragmentation was declined on this entry for exactly this reason, and the same test applied here fails it.
PROPOSED-UPDATE 5 — knowledge/story-threads.md (research-evidence)
Anchor: Project OT, the Next check criteria list.
Criterion (a′), added 2026-09-20 (Stafford), ranked above the existing (a): any account of why the second wave was cancelled on 19 May — a named source, a dated document, or a first-person statement. The report attributes no cause and the quantified internal figures post-date the decision by roughly two weeks. This is the fact that classifies the case — a company that measured well and corrected, or a company whose workforce supplied the brake its design left out — and every existing criterion could be satisfied without it. Note the failure mode: the October 2026–January 2027 window will generate commentary on the capability claim and is unlikely to generate anyone revisiting 19 May, so the absence should be recorded as a result rather than left open.
PROPOSED-UPDATE 6 — knowledge/emerging-patterns.md (research-evidence)
Anchor: The Defensive Gate, the pre-registered second-member bar.
Clause added 2026-09-20 (Stafford): the qualifying instrument must also measure the decision-maker's beliefs about the technology (perceived benefits and risks, or equivalent), not only behaviour against their own unit's stake. WP 26/23 measured belief channels and found they carry ~70–82% of the adoption and advocacy effect while explaining almost none of the staffing effect. A behavioural second member without belief measures will reproduce the same ambiguity at higher cost — it will show the pullback and remain unable to say whether the manager was protecting his position or had revised his estimate of the tool.
Source candidates
None proposed. DORA/Google Cloud considered and deliberately not logged — a research programme to read, not an outlet to monitor.
STAGE-CANDIDATE
None. Nothing was read end to end today that the scan does not already hold; the exploratory slot returned a verified gap and one unverified figure, and this base does not file either.