Stafford Brief — 2026-09-25

Friday. Delta layer on the 12:00 research-evidence scan. Yesterday's run merged cleanly to main — no OPS flag. Weekly Synthesis at the end.

Today's scan told you not to quote a number. I went and read the paper, and the number is fine — it's the warning that was wrong, and it has been wrong in your base for six weeks.

The scan's Friday staleness check flagged Ewens and Giroud's Corporate Hierarchy — the study that firms flatten after adopting AI, which is load-bearing in your flattening argument and cited in two of your drafts. It found a revised version whose description didn't match what the base holds, and instructed: until this is reconciled, do not quote the 16-million-résumé and 2,500-firm figures. Correct instinct. The route it queued would not have worked, and the base's own six-week-old note about this paper says the opposite of the truth.

And the second thing, which cost more to get and is worth more. The scan named five unread citations and ranked one first: the study behind the claim that human oversight wears out through use. I read it. It is real, it is in one of the best medical journals in the world, and it measures three months — not the "extended use" the argument built on it needs.

What I am not repeating from the scan's brief: the Danish adoption instrument and its 23% don't-know figure, the Mitchell oversight paper, the PwC board contradiction, the new Don't-Know Denominator pattern, and the five thread advances. Read research-evidence/briefs/2026-09-25.md first. This is what changes because of it — plus two corrections to it.

📖 TODAY'S READ

The same study exists in two versions. Your drafts cite the old one. The base's warning label says the new one is a forgery.

The paper is Michael Ewens and Xavier Giroud, Corporate Hierarchy, National Bureau of Economic Research working paper 34162. It is the study that measures how many management layers American public companies have, and finds that companies cut layers after adopting AI. Your Flattening Bet essay cites it. So does the instrument audit.

I downloaded and read both versions in full today — 65 pages each, side by side. Here is what actually differs.

And here is what does not differ, which is the part your argument rests on. All four AI results are significant at the 10% level only in both versions — that is the weakest threshold researchers use, and the authors say why themselves, in words identical across the two texts: the test has little power because layer counts barely move within a company. The AI measure is still borrowed from another team's paper (Babina and colleagues) rather than built by these authors. There is still no measure of how many people each manager supervises, and still no measure of whether any of the flattening did the companies any good.

So the essay's argument survives untouched, and one sentence in its citations block did not. I corrected both drafts in place — drafts/the-flattening-bet.md and drafts/four-papers-one-method.md — citing the October revision with the superseded figures kept in an explanatory clause.

Now the thing worth more than the reconciliation. The base's existing file on this paper carries a warning, dated 14 August, headed "Citation warning, verified today." It says that accounts reporting 3,100 firms and 7 million résumés and pharmaceutical are fabrications produced by automated summaries, and that the true figures are 2,500 and 16 million and biotech. Those "fabrications" are the authors' own revision. What happened is that someone compared a summary of the new version against a full read of the old one, found a mismatch, and — reasonably, given that this base has caught seven genuine invented methodology details — called it a fake.

And the route the scan queued today would not have caught it either. The scan said to read nber.org/.../revisions/w34162.rev0.pdf. That URL still serves the August text — I downloaded it, the cover page says August 2025 and the abstract says 2,500 firms. The path is named "revisions" and delivers the original. The working route is the authors' own copy at columbia.edu/~xg2285/Hierarchy.pdf.

What you should take from it, in one sentence: the instrument is not weaker or stronger than you thought, but the base's method for spotting invented numbers needs one more step — check whether the paper was revised before calling the difference a lie — and that step is now hard rule 14 in my instructions.

One substantive gain, though, and it belongs in the essay. The revision covers 25% more companies using less than half the résumés — from roughly 6,400 résumés per company down to roughly 2,260. The measure works by watching people move between jobs and inferring the ladder from the pattern. A ladder inferred from a third as much evidence per company, producing bigger effects at the same weak significance level, is not a strengthening. That is a good paragraph for your instrument audit and a better illustration of the essay's own thesis than the line it replaces.

Full reconciliation, both texts, access record: knowledge-base/ewens-giroud-corporate-hierarchy-october-2025-revision-2025-10.md.

🔭 THINKER TO WATCH

Michał F Kamiński, Mette Kalager, Michael Bretthauer and Yuichi Mori — the group that produced the one hard number underneath an argument the whole governance field is about to start making.

Kamiński is at the Maria Skłodowska-Curie National Research Institute of Oncology in Warsaw; Kalager and Bretthauer run the Clinical Effectiveness Research Group at the University of Oslo; Mori is at Oslo and in Japan. They are senior authors on the Budzyń study below, and they are not AI researchers — they are the people who built the modern quality standards for cancer screening. That is exactly why they matter here.

Why them, this week. Today's scan featured Mitchell, Ghosh and Passi, whose argument is that human oversight of AI degrades through the use of AI. It is the sharpest governance argument in your base and it has no data of its own. Every empirical claim in it is borrowed. This group produced the one borrowed claim that is a real clinical outcome measured in real hospitals, and the scan ranked chasing it first of five.

Brandon's differentiation, and it is unusually clean. This group measures whether a screening programme is performing to standard. They are not making an organizational argument and would probably decline to make one. What they have is the measurement design your framework has been asking for all month and could not find: they compared the same clinicians doing the same procedure with the machine off, before and after the machine arrived. That is a before-and-after on the human capability, holding the task constant — which is precisely the shape of test the verification-cost thread has failed to find in nine instruments across six functions. Your contribution is not the finding. It is noticing that the design generalizes far outside endoscopy, and that no one outside medicine is running it.

Handle with care, and the care is the point of this entry. They are a signal-one entry and stay on probation. The promotion bar I am setting, deliberately narrow: a second signal counts only if this group — or anyone — publishes a measurement of review performance (catching an error the machine made) against tenure of AI use, with a comparison group. Another study of unassisted task performance does not count, and neither does commentary citing this one.

🏢 CASE IN THE WILD

Four hospitals in Poland turned on an AI polyp detector. Three months later their doctors were worse at finding polyps without it.

The organizations are four endoscopy centres taking part in the ACCEPT trial. They introduced AI polyp-detection tools at the end of 2021. From then on, whether any given colonoscopy got AI help was decided by the date — so there is a steady supply of examinations done without the machine, before and after it arrived.

The finding, in plain terms. Across 1,443 colonoscopies done without AI help, the rate at which doctors found pre-cancerous growths fell from 28.4% before the AI arrived to 22.4% after it — 6 patients in every 100 fewer getting a growth detected. The drop is statistically solid (p=0.0089) and survives adjusting for patient age and sex. (Technical detail: 795 examinations before, 648 after; 95% confidence interval on the drop −10.5 to −1.6 points; adjusted odds ratio 0.69.)

The breakpoint mechanism is Technology Illusion, and this is the cleanest arithmetic statement of it this base holds. The framework's definition is deploying technology on top of organizational conditions that have not been dealt with. Here the condition is a human skill the technology quietly spends. The organization sees the assisted number, which went up. It does not see the unassisted number, which went down. Both are outputs of the same deployment and only one of them appears on any dashboard. An organization measuring only the assisted figure would record an unambiguous success.

Where I have to push back on the argument this find is about to be used for — including my own. The study measures doctors doing the procedure themselves without help. That is not the same skill as checking whether the machine got it right, which is the capability every governance framework in your base depends on. They are related. They are not identical, and the difference is exactly the load-bearing one. The window is also three months, not the years of "extended use" that the Mitchell argument invokes. And there is no comparison group — no similar hospitals without AI measured over the same winter — so a general decline across that period is not ruled out. The authors write "might reduce" and "suggesting a negative effect." Quote their hedge, not a cleaned-up version of it.

The open question, and it is answerable. These four centres still have the data. Did the drop concentrate in particular doctors — and if so, in the ones who used the AI most? That single cross-tab would tell you whether this is a property of exposure (a dose-response curve, which would be devastating and general) or of four hospitals in one winter. Nobody has published it.

Full entry with every caveat: knowledge-base/budzyn-lancet-endoscopist-deskilling-ai-colonoscopy-2025-08.md.

⚙️ FOUR FORCES CONCLUSION

Capability — and my job today is to price the proposal the scan made this morning, not repeat it.

The scan proposed adding a clause to the Capability diagnostic: not only can the organization execute, but is the capability it is spending being replenished? I agree with the proposal and I am going to make it more expensive to hold.

Here is what the proposal now costs. Before today, "oversight capability depreciates" was an assertion in an unrefereed position paper. Now it has one measurement attached: six percentage points, over three months, on a related-but-different skill, in four hospitals, with no comparison group. That is enough to raise the question in front of a board and nowhere near enough to survive a hostile reviewer. If you put the replenishment clause into Paper 2 citing Budzyń as its evidence, the first serious critic will say — correctly — that you have generalized from a single before-and-after study of a task that isn't review, in a specialty that isn't theirs, over a quarter.

So state it as a prediction rather than a finding, and name the test. The version that holds: a capability spent through use will decline unless something replenishes it, and no organization in this evidence base measures the rate. The test that would settle it is already written down — the scan added it this morning as criterion (6) on the verification-cost thread: measure review quality (does the reviewer still catch a seeded error?) against how long that reviewer has been working with AI. That criterion, written this morning, now has its first real-world precedent in a design published thirteen months ago.

What it changes for you, concretely. The Capability force gains a clause you can defend today: capability has a rate of change, and nobody is measuring theirs. It does not yet gain the number that would let you say how fast.

💭 OPEN QUESTION

How many other load-bearing entries in this evidence base are frozen on a version the author has since replaced?

Today's find was not a fabricated figure or a bad tag. It was a stale version, carrying a warning label that made it look freshly verified. The entry said "verified today." It had been verified — against the wrong document.

Why this is worth your attention rather than mine alone. Look at what this base is mostly built from. Working papers: NBER, SSRN, IZA, arXiv. That is the genre of publication that gets revised silently and repeatedly. A journal article gets a version of record; a working paper gets quietly swapped, keeping the same number, the same URL and the same title. Today's case shows the failure is not even hypothetical: the publisher served the new abstract on the landing page and the old PDF at a path literally named "revisions," and the base took the mismatch for a forgery.

And notice the asymmetry that makes this expensive. A wrong tag is visible to anyone who reads the entry. A stale version is invisible by construction — the entry is internally consistent, correctly quoted, and wrong only relative to a document nobody has re-fetched. The staleness rule that caught this one fires at twelve months. Ewens and Giroud was revised at two months.

The observation that would settle it, and it is a morning's work, not a project. Take every working paper in this base that is load-bearing in a paper case file — my guess is fifteen to twenty-five entries. For each, fetch the publisher's current landing page and compare the abstract's headline numbers against what the base holds. You are not re-reading the papers. You are checking one field. My prediction, worth recording so it can be wrong: two to four will have moved, and at least one will be in a draft.

What I actually don't know, and it is the interesting half. Whether a revision that changes the sample by this much should count as the same instrument for pattern-counting purposes. The Résumé-Panel Consensus pattern holds nine members. If one of them quietly changed its firm count by 25% and its résumé count by more than half, is that the same member, a strengthened member, or a new one? I don't have a principled answer, and the answer determines whether an escalated pattern's count means anything.

🥊 CHALLENGE

The control you are about to prescribe is consumed by the thing it supervises. Logged at full strength, and it costs you something real.

The objection. Your framework diagnoses Technology Illusion and then prescribes, in substance, the human check the technology assumed was there. Asset 5 argues the human-in-the-loop control is nominal; asset 7 argues the missing verification rung should be added to the ladder. Mitchell, Ghosh and Passi say the rung cannot be added as a fixture, because what it is made of wears out through the exposure it exists to supervise. In their words: the cognitive capacities required for oversight "are themselves degraded by extended use of AI systems." If that is right, you are not naming a design gap to be closed once. You are selling a control with an unpriced depreciation schedule.

Why this is not easily dismissed. Today's Budzyń study puts a real clinical outcome under it — the first in this base where a professional capability measurably declined after an AI deployment, measured against patient harm rather than a questionnaire.

What limits it, stated by me rather than left for a reviewer. The construct is not review. The window is three months. There is no comparison group. Mitchell et al. is a position paper with no data. The objection is live, not won.

What it would take for it to win. Review quality — false-approval rate, or catch rate against deliberately seeded errors — measured against how long the reviewer has been working with AI, with an unexposed comparison group.

And here is why you should concede it early rather than fight it. Conceding costs you the sentence "install the verification rung" and buys you a much better one. Technology Illusion gets its sharpest possible form: a deployment that improves the assisted number while degrading the unassisted one, both produced by the same decision, only one of them on any dashboard. And the remedies that would replenish the capability — rotating people off assisted work, keeping unassisted practice alive, paying reviewers independently of how much they approve — are all recurring throughput costs, which is to say they are exactly what the Commitment force predicts an organization abandons in the first bad quarter. That is a falsifiable prediction nobody else in this market is making, and it is stronger than the prescription it replaces.

Logged: knowledge/thesis-challenges.md, "The mitigation depreciates," 2026-09-25.

🧱 ASSET SUGGESTION

No new asset. Two amendments, and the first one changes what asset 7 is for.

Asset 7 — "The Ladder With No Verification Rung." The essay currently says the ladders are missing a rung and implies you should add it. Today says the rung depreciates. Three changes: the diagnostic goes from two questions to three (who checks the output, how much AI-assisted work have they done this year, and when did they last do the work unassisted? — answerable from a rota, costs nothing); the ending changes from "add the rung" to "the rung is a running cost"; and the asset gains an explicit tie to Commitment, which it currently lacks. The stronger ending is the second one — and it rhymes with the asset's own existing evidence, that 73.8% of committed AI configuration files are never touched again after the day they are written. A rung that decays, in a population that never revisits what it wrote, is the same finding twice.

Asset 2 — "The Flattening Bet." Citation corrected in place, both drafts. This one would have shipped: a reader checking your instrument audit against the paper's current abstract would have found different numbers, and the audit is the essay's credibility.

Both specs: knowledge/asset-suggestions.md, 2026-09-25.

🎓 SPARRING NOTE

None issued, and the reason is a rule I broke rather than a quiet week.

The cap is two sparring notes per week. Three went out in the working week beginning 09-21 — on the 22nd, 23rd and 24th. Each run checked the cap and found it clear; each was counting from its own memory of the week rather than from the file. No single run did anything wrong and the rule still broke, which is the failure mode worth naming: a per-run check against a per-week limit drifts upward whenever runs are consecutive. Fixed: the check now counts entries in the file since the most recent Monday, before a note is drafted.

The consequence matters more than the bookkeeping. Six notes now stand unanswered — 08-31, 09-01, 09-17, 09-22, 09-23, 09-24. A prompt backlog is not a development plan. My recommendation is not another prompt: pick one — 09-17 has the most in it — answer it, and let me retire the rest. Today's material would have made a legitimate note on growth edge #1 twice over; both hooks are parked in the file rather than issued.

knowledge/brandon-development.md, 2026-09-25.

SOURCES & THREADS

Exploratory query: spent on a retrieval, for the second day running, and for the same reason. The scan's Friday staleness sweep flagged Ewens & Giroud and queued the reconciliation for "the next scan" while issuing a standing instruction not to quote figures that are live in one of Brandon's drafts. Operating rule 2 puts instrument verification in my lane; a live error in a drafted essay outranks a new search. Both versions downloaded and text-extracted in full. Do not re-run. Logged with the full side-by-side: knowledge/query-log.md, 2026-09-25.

Second slot — a citation chase, not the exploratory query, and recorded separately so the count stays honest. Budzyń et al., the scan's top-ranked unread chase of five. Route: web search to identify, then the Europe PMC REST API for the structured abstract — PubMed's HTML remains cookie-gated for this container, per the 09-04 note, and the API is not. thelancet.com full text remains unobtained (subscription). Everything above rests on a structured abstract read in full, and nothing about the paper's methods section is characterised beyond it.

Threads: 0 due, 0 advanced by me, 0 opened, 0 closed — and I checked rather than taking the scan's word. Nearest due is Agents in production without a governance model, 2026-09-30. One observation for the thread the scan did advance: its new criterion (6) on Does anyone design for verification cost? — review quality against tenure of AI use — already has a published precedent in a different field, thirteen months old. The criterion is not waiting on a design; it is waiting on someone to run the design outside medicine.

Source candidates: 1 added (proposed), 0 promoted, 0 expired. The Oslo Clinical Effectiveness Research Group / ACCEPT trial group, on probation, with the narrow bar stated in the Thinker section. Also worth recording as a route, not a source: the Europe PMC REST API is now the base's second successful retrieval of a gated clinical primary in two days.

Karpathy: no action needed from me, and the reason is itself a finding. The scan's synthesis this morning says the standing 403 "should be dispositioned" and has "now been carried for five weeks." I dispositioned it yesterday — diagnosed as a Cloudflare challenge, found the only working route reaches his abandoned blog, and proposed retirement from the daily list. That proposal sits in yesterday's Routing section where the scan cannot see it. This is the proposal-channel problem stated as a single concrete instance, and it is in the synthesis below.

Weekly Synthesis

1. Patterns — no promotions, and one bookkeeping question I cannot answer. The scan opened The Don't-Know Denominator at 2 and correctly declined to escalate it; I agree with both the opening and the refusal, and the reason it gives — that the two members' constructs are unified by the reader rather than by either author — is the right discipline. Nothing from my work today crossed a threshold. What today does raise, and what I am handing you unresolved, is the question in the Open Question above: the Résumé-Panel Consensus pattern holds nine members, and one of them has silently changed its sample by 25% of firms and more than half of its résumés. Is that the same member? An escalated pattern's count is load-bearing; I do not have a principled rule and would rather say so than invent one.

2. Staleness — the scan's rule caught a real one, and the rule's threshold is wrong. The 12-month sweep found Ewens & Giroud at thirteen months. The paper was revised at two months. A working-paper-heavy evidence base needs a version check that fires on the genre rather than on the calendar. Proposal, not executed: for any working paper that is load-bearing in a case file, re-fetch the publisher's landing page and compare headline sample figures quarterly, not annually — it is one field per entry, not a re-read. Prediction on record: two to four of fifteen to twenty-five entries have moved.

3. Paper evidence — the weakest claim got a mechanism and still has no evidence, and that distinction is the week's most important one. Claim 4 (the failure mode is harder to detect) is the weakest claim in the file. This week supplied a mechanism for detection failure (oversight degrades) and, today, the first real-world measurement adjacent to it — and still zero instruments measuring an organization failing to notice. The scan's exploratory sweep established that the market literature covers agents failing extensively and organizations not noticing not at all, and handed forward the right vocabulary: auditing, not engineering — false-approval rate, sampling, detection lag, control testing. I endorse that and add one thing: today's Budzyń design is the template. Measure the humans' unaided performance before and after, holding the task constant. In auditing that is a re-performance test and the profession has run it for decades. The next exploratory slot aimed at Claim 4 should search the audit literature for re-performance studies of AI-assisted reviewers, not the agent literature.

4. Challenges audit — nothing has yet crossed the 30-day line, and the first cohort arrives in three days. Oldest OPEN challenge is Unfalsifiability risk, logged 2026-08-29, 27 days old; it crosses on 2026-09-28, followed by The measurability confound (09-01) on 10-01. Pre-registering the disposition now, so it is not decided under the pressure of a deadline: an OPEN challenge at 30 days gets one of three outcomes — a named response plan, a demotion to a watch item, or an explicit statement that it is unanswerable with the evidence that exists. What it must not get is another month of carrying. One new challenge logged today (The mitigation depreciates); one substantive update (Coordination-primacy, whose supporting instrument is the paper reconciled above — bigger panel, bigger coefficients, same 10% significance, no change to the bar the objection must clear).

5. Windows — the scan executed three overdue decisions and left the one I asked for undone, and I am not going to ask a third time. Accenture recorded LOST, the three dead windows moved into a disposition table non-destructively, no new window opened for a seventh consecutive scan. All correct. The supervisor-pattern window (~2026-10-06) is still open and still blocked on a single amendment paragraph, for the tenth consecutive scan. My 09-24 recommendation was to close it and re-file its content as an asset. The scan's own escalated question — should this file keep opening windows — is a decision only Brandon can make, and the market routing has now carried it for seven weeks. My answer is unchanged and I will state it once more and then stop: stop opening windows; record time-sensitivity as a property of assets. Today's evidence is the argument — asset 7 got materially better from a citation chase nobody was in a hurry to run, and would have been worse under a deadline.

6. Sources hygiene — the scan's proposed rule split is right, and the deadline is real. The first candidate-expiry cohort — roughly two dozen entries first seen 08-09 to 08-20 — comes due 2026-10-08, thirteen days out. Several are annual report series, and an annual series cannot produce a second signal in sixty days no matter how good it is. The scan's proposal — 60 days for outlets and authors, one full publication cycle plus 30 days for annual and semiannual series — is mechanical, since every entry already records its cadence. Endorsed without amendment. It needs to be applied before 10-08, and applying it is a research-evidence write, not mine.

7. The proposal channel, week five, and this week it produced a measurable waste rather than a theoretical one. The audit stands unchanged: checked again today, no (Stafford) attribution anywhere in research-evidence/knowledge/ is newer than 2026-08-30, and research-evidence/CLAUDE.md still contains no reference to this repo or to the proposal format. The concrete instance this week: the scan's Friday synthesis asks for a Karpathy disposition it calls five weeks overdue. I delivered that disposition yesterday. It is in yesterday's Routing block, unreadable to the only party who can act on it. That is not a backlog any more; it is two systems independently spending time on the same open item. The one-line fix is unchanged and is in knowledge/pending-proposals.md: add a step 0 to research-evidence's post-brief filing that drains this queue. Nine Tier-1 items — each one a currently-live claim that is wrong — are now joined by a tenth today. Brandon, this is the single highest-leverage edit available in either repository and it is four lines of markdown in a file I cannot write to.

Routing

strategy — PROPOSED-UPDATE 1 → research-evidence/knowledge-base/ewens-giroud-corporate-hierarchy-2025-08.md, replacing the block headed "Citation warning, verified today." This is a Tier-1 correction: the warning currently tells any reader that the paper's current figures are fabrications. Proposed replacement:

Version note, replacing the 2026-08-14 citation warning (Stafford, 2026-09-25). The 2026-08-14 warning on this entry was a false positive and is withdrawn. It declared the figures 3,100 firms, 7 million resumes and pharmaceutical to be fabrications of automated summaries. They are the authors' own October 2025 revision. Both versions were downloaded and text-extracted in full on 2026-09-25. August 2025 (this entry's text): over 2,500 firms, 16 million resumes, biotech, Table 12 = −0.50/−0.62/−0.58/−0.13 on 4,606 obs across 1,250 firms, R2 0.99. October 2025 revision (current; columbia.edu/~xg2285/Hierarchy.pdf): over 3,100 firms, 7 million resumes, pharmaceutical, "operating asset volatility" replacing "more stable cash flows", Table 12 = −0.61/−0.75/−0.68/−0.14 (SEs 0.34/0.40/0.36/0.074) on 5,505 obs across 1,491 firms, Within R2 0.01, mean dep. var. 2.44 → 2.41. Unchanged in both: all four coefficients starred at the 10% level only; RHS measures taken from Babina, Fedyk, He & Hodson (2024); the authors' verbatim power caveat; the long-differences robustness check; no span-of-control measure; no outcome variable. Access warning, and it is the reason this went undetected: nber.org/system/files/working_papers/w34162/revisions/w34162.rev0.pdf still serves the August 2025 text despite the path name, and .../w34162/w34162.pdf returns HTTP 403; the NBER landing-page abstract serves the October revision. Cite the October revision; the working route is the author copy. Reconciliation: stafford-research/knowledge-base/ewens-giroud-corporate-hierarchy-october-2025-revision-2025-10.md.

strategy — PROPOSED-UPDATE 2 → research-evidence/knowledge/emerging-patterns.md, the Résumé-Panel Consensus pattern.

Member-version note added 2026-09-25 (Stafford). Ewens & Giroud (WP 34162) has a October 2025 revision the base had not read: firm coverage rises from over 2,500 to over 3,100 while the resume count falls from 16 million to 7 million — roughly 6,400 resumes per firm down to roughly 2,260, a ~65% thinning of the evidence under each firm for a measure that infers layers from a network of observed job transitions. AI coefficients rise ~20% and remain 10%-significant only. The pattern's shared-right-hand-side caution is untouched — Table 12 still uses Babina et al.'s measures — so these two members remain one instrument and must never be cited as independent confirmations. Open bookkeeping question this raises and does not answer: whether a revision changing a member's sample this much counts as the same member for the pattern's recurrence count. An ESCALATED pattern's count is load-bearing; the base should decide this rule rather than discover it.

strategy — PROPOSED-UPDATE 3 → research-evidence/knowledge/sources.md, NBER promotion note, and to the staleness procedure in §Weekly Synthesis.

Version-check procedure added 2026-09-25 (Stafford). The 12-month staleness rule found the Ewens & Giroud discrepancy at thirteen months; the paper was revised at two. Working papers (NBER, SSRN, IZA, arXiv) are revised silently under the same number, title and URL, and a publisher may serve the new abstract on its landing page while an older PDF remains at a path named "revisions." Proposed: for every working paper load-bearing in a paper case file, re-fetch the publisher's landing page quarterly and compare the abstract's headline sample figures against the KB entry — one field per entry, not a re-read. Estimated scope: 15–25 entries.

personal / governance — STAGE-CANDIDATE, Budzyń et al. Ready to run in research-evidence:

python3 build/add-research-json.py --json '{
  "title": "Endoscopist deskilling risk after exposure to artificial intelligence in colonoscopy: a multicentre, observational study",
  "publisher": "The Lancet Gastroenterology & Hepatology",
  "url": "https://pubmed.ncbi.nlm.nih.gov/40816301/",
  "published_date": "2025-08-12",
  "credibility": "high",
  "breakpoints": [
    {"name": "technology_illusion", "rationale": "The adenoma detection rate of unassisted colonoscopy fell from 28.4% to 22.4% in the three months after AI polyp-detection was introduced at the same four centres, so the deployment improved the assisted figure while degrading the unassisted one - a technology installed on top of a human capability it silently spends, with only one of the two numbers visible to the organization."}
  ],
  "forces": [
    {"name": "Capability", "rationale": "First real-world clinical measurement in this base of a professional capability declining as a function of exposure to the system meant to support it: adjusted odds ratio 0.69 (95% CI 0.53-0.89) for AI exposure on unassisted detection."}
  ],
  "key_findings": [
    "Unassisted adenoma detection rate fell from 28.4% (226 of 795) before AI introduction to 22.4% (145 of 648) after, absolute difference -6.0% (95% CI -10.5 to -1.6; p=0.0089)",
    "Multivariable logistic regression: AI exposure OR 0.69 (0.53-0.89); male patient sex 1.78 (1.38-2.30); patient age 60+ 3.60 (2.74-4.72)",
    "The post-exposure window is three months, not extended use; the design is retrospective before-and-after within the same four centres with no contemporaneous unexposed control group",
    "The outcome is unassisted task performance, NOT the capacity to review and overrule machine output - the two are related and must not be conflated when this is cited for oversight degradation",
    "Authors state the hedged conclusion verbatim: continuous exposure to AI might reduce the ADR of standard non-AI assisted colonoscopy, suggesting a negative effect on endoscopist behaviour"
  ],
  "sample_size": 1443,
  "instrument": "Retrospective observational study at 4 Polish endoscopy centres in the ACCEPT trial; 1,443 diagnostic colonoscopies performed WITHOUT AI assistance (795 pre-implementation, 648 post), 8 Sept 2021 to 9 March 2022, comparing the 3 months before and 3 months after AI polyp-detection introduction at end-2021; primary outcome adenoma detection rate; multivariable logistic regression. Number of endoscopists not reported in the structured abstract; clustering at clinician level not described. Structured abstract read in full via Europe PMC REST API (EXT_ID:40816301, resultType=core). FULL TEXT NOT READ - thelancet.com subscription-gated; do not characterise the methods section beyond the abstract."
}'

*personal / governance — PROPOSED-UPDATE 4 → research-evidence/knowledge/story-threads.md, Does anyone design for verification cost?, criterion (6).*

Precedent noted 2026-09-25 (Stafford). Criterion (6) — review quality against tenure of AI use — already has a published design in another field. Budzyń et al., Lancet Gastro & Hepatol 2025;10(10): same clinicians, same procedure, machine off, measured before and after the machine arrived; unassisted detection −6.0 points over three months across four centres. The criterion is therefore not waiting on a method to be invented; it is waiting on someone to run this method outside medicine. Two bounds to carry: the construct is unassisted task performance rather than review capability, and the window is three months.

personal / governance — PROPOSED-UPDATE 5 → research-evidence/knowledge/field-map.md. Add the Oslo Clinical Effectiveness Research Group / ACCEPT trial author group (Kamiński, Kalager, Bretthauer, Mori) at signal 1, first seen 2026-09-25 via the Mitchell citation chase, with the narrow promotion bar: a second signal counts only for a measurement of review performance against tenure of AI use with a comparison group, not another unassisted-performance study and not commentary citing this one.

governance — the diagnostic, and it is a rota question rather than a survey question. Your reviewers of AI output: how much of their own work this year did they do unassisted? No organization in this base can answer it, every organization can find out in an afternoon, and on today's evidence it is the input to a capability that is being spent without a meter. Pair it with the existing verification-rung questions and it is a three-question card that costs a client nothing to answer and tells you more than a readiness assessment.

market — no window, eighth consecutive brief, and I am not going to keep filing this line. The recommendation is in the synthesis; it needs a decision from Brandon, not a ninth restatement from me. The one thing that genuinely needs him this week is in §7: four lines of markdown in research-evidence/CLAUDE.md that would let three weeks of corrections reach the file they correct.