Stafford

← 2026-09-26 · all briefs · 2026-09-28 →

Stafford Brief — 2026-09-27

Sunday. Delta layer on the 12:00 UTC scan (2 staged, 11 skipped) — read that brief first; this one does not repeat it.

Three things, and the first two are corrections to the scan's best work today rather than additions to it.

One. The scan's closing decision — should Brandon re-audit every prevalence number in the base that came from public reporting? — rests on a sentence that misclassifies the largest part of its own evidence base. Consulting surveys are not the blind channel. They are the same kind of instrument as the half of today's paper that did see organizational causes. I counted the base: the genuinely affected slice is roughly a fifth of it, not "a large share" — which turns an expensive open decision into a small, specific one.

Two. The scan opened a new pattern at one source and was right to refuse to inflate it. But the pattern as written runs two different mechanisms together, and the second one already has two members — both found by this repo, both verified from primary texts, on 23 and 24 September. The scan cannot see them because the channel that would carry them to it has never been read by anything. That is now distorting the shared evidence base, not just delaying it.

Three. Today's retrieval failed, and the failure is worth more than the thing I was retrieving. The organizational data inside the best instrument in this base is the least reachable part of it — not paywalled, exactly: bot-blocked.

📖 TODAY'S READ

Somebody finally measured what the "a human reviews it" safeguard is actually being asked to do — and in a quarter of cases, the thing it has to catch is a passage that reads perfectly well.

"One note in three: a verified census of three deployed AI scribes, and the instrument that counted it" — Fox, Markham, Lail & Karotsieris; arXiv:2608.31017, 31 August 2026. Not peer reviewed, no venue named, author affiliations not established and deliberately not guessed here.

What they did, in plain terms. Ambient AI scribes listen to a doctor's appointment and write the clinical note. The entire safety story for these products is one sentence, which the paper quotes as its premise: the clinician signs every note. The authors ran three commercial scribes over the same 142 consultations — UK primary care and US ambulatory appointments, plus some written test scenarios — producing 565 notes, and then went looking for errors.

What they found. About one note in three contains a real error (31%). Strip out two error types that only happened because the products weren't connected to a patient record in this test — invented names and dates — and the honest figure, which is the authors' own, is roughly one note in four (24.8%). The errors cluster in allergies and medications, and in one category that should stop Brandon mid-sentence: a physical examination written into the note for a telephone appointment, where no examination can have happened. A further type came from other researchers' error lists: a treatment the doctor decided against, recorded as treatment given.

Why that last detail is the whole value of the paper, and it is the sentence to remember. Those errors are not garbled text. A fabricated normal examination reads like a competent clinical note. So the control — a tired doctor skimming before signing — is being asked to catch the errors specifically designed by accident to survive skimming. This is Technology Illusion in its cleanest measured form: the product assumed an organizational condition (someone with the time and retained skill to re-derive the note against the appointment) and the organization was never asked to supply it, resource it, or test it.

What Brandon should take from it, in one sentence: he now has a number for the workload that every "human in the loop" claim in his base quietly assumes is free.

The bound, and it is load-bearing — get this wrong and the argument becomes a liability. This paper audited the drafts the scribes produced, not the notes doctors signed. It does not show doctors missing anything. It prices the control; it does not test it. "A third of AI clinical notes contain errors doctors sign off on" would be false, and quotable against him. "The signature is being asked to catch a verified error in roughly one note in four, including fabricated examinations that read as normal — and nobody has measured whether it does" is exactly right, and the unmeasured part is the strongest half.

Technical detail: errors were found by twelve language-model passes and judged by an adversarial panel of two models from different families, each told to knock down what it could; 618 of 13,678 candidates survived. Two clinicians checked blind samples and upheld 20 of 21 and 12 of 12 findings. So this is a model-run census with a 33-finding human spot-check, not a doctor-led chart review, and the 31% must never be cited as though clinicians counted it. File: knowledge-base/fox-markham-lail-karotsieris-scribe-note-error-census-2026-08.md.

🔭 THINKER TO WATCH

Rebecca G. Mishuris and the Rotenstein author group — not for what they published, but for the table they have and nobody has seen.

Mishuris is Chief Health Information Officer at Mass General Brigham and senior author of the JAMA multisite AI-scribe study this base leans on more heavily than any other single instrument; the co-author list of 32 includes Julia Adler-Milstein and David W. Bates. (Mishuris's title is taken from the paper's own author line, read in full on 2026-09-24; the others' current affiliations are not established here.)

Why they are the people to watch right now. Their paper studied five named academic health systems — Mass General Brigham, Emory, UCSF, Yale New Haven, UC Davis — and reported one pooled answer for all five. The differences between those five organizations, which is the entire subject of Brandon's work, are documented in a supplementary table called eTable 1 and are absorbed out of the results by design. These authors are the only people in the world who can publish the number Brandon's thesis most needs, they can do it without a new study, and they are each other's colleagues. They have not.

How Brandon's position differs, and it is not a disagreement about facts. Their discipline treats the difference between five organizations' rollouts as noise to be averaged away in order to see the technology clearly — the paper says pooling makes the estimate "less likely to be biased by site-specific implementation details." Brandon's position is that the site-specific implementation details are the finding. Same data, opposite view of what counts as signal. That is a more useful disagreement to have in public than any argument about whether AI works.

🏢 CASE IN THE WILD

Five American health systems bought the same class of tool, and at least one of them decided who was allowed to have it by asking leadership to nominate people.

This is in the JAMA paper's Methods, verbatim: sites "defined the clinicians who were eligible for gaining access to the tool (eg, ranging from no restrictions on eligibility to restrictions based on clinical volume or eligibility based on leadership nomination)." Three different rules for the same decision, inside one study, across five peer institutions.

The breakpoint mechanism, named. The tool's entire justification is reducing documentation burden — the paperwork that follows doctors home. So there is one obviously correct allocation rule: give it to the people carrying the most paperwork. One site approximated that (eligibility by clinical volume). One declined to ration at all. And at least one allocated it by managerial nomination — a burden-relief tool handed out as a discretionary favour, disconnected from where the burden actually sat. That is Incentive Fragmentation as a live administrative act, not as a metaphor: the thing being optimized at the point of allocation was not the thing the purchase was justified by.

The open question, and it is answerable. The study found the benefit was concentrated in doctors who used the tool in half or more of their appointments — depth of use, not breadth of access, is where the gain lives. So: did the sites that rationed by nomination end up with fewer deep users than the site that rationed not at all? If yes, an organization's allocation rule set the return on its own purchase, and the paper's five sites are a natural experiment on Brandon's thesis that ran already and was reported as a robustness check. Nobody has asked. It requires no new data.

⚙️ FOUR FORCES CONCLUSION

Capability — and today is the day it stops being a stock and becomes a running bill.

Every Capability argument in this base has treated review capacity as something an organization either has or lacks. Today's paper says the bill is roughly a quarter of all outputs needing a genuine catch, in error classes that defeat skimming. The challenge logged on 25 September (from the Polish colonoscopy study, where doctors' unassisted detection rates fell after AI was introduced) says the person paying that bill gets worse at paying it the longer they pay.

Put the two together and Capability needs a second clause it does not currently have. Not only can the organization execute — but is the capability it is spending being replenished? The replenishment is concrete and costly: rotating people off AI-assisted work, maintaining unassisted practice, paying reviewers on something other than throughput.

What it means for the paper-2 argument. This is the sharpest available version of Technology Illusion — a deployment that improves the assisted number while degrading the unassisted one — and it is stated as arithmetic rather than as a worry. But it costs the prescriptive half of the framework, because "install a verification step" is now visibly a recurring expense with a decay schedule, and the Commitment force predicts exactly what an organization does with a recurring expense in a bad quarter. Brandon should publish that concession before a reviewer extracts it from him. It makes him look like someone who has thought about it, which he now has.

💭 OPEN QUESTION

The scan asked Brandon to decide whether to re-audit the whole back catalogue. Before he decides, the question needs fixing in two places — and once fixed, the honest answer is "no, do this much smaller thing instead."

The scan's reasoning was: today's paper shows that public-record sources under-report organizational causes of AI failure by roughly seventeen to one; a lot of this base is "consulting surveys, press releases and public reporting"; therefore a lot of this base is biased against Brandon's own explanation.

First correction: consulting surveys are on the other side of that finding, not this one. The paper compared a public incident database (organizational causes: 3%) against asking practitioners directly (organizational causes: 53%). A McKinsey or Deloitte survey of executives is the second kind of instrument — insider self-report with a disclosed frame. On the specific question of whether the organizational cause is visible, those surveys are the channel that can see it. They have other well-documented problems in this base — no stated frame, no sample, aggregators restating each other — but invisibility of the organizational cause is not one of them, and the scan's sentence sweeps them into the wrong bucket.

Second correction: the audit cannot be run on the publisher, because the outlet is a delivery van, not an instrument. A Fortune article reporting KPMG's Adaptability Index is an executive survey wearing a media byline. Sorting this base by outlet would mislabel most of it.

So I counted it, with the crudeness stated. Of 623 entries: about 68 are academic or primary sources; 20 are press releases; 171 come from consultancy or analyst domains directly. Of the remaining 397 that look like "media," 279 name a survey, index or research firm behind the number — media carrying an insider instrument. That leaves roughly 118 entries of genuine public-record or incident reporting, which with the press releases makes the affected slice about 138 of 623 — a fifth, not "a large share." (Method: keyword classification of outlet plus the text of findings and rationales. It is coarse and will misfile at the edges; it is an order-of-magnitude scoping to price a decision, not an audit result.)

And the reason I can only estimate it is itself the answer. The base has a field for exactly this — instrument, made mandatory for structural findings, recording data source, inference method and observation window. It is populated on 45 of 623 entries. Seven percent. Today's two staged entries both carry it, so the rule is working going forward and was never applied backwards.

Which makes the recommendation to Brandon concrete, and cheap. Don't re-audit 623 entries by publisher; it would mostly mislabel. Backfill the instrument field on the entries that are actually load-bearing — the ones cited in the six papers and the case files — and let the rest stay as they are with a standing caution. That is tens of entries, it produces the classification the re-audit was trying to approximate, and it is the same work whether or not today's paper had appeared.

The question I genuinely cannot answer, and it is the uncomfortable one. The scan's Open Question contained a sharper point than its recommendation, and it survives all of the above: a systematic bias in a channel is a reason to distrust that channel in both directions, including when it agrees with you. The 118 public-record entries under-report organizational causes. Fine — that cuts Brandon's way. But a fifth of his base being drawn from a channel demonstrably blind to his central variable means those entries cannot be cited for the prevalence of his mechanism either. Has he ever cited a public-reporting entry as evidence that breakpoints are common? If yes, that citation is now weaker in a way that has nothing to do with the bias running in his favour. I can't check his drafts against 118 entries today; he can check three drafts against one question.

🥊 CHALLENGE

Full statement and status: knowledge/thesis-challenges.md, The mitigation depreciates — OPEN, unchanged, and now half-armed rather than asserted.

The objection: Brandon diagnoses organizations for deploying technology on top of a missing human check, then prescribes installing the check — and the capability the check consists of is consumed by the thing it supervises. Logged 25 September on the Polish colonoscopy finding.

What today adds: the objection's cost premise was previously unquantified anywhere in this base. It now has a workload figure in a domain where this base also holds two deployment instruments — roughly one note in four needing a genuine catch, in classes that read as correct. High and adversarial workload (today) absorbed by a reviewer whose unassisted skill measurably declines with exposure (25 September). That is the strongest version of the prescriptive problem this base has assembled.

What today does not add, stated as firmly: nothing about whether the control works. Fox et al. audits drafts. The discriminating test — catch rate against deliberately seeded errors, measured against how long someone has been using the tool, with a comparison group that hasn't — remains unrun by anybody. The objection is better supplied, not closer to winning.

📈 PATTERN BUILDING

The pattern the scan opened this morning is two patterns, and the second one already has two members sitting in this repository.

The scan opened The Invisible Cause at one source and pre-registered what a second member would have to be: another instrument comparing an insider channel against a public channel on the same AI outcome, reporting divergence. That bar is correct and it is right that today's source is alone against it. I am not proposing an escalation. I am proposing a split, because one heading is currently doing two jobs.

Mechanism A — the event leaves no trace. An AI project killed by a scandal produces a news story; one killed because its sponsor moved on produces nothing. The public record does not under-weight the organizational cause; it cannot observe it. One source. The scan's pre-registered bar applies to this and should stay exactly as written.

Mechanism B — the trace exists, was collected, and the discipline routes it out at the interpretation step. Two members, both verified from primary texts by this repo, neither visible to the scan:

  1. Rotenstein et al., JAMA, read in full 2026-09-24. Five named health systems whose access rules differed. The between-organization differences were computed at least five times (a leave-one-site-out check) and reported as a robustness note. The paper's own justification for pooling: pooled estimates are "less likely to be biased by site-specific implementation details." The organizational variable was in hand and was classified as bias.
  2. Ohde et al., npj Digital Medicine, read 2026-09-23. Compiles the same tool class across three named health systems producing wildly different results — Mass General Brigham saved 5.6 minutes per appointment, Permanente 18 seconds, Intermountain nothing significant, an almost nineteen-fold spread. The authors attribute all of it to "study design–related factors" and name no organizational variable at all.

Why the split matters rather than being tidiness. Mechanism A is bad news that flatters Brandon: the evidence is missing, so his explanation is under-counted, and there is nothing to do but prefer better channels. Mechanism B is the publishable one, because it is fixable and because it is an indictment. The data exists. It is in print or in a supplement. The reason the organizational number is unavailable is not that nobody collected it — it is that the field's reporting convention treats organizational difference as contamination to be removed before the technology's true effect can be seen. That is Technology Illusion operating inside the research literature that studies technology. It is also the one version of this argument that names a remedy a reader can act on, which is what makes it an essay rather than a complaint.

And it needs one bound of its own. Both members are ambient clinical documentation. Two instruments in one technology class in one professional field is a pattern about medicine until a third arrives from somewhere else. Proposed as a 2-source pattern with that scope condition stated in the heading, not escalated.

The operational fact underneath this, which is now costing evidence rather than just time. Both members were found by this repo and written into briefs on 23 and 24 September. The channel that carries this repo's findings to the shared base — the PROPOSED-UPDATE blocks in these Routing sections — has never been read by anything. Audited 21 September: 81 blocks emitted since 30 August, zero applied, the receiving end never built. Re-checked today at ac5d497: still nothing. The shared repo's operating rules contain no reference to this channel; the most recent Stafford attribution in its knowledge files is still 30 August, the day of the split. The consequence stopped being procedural today: a correctly-disciplined scan opened a pattern at one source and pre-registered a bar for the second, while two candidate members sat eight days old in files it is not instructed to open. The one-line fix is in knowledge/pending-proposals.md and has been since 21 September.

🪟 WINDOW FLAG

None. Nothing today is time-priced. The scribe paper is four weeks old, the JAMA study five months, and the pattern split is argument value. Manufacturing a window here is the specific failure the 18 September ledger note flagged this file for.

🧱 ASSET SUGGESTION

Asset 5 ("The Human-in-the-Loop Control That Isn't") should gain a second half rather than a footnote, and the second half is where the numbers are. Full spec: knowledge/asset-suggestions.md, amendment 2026-09-27.

The pitch in two sentences. Asset 5 has argued since 3 September that "human in the loop" is asserted as a control without anyone specifying what the human is supposed to catch — a good argument that has never had a number. Ambient clinical scribes are the cleanest human-in-the-loop deployment in commercial existence (the control is literally one signature, and the vendors advertise it), and this base now holds three instruments on the same technology: what the deployment saves (Rotenstein), how wildly that saving varies between organizations (Ohde), and what the only control is asked to catch (Fox et al.). Nobody has put those on one page. That page is the essay.

Form: the existing Substack long-form plus the five-question audit card, with the scribe material as its worked example. Effort low — all four sources are in the base and the work is assembly. Risk moderate, and entirely in one place: the clinical setting invites overclaim, and one careless sentence about doctors signing off on errors would be both wrong and permanently quotable. Recommendation: second half, not replacement — the customer-support contrast is what makes the argument general; the scribe census is what makes it concrete.

🎓 SPARRING NOTE

None, and the cap check is the reason. Run per the procedural fix this file adopted on 25 September — count notes issued since the most recent Monday, from the file rather than from memory. Since Monday 21 September: three issued (09-22, 09-23, 09-24) against a cap of two. The cap is already breached for this week and six notes stand unanswered, the oldest from 31 August. Today's material would have made a legitimate note on growth edge #1; it stays parked. The recommendation from 25 September stands and is a /coach item, not another prompt: pick the 09-17 note, answer it, and retire the rest.

SOURCES & THREADS

Exploratory query — spent on retrieval, returned a dated null, and the null is the more useful result. The target was pre-registered by this repo's own log on 24 September as "the one remaining acquisition, named and narrow": eTable 1 in Supplement 1 of Rotenstein et al., the table the JAMA paper says holds the five health systems' institutional characteristics and rollout designs — the only organizational variables in the base's strongest instrument. Four routes, all documented, none successful. (1) The PMC supplement path that 24 September recorded as "an interstitial" — partly solved: the real path uses an internal numeric instance id (13152022) that must be scraped from the article page, and the PMC-ID-based path returns 404, which is why it read as an interstitial. (2) That correct path returns an NCBI proof-of-work bot challenge instead of the file. Not solved, deliberately — reverse-engineering a publisher's bot-mitigation is not a route this system takes. (3) Europe PMC's supplementary-files API, the route that has worked five times since 4 September: refused, "not open access" — the supplement sits outside the free subset even though the full text is free. (4) jamanetwork.com → 403. One confirming search found no legitimate mirror.

What the null establishes, and it is better than the table probably was. In the best deployment instrument this base holds, the outcome data is free and open — read in full on 24 September — and the only organizational variables in the paper are behind a non-open-access flag, a bot challenge and a publisher 403. The organizational half of the best instrument in this evidence base is the least reachable part of it. Access-route corrections now logged: the PMC instance/<id> path differs from the PMC ID; the proof-of-work gate is new since the 4 September access notes and will block every future PMC supplement retrieval by this method, while article full texts are unaffected. Remaining routes are a co-author request or an institutional-repository deposit — neither is a search. Full row in knowledge/query-log.md.

A search intermediary again asserted what eTable 1 contains. It only paraphrased the main text, added nothing verifiable, and was not treated as verification — the eighth time this base has declined an intermediary's claim about a methods section. The 24 September precedent (an intermediary inventing "site-specific estimates and random effects models" for this same supplement) is why.

One KB file, from the confirming search rather than from the target: knowledge-base/fox-markham-lail-karotsieris-scribe-note-error-census-2026-08.md. Judgment files updated: thesis-challenges.md (The mitigation depreciates, update), asset-suggestions.md (amendment to asset 5 — no new number, no renumbering), query-log.md. brandon-development.md deliberately untouched — no sparring note ran.

Threads: none due to this repo today. The scan reported the nearest next-check as Agents in production without a governance model, 30 September, and correctly ran no token search. I add no thread proposals beyond the pattern split below.

Source candidates: one, at one signal — see Routing.

Routing

STAGE-CANDIDATE — knowledge-base/fox-markham-lail-karotsieris-scribe-note-error-census-2026-08.md

Ready to run in research-evidence:

python3 build/add-research-json.py --json '{
  "title": "One note in three: a verified census of three deployed AI scribes, and the instrument that counted it",
  "publisher": "arXiv (preprint, not peer reviewed)",
  "url": "https://arxiv.org/abs/2608.31017",
  "published_date": "2026-08-31",
  "credibility": "moderate",
  "breakpoints": [
    {"name": "technology_illusion", "rationale": "The only safeguard in this deployment is the clinician signature, and this audit prices what that signature must absorb: a verified failure in 31.3% of notes (24.8% netting out the two classes a patient record would have prefilled), concentrated in allergy and medication data — a control whose required catch rate is a quarter of all outputs, asserted rather than measured."},
    {"name": "process_friction", "rationale": "The named failure classes are the ones plausibility-based reading cannot catch — history written up as examination on telephone consultations that can contain none, and a retracted treatment recorded as delivered care both produce clinically coherent prose — so catching them requires re-deriving the note against the encounter, reinstating the documentation work the deployment was bought to remove."}
  ],
  "forces": [
    {"name": "Capability", "rationale": "Supplies the workload denominator every human-in-the-loop argument in this base has lacked: the volume and class of error a reviewer must detect, though not a catch rate."}
  ],
  "key_findings": [
    "Three commercial ambient AI scribes audited on the same 142 consultations, 565 notes: 31.3% [27.0, 35.6] carry a verified failure; 24.8% [20.8, 29.0] after excluding the two classes a patient record would have prefilled (invented identity and dates) — the authors own adjusted figure",
    "Failures concentrate in allergy and medication information, invented patient identity, and history written up as examination on telephone consultations that can contain none",
    "One failure mode taken from published scribe-error taxonomies: a treatment the clinician retracts, recorded as delivered care",
    "Bound: the audit covers scribe DRAFTS, not signed notes — it prices the human-in-the-loop control and does not test whether clinicians caught anything"
  ],
  "sample_size": 565,
  "instrument": {
    "data_source_genre": "product audit — three commercial AI scribes run over the same 142 recorded UK primary-care and US ambulatory consultations plus authored scenarios; no product given a patient record",
    "inference_method": "Model-discovered and model-adjudicated census: twelve discovery passes proposed 13,678 candidate errors; 5,898 cleared an importance filter; an adversarial panel of two models from different families, each instructed to refute, left 618 survivors. Human validation limited to two blind disjoint samples — a physician author upheld 20 of 21, an independent clinician 12 of 12 (33 findings total). NOT a clinician chart review; the 31.3% must not be cited as clinician-counted.",
    "observation_window": "not stated in the abstract; submitted 2026-08-31",
    "caveats": "Unreviewed preprint, no venue in the comments field, author affiliations unestablished. No per-product breakdown is reported, so no vendor may be named as better or worse. Drafts, not signed notes."
  },
  "tags": ["paper-2", "practitioner", "COUNTERARGUMENT"]
}'

PROPOSED-UPDATE 1 — knowledge/emerging-patterns.md, the Invisible Cause entry

Split the heading into two mechanisms. Keep the existing entry as Mechanism A, unchanged, at one source with its pre-registered bar intact. Open Mechanism B as a separate 2-source pattern, not escalated, with the scope condition in the heading. Proposed text:

Pattern: The Interpretive Discard — the organizational variable was collected, and the discipline reported it as bias

First seen: 2026-09-23 (Ohde) | Recurrence count: 2 | NOT escalated | Scope condition, in the heading because it binds: both members are ambient clinical documentation. This is a pattern about one technology class in one professional field until a third member arrives from outside medicine. Sources: Ohde et al., npj Digital Medicine 2026;9 (doi:10.1038/s41746-026-02554-0) — knowledge-base/ohde-npj-ambient-scribe-scaling-heterogeneity-2026-03.md; Rotenstein et al., JAMA 335(16):1408–1417 (doi:10.1001/jama.2026.2253) — knowledge-base/rotenstein-jama-ai-scribes-full-text-site-level-verification-2026-04.md. What it says. Distinct from The Invisible Cause, which is about events that leave no record. Here the record exists and the organizational variable was collected — and is routed out at the interpretation step. Rotenstein names five health systems, documents that their access rules differed (from no restriction, to restriction by clinical volume, to "eligibility based on leadership nomination"), computes between-site variation at least five times in a leave-one-site-out check, absorbs it into clinician fixed effects, runs four person-level subgroup analyses and no organizational one, and justifies pooling because pooled estimates are "less likely to be biased by site-specific implementation details." Ohde compiles the same tool class across three named systems at roughly nineteen-fold dispersion (MGB −5.6 min/appointment; Permanente −18 s; Intermountain no significant gain) and attributes all of it to "study design–related factors," naming no organizational variable. Five Breakpoints angle. Technology Illusion, applied to the literature that studies technology: the object of measurement is taken to be the capability, and the organization enters the model as error. Unlike Mechanism A this is remediable without new fieldwork — Rotenstein's own co-authors could publish eTable 1 against the outcome table. Discipline. Ohde is a Perspective, not an empirical study, and its three deployment figures are second-hand against its own references 3/11/12, none of which has been read. Rotenstein is one instrument read twice (abstract entry plus full-text file) and must never be cited as two. Do not count this pipeline's own nulls here either — the same refusal the Invisible Cause entry correctly applied. What would escalate it: a third member outside ambient clinical documentation — any field where a multi-site AI study collects an organizational variable and reports it as a robustness check or as noise.

PROPOSED-UPDATE 2 — knowledge-base/rotenstein-jama-ai-scribes-clinician-time-multisite-2026-04.md, Sources & Confidence

Add the access record, so the next run does not re-spend a slot on it: eTable 1 in Supplement 1 remains unobtained after four routes on 2026-09-27 — the PMC-ID-based supplement path 404s; the correct articles/instance/13152022/bin/... path returns an NCBI proof-of-work challenge; Europe PMC's supplementaryFiles API refuses with "not open access"; jamanetwork.com returns 403. Its contents must not be characterised beyond the main text's own description of it. Remaining routes are a co-author request or an institutional-repository deposit.

PROPOSED-UPDATE 3 — knowledge/paper-2-case-file.md, Claim 4 / the gap ranking

A cheap, named, no-new-data test of the organizational-variance question exists inside an instrument the case file already cites. Rotenstein reports that the benefit concentrated in clinicians using the tool in ≥50% of visits, and separately that sites' eligibility rules ranged from unrestricted to clinical-volume-based to leadership nomination. Whether allocation rule predicts depth of use is therefore estimable from data already collected at five named sites by authors who are colleagues. Log as the highest-leverage unasked question currently identified, distinct from the standing "no instrument decomposes within- against between-firm variance" gap — this one needs no new instrument, only a subgroup the authors declined to run.

PROPOSED-UPDATE 4 — knowledge/source-candidates.md

Fox, Markham, Lail & Karotsieris (arXiv:2608.31017) — 1 signal, first seen 2026-09-27, surfaced by this repo's exploratory retrieval. Unaffiliated-as-far-as-established authors running adversarial multi-model audits of deployed commercial clinical AI. Pre-registered second-signal bar: a further audit by the same group whose unit of analysis is a deployment rather than a product — i.e. that names the organization or its review policy. A second product-comparison audit does not count. Expiry 2026-11-26 per the 60-day rule.

PROPOSED-UPDATE 5 — research-evidence/CLAUDE.md, §Post-Brief Filing

Unchanged from 2026-09-21 and now materially costing evidence rather than time. Add as step 0: Drain the proposal queue — read stafford-research/knowledge/pending-proposals.md (clone read-only), apply or explicitly decline each Tier-1 item, record the disposition. Today is the first day the gap produced a wrong state in the shared base rather than a missing one: a pattern was opened at one source, with a pre-registered bar for a second, while two candidate members for its sibling mechanism sat eight days old in files the scan is not instructed to open. Tier-1 item 10 is also still live — knowledge-base/ewens-giroud-corporate-hierarchy-2025-08.md still tells a reader that the authors' own October 2025 figures (3,100 firms / 7 million résumés / pharmaceutical) are fabrications of automated summaries. Verified still present today at ac5d497, two days after the reconciliation. Of the ten Tier-1 corrections, that is the one that actively instructs a wrong action.

Audience-tagged routing

Ship: see the first line of this run's final message.