Stafford Brief — 2026-09-15

Delta layer on the 12:00 UTC scan (research-evidence briefs/2026-09-15.md). Tuesday: no falsification sweep, no weekly synthesis. One exploratory query, 2 KB files, 2 challenges logged, 1 asset amended.

Read the scan's brief first — it is the better piece of work today and this one does not repeat it. Its primary (Agachan et al., flat agent teams beat hierarchical ones at 51.5% fewer tokens) is real and its window is well-chosen. This brief is about the one thing it got wrong, which it got wrong in a way that is worth more to Brandon than the thing it got right.

One operational note before the content. Stafford did not run on 09-11, 09-12, 09-13 or 09-14. Nothing was lost from the scan — that pipeline ran all four days — but the judgment layer has debt, and one specific piece of it is cleared below: the HealthPlix counterargument from 09-13 should have been mirrored into the challenge ledger two days ago and wasn't. That is the second late mirror in a week, from the same cause. There is a cheap fix and it needs your call; it is in Routing.

📖 TODAY'S READ

Medicine already ran today's experiment. It got the opposite answer, and it got it in 2018.

The scan opened a new pattern this morning called The Binding-Gate Penalty: the idea that when you force someone to respond to a check, you make things worse, and the damage comes from the compulsion rather than from the check. Two sources, two days apart — the HealthPlix prescribing study on 09-13 and today's agent-team experiment. The scan was careful about it: it flagged the pattern as probation-only, and it wrote down in advance what would kill it. Any study finding that a mandatory check beat an advisory one, where the checker could genuinely check, counts against the pattern. And it said the third member had to be human beings, not software agents.

Both conditions were met eight years ago, in the literature the pipeline started reading two days ago, and the answer runs the other way. Powers and colleagues at the Yale Center for Medical Informatics reviewed 32 studies of hospital "hard stops" — alerts that flatly prevent a doctor from placing an order, or let them proceed only if a third party overrides. Four of those studies put hard stops head-to-head against soft stops, the kind a doctor can click past. Three of the four favoured the hard stop. Peer-reviewed, human subjects, published 2018.

So the pattern, as the scan worded it this morning, does not survive the day it was opened.

But the mechanism underneath it does survive, and this is the part that matters. A drug-interaction alert is the best case there is for a checker that can actually check: the computer compares the order against a known interaction table and is either right or wrong. That is exactly the condition under which today's agent paper predicts the mandatory version should win — its authors' rule is that a supervisor earns its cost when it can verify and becomes a liability when it can only opine. Medicine's hard stops win because the checker has something to check against. The agent Manager loses because it doesn't.

So the pattern is pitched at the wrong altitude, and one word fixes it. The variable is not compulsion. It is compulsion without a stated standard for what counts as acceptable. Say it that way and it survives all four studies — HealthPlix, the agent teams, Powers, and Strom below — and it is still falsifiable. Say it the way the scan said it this morning and a clinical informaticist kills it in one email.

Files: knowledge-base/powers-jamia-hard-stop-alerts-systematic-review-2018-11.md, knowledge-base/strom-nearly-hard-stop-alert-cpoe-rct-2010-09.md. Both verified from the PubMed and ClinicalTrials.gov APIs, not from summaries — every rendered page in this chase returned a cookie wall, a 403 or a 405.

🔭 THINKER TO WATCH

Richard Shiffman, Edward Melnick and the Yale Center for Medical Informatics — and the thing to watch is not them, it is how late Brandon is to their question.

Yesterday's scan named Sittig & Singh as a mature competing framework with twenty years of evidence and no awareness of Brandon's work, and called that a risk of being scooped by a tradition. Today is the same tradition and the risk is worse than being scooped. These are not people with a rival framework. These are people who have been measuring the exact thing Brandon's framework is now starting to assert, for three decades, with randomized trials and systematic reviews, in a field where getting it wrong kills someone.

Brandon's differentiation is real but it is narrower than it was yesterday, and it should be stated narrowly. Clinical informatics knows an enormous amount about one check inside one workflow. It has no theory of the organization the check sits in — no account of why the same alert works in one hospital and gets routed around in another, beyond "user involvement and iterative design," which is the paper's own conclusion and is a description rather than a mechanism. Five Breakpoints is a theory of that. What Brandon should stop claiming is that nobody is looking at verification. They have been looking at it since before he started, and they have numbers.

The practical move: cite them. An argument about AI oversight that leans on a thirty-year clinical evidence base is dramatically harder to dismiss than one that leans on a two-week-old preprint with 43 pairs. The evidence is free, it is public, and right now nobody writing about AI governance is using it.

🏢 CASE IN THE WILD

Two Philadelphia teaching hospitals switched off a safety control that was working. They were right to.

In 2006 the University of Pennsylvania ran a randomized trial across two academic medical centres. 1,981 clinicians. Half got a near-hard-stop alert when they tried to prescribe an antibiotic against a patient's active blood thinner — a genuinely dangerous combination. Half got standard practice.

The control worked, and worked big. Doctors did the right thing 57.2% of the time with the hard stop, against 13.5% without it — more than four times better, on the exact behaviour the alert was built to change.

They stopped the trial anyway. Four patients in the intervention arm had needed treatment delayed: two waiting on the antibiotic, two waiting on the blood thinner. The gate had no way to distinguish a careless order from an urgent one, so it blocked both.

The breakpoint is Technology Illusion, and this is the cleanest version of it in the base because nothing was broken. The alert was well-built. It fired correctly. It changed behaviour by a factor of four. The organizational reasoning around it was what failed: nobody had asked what happens to the patient who needs both drugs right now. Every number on the dashboard said this control was a success. The four harms were visible only because a randomized trial with outcome monitoring happened to be running underneath it.

The open question, and it is the one to put to a client. Every AI approval gate being sold this year is this architecture. How many of them have anything running underneath, watching for what the gate does outside its own success metric? Penn found out because they were running a trial. Nobody deploying an AI review gate in 2026 is running a trial.

⚙️ FOUR FORCES CONCLUSION

Capability, and today's evidence says something uncomfortable about it: the same control helps or harms depending on the organization, and the papers can't tell you which in advance.

Powers et al. do not conclude that hard stops work. They conclude that hard stops work when the organization can keep working on them — their words are "implemented judiciously with continuous user feedback informing rapid, iterative design," and the studies reporting good user experience are the ones citing heavy user involvement. Strom et al. end the same way: the lesson is "the importance of formal evaluation and monitoring for unintended consequences."

Both papers, independently, locate the deciding variable in the deploying organization rather than in the control. That is Capability in Brandon's sense, arrived at by two groups who have never heard of the Four Forces. It is also the strongest available answer to the objection that the framework just relabels things: here are two clinical papers whose own conclusions name an organizational capacity as the thing that separates a control that helps from a control that harms, and neither has a word for it.

What it changes: Claim 5 (readiness is the prerequisite, not the technology) has been carried in this base mostly by survey evidence about executives. It now has a thirty-year clinical literature saying the same thing about a specific control, with harms attached.

💭 OPEN QUESTION

Is a governance control that works on its own metric and harms the outcome it protects a failure in Brandon's framework — or is it invisible to it?

The framework diagnoses organizations where things went visibly wrong. Strom is a case where everything went visibly right. The alert worked. Compliance quadrupled. Had Penn deployed it instead of trialling it, it would be a case study in successful clinical decision support, and the four delayed patients would be four unremarkable bad days in a large hospital.

The framework currently has no way to flag that organization. Momentum Mirage is the closest fit — the appearance of progress without the movement — but that breakpoint is about progress that isn't real, and here the progress was real; it just wasn't the whole picture. Nothing in the five names the case where a control succeeds at what it measures and the harm lands where nobody looks.

The sharper form, which is a design question rather than a taxonomy one: if the answer to "what can your reviewer verify?" is a great deal — the checker is genuinely competent, the criterion is genuinely verifiable — does the framework have anything to say about whether the gate should bind? Today's evidence says that is a separate question with a separate answer, and that getting the first one right does not settle the second. Brandon's published position does not currently distinguish them.

🥊 CHALLENGE

Two logged today in knowledge/thesis-challenges.md, and the second is the serious one.

(1) The light-touch counterexample — HealthPlix — OPEN, mirrored two days late. The 09-13 scan filed it with a COUNTERARGUMENT tag and said it should appear in the paper in the critic's voice. Stated plainly: 1,701 doctors, 2.81 million prescriptions, a randomized field experiment where making a drug-interaction flag non-binding cut errors 8.6% — with no purpose work, no incentive redesign, no leadership involvement, no training. A pure interface change produced exactly the durable improvement the framework says requires organizational work first.

The defence is that solo practitioners have no organization to break: no handoffs, no second party, no strategy layer. That defence is correct and it is also dangerous, because the framework does not say so anywhere. If every successful light-touch intervention gets answered with "that setting was too simple," the framework has the unfalsifiability problem already logged against it. The fix costs nothing Brandon is actually claiming: state that Five Breakpoints is a theory of multi-party execution and makes no claim about single-actor decisions with unambiguous right answers. Written in advance it is a scope condition. Written in response to a critic it is a retreat.

(2) The binding gate is not the variable — OPEN, new. Beyond refuting the pattern, this one has a second half aimed at the framework. The scan's own Open Question named it: much of what this pipeline has spent eight weeks filing as organizations neglecting verification may have been organizations correctly dropping a check that couldn't check anything. If that's right, the verification gap isn't a breakpoint — it's triage, and the framework has been reading rational behaviour as pathology.

This is not hypothetical, and the case is already in the base. The 2× mandate firm (logged 09-10) cut human review of pull requests from 89% to 68%, doubled per-reviewer load, and showed no quality degradation on any metric anyone measured. That is currently filed as a separate objection. It and this one are the same objection and should be read together. What would defeat both: a single measured case where a review step with no stated criterion was removed and harm followed. The base holds none.

📈 PATTERN BUILDING

Not a new pattern — a correction to one that is eight hours old, plus the reason it happened.

The Binding-Gate Penalty should be restated as: a mandatory gate beats an advisory one where the gate holds a criterion it can actually check, and loses where it holds only judgement. Four members, two directions, and it survives all of them:

The generalization the scan wrote — "the cost of a check is not the check, it is the compulsion" — is contradicted by row one and complicated by row two. The restatement survives all four.

And the process lesson is worth more than the correction. The pipeline reached clinical informatics on 09-13 and opened a "novel" pattern on 09-15 about a question that field settled in 2018. The base's default move on a new mechanism is to look for the second instance in its own categories. The better first move is to ask which older field already ran the comparison — because a mature literature does not just add a third source, it tells you whether your variable is the right one. Logged as a standing rule in knowledge/query-log.md.

🧱 ASSET SUGGESTION

Asset 7 — "The Ladder With No Verification Rung" — needs one word changed before it ships, and the change makes it a much better essay. Full amendment in knowledge/asset-suggestions.md.

As proposed on 09-10, it argues that AI maturity ladders contain no rung for checking the machine's output, and that the absence is the failure. As written, the first reader who knows clinical informatics takes it apart: medicine built the rung, made it mandatory, and stopped a trial over it.

The fix: the missing rung is not review. It is a stated acceptance criterion. The ladders escalate authority — more agents, more delegation, more orchestration — and never once specify what a level's output has to satisfy to be accepted. That claim is true, checkable against the ladders themselves, and it gains a thirty-year evidence base instead of resting on an absence.

The opening writes itself: Medicine spent thirty years learning that a mandatory check is worth its cost only when the checker can actually check. The AI maturity ladders being sold this year skipped that entirely — they specify what the machine may do at each level, and never what its output has to satisfy.

This was already one of two assets in the ledger that ship without a decision from you. It still is. But don't draft the current framing.

🎓 SPARRING NOTE

Withheld — fourth consecutive withholding, and the cap is the reason. Two notes have been outstanding and unanswered since 08-31 and 09-01, both on growth edge #3. The cap is two. A fifth parked hook stacked on an unanswered stack is noise.

Recorded rather than issued, because it is the best hook growth edge #2 (the readiness instrument sketch) has had: today supplies a readiness question that is one sentence long, costs nothing to ask, and has a verifiable answer — for this review step, what can the reviewer check the output against? Every instrument the readiness literature offers is a multi-item survey with no demonstrated predictive validity. This one is a single question with a hard answer, and two clinical literatures suggest it predicts whether the control helps or harms. If an ex-ante readiness measure can be that cheap, the unfalsifiability challenge gets materially easier — and edge #2 has never had a candidate item this concrete.

Standing item for the first /coach session, now five hooks deep (09-03, 09-04, 09-05, 09-06, today). The notes are stale by waiting, not by content. Brandon: the backlog is the signal. One session clears it.

🔍 SOURCES & THREADS

Threads: none were due and none were chased. The scan advanced two (verification cost, team-agent interaction) and opened one (the supervisor pattern, next check 2026-10-15). The nearest due date is team-agent interaction on 2026-09-23. One criterion needs amending on the strength of today's find, and it is in Routing — the new supervisor thread's highest-value criterion is "a human-subject version varying whether a reviewer holds a verifiable criterion or only judgement," which Powers et al. partially supplies from 2018. The criterion should be narrowed rather than left to be met by something already in hand.

Candidates: one proposed, not added — Stafford can't write that file. The Yale Center for Medical Informatics group (Powers, Shiffman, Melnick, Hickner, Sharifi) is the outlet candidate today, and the promotion condition should be stated tightly, because the risk with a thirty-year literature is ingesting all of it: a second find only counts if it measures a verification step against an outcome, not if it describes alert design. JAMIA itself is the larger candidate and the same condition applies.

Exploratory query: spent, and it refuted a same-day pattern. Target was not a case-file gap — it was the falsification condition the scan pre-registered this morning and had no slot left to test. Coverage check first across all 531 KB files in both repos: two hits on hard stops, both observational, no head-to-head comparison anywhere in the base. Full entry in knowledge/query-log.md.

Provenance. Every rendered page in this chase failed — PubMed returned a cookie wall, Europe PMC and healthexec 403, AHRQ PSNet 405, ClinicalTrials.gov served a JS shell. Both primaries were read from the PubMed E-utilities and ClinicalTrials.gov v2 APIs, which return structured records and are the route to prefer for anything clinical from here on. Added as a standing note for the next session.

Routing

strategy

Brandon has been handed a thirty-year evidence base for an argument he is currently making from a two-week-old preprint. The scan's window piece (agent org charts, closes ~2026-10-06) is well-chosen and should still be written. But add one paragraph to it: medicine has been measuring mandatory-versus-advisory checks since the 1990s, and the answer is that it depends on whether the checker can check. That paragraph does three things — it makes the piece unkillable by the "that's just agents, not people" objection, it separates him from the dozen others who will write up the same preprint, and it costs him one paragraph. The clinical evidence is free, public, and nobody on the AI governance beat is using it.

governance

Refine the question the scan proposed this morning. It offered: for every human review step, what can the reviewer actually verify? Good question, incomplete. The two-part version is the diagnostic: (1) what can this reviewer check the output against? and (2) if the answer is "nothing specific," why is the step mandatory? Strom is the reason part two matters — a gate that can verify, that worked, and that still harmed patients, because nobody was measuring outside the gate's own success metric. Third question for a board, and it is the one nobody asks: what is running underneath your AI review gate, watching for what it does that its own metrics don't capture?

personal

The 09-13 counterargument sat unmirrored for two days because Stafford didn't run, and this is the second time in a week. The mirror rule assumes a daily Stafford run and fails silently without one. It is not a content problem — the scan caught both finds correctly and said so in its own briefs. Your call on the fix, below. Separately: the /coach backlog is now five parked hooks deep and two sparring notes have been unanswered since 08-31. Everything in that file is waiting on one session.

market

No window opened here. The scan's window is the day's window and its four conditions are correct — add nothing to it except the clinical paragraph above.

STAGE-CANDIDATE 1 — Powers et al. 2018

python3 build/add-research-json.py --json '{
  "title": "Efficacy and unintended consequences of hard-stop alerts in electronic health record systems: a systematic review",
  "publisher": "Journal of the American Medical Informatics Association (JAMIA)",
  "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC6915824/",
  "published_date": "2018-11-01",
  "credibility": "high",
  "instrument": {
    "data_source": "32 studies of computerized hard-stop alerts in healthcare settings, from biomedical and computer-science databases, grey literature, reference lists and prior reviews",
    "inference_method": "systematic review, narrative synthesis, NO meta-analysis (heterogeneity); 3 RCTs, 3 retrospective cohort, 18 pre-post, 5 time-series",
    "observation_window": "no date restriction; searches run 2017-03-09, updated 2017-12-09",
    "limits": "no study rated excellent quality; heterogeneity prevented pooling; likely publication bias favouring positive results; underreporting of health outcomes and implementation harms"
  },
  "breakpoints": [
    {"name": "technology_illusion", "rationale": "16 of 32 studies reported process outcomes only — half the literature on a mandatory control designed to prevent patient harm never measured patient harm, scoring the gate on whether it fired rather than on whether it helped."},
    {"name": "momentum_mirage", "rationale": "Eleven studies reported unintended consequences including the authors own named category avoidance of hard-stopped workflow — clinicians routed around the gate, so compliance statistics improved while the governed behaviour relocated."},
    {"name": "process_friction", "rationale": "Delay to care is named by the authors among recurring unintended consequences: a control inserted to stop one class of error introduces structural friction into the clinical path it sits on."}
  ],
  "forces": [
    {"name": "Capability", "rationale": "The authors conclusion makes effectiveness conditional on the deploying organizations capacity to iterate — implemented judiciously with continuous user feedback informing rapid, iterative design — and the studies reporting good user experience cite heavy user involvement."}
  ],
  "key_findings": [
    "Four studies directly compared hard-stop to soft-stop alerts; three of the four favoured the hard stop on the desired process outcome — a direct counter-example to the Binding-Gate Penalty pattern opened 2026-09-15",
    "Of 32 studies, 15 evaluated patient health or healthcare delivery outcomes and 11 showed improvement; 16 reported process outcomes only",
    "Eleven of 32 studies reported unintended consequences: avoidance of hard-stopped workflow, increased alert frequency, and delay to care",
    "Authors conclusion is conditional not directional: hard stops can be effective but must be implemented judiciously with continuous user feedback informing rapid, iterative design"
  ],
  "sample_size": 32
}'

STAGE-CANDIDATE 2 — Strom et al. 2010

python3 build/add-research-json.py --json '{
  "title": "Unintended effects of a computerized physician order entry nearly hard-stop alert to prevent a drug interaction: a randomized controlled trial",
  "publisher": "Archives of Internal Medicine",
  "url": "https://pubmed.ncbi.nlm.nih.gov/20876410/",
  "published_date": "2010-09-27",
  "credibility": "high",
  "instrument": {
    "data_source": "1,981 clinicians at 2 academic medical centers, Philadelphia PA (ClinicalTrials.gov NCT00870298 gives actual enrolment 1,971 — discrepancy recorded, not reconciled); 194 intervention-arm alerts, 148 control-arm alerts",
    "inference_method": "randomized controlled trial, parallel assignment, single-masked — a randomized manipulation of the coerciveness of a check, not an observational comparison",
    "observation_window": "2006-08-09 to 2007-02-13; TERMINATED EARLY, registry reason: four adverse events encountered",
    "limits": "harm finding is n=4 events and a stopping decision, not a powered estimate; exact content of the control-arm alert not established from sources read"
  },
  "breakpoints": [
    {"name": "technology_illusion", "rationale": "The alert performed exactly as designed against its own metric — desired response 57.2% vs 13.5%, a more than fourfold improvement — and still had to be switched off, because the harm arrived in a dimension the control did not model."},
    {"name": "momentum_mirage", "rationale": "Every process metric showed success while four patients had needed treatment delayed; the appearance of a working control and a working control were separated only by the presence of a randomized trial with outcome monitoring."},
    {"name": "process_friction", "rationale": "The intervention precipitated clinically important treatment delays in 4 patients who needed immediate drug therapy — a mandatory gate on an urgent clinical path shifted friction onto patients rather than onto the clinician whose behaviour it scored."}
  ],
  "forces": [
    {"name": "Capability", "rationale": "The authors close by making organizational evaluation capacity the lesson — the importance of formal evaluation and monitoring for unintended consequences of programmatic interventions — since what distinguished this deployment from a harmful one was the instrument watching outside the alerts own success metric."}
  ],
  "key_findings": [
    "Desired response (not reordering within 10 minutes) 57.2% (111/194) with the near-hard-stop vs 13.5% (20/148) control; adjusted odds ratio 0.12 (95% CI 0.045-0.33)",
    "Trial terminated early after 4 unintended consequences in the intervention arm: treatment delayed for 2 patients on trimethoprim-sulfamethoxazole and 2 on warfarin",
    "The only randomized member of Powers et al 2018 four hard-versus-soft comparisons, and the one running against the mandatory gate",
    "Demonstrates Claim 4 experimentally in human subjects: a control succeeded on its own success metric while the outcome it existed to protect moved the other way"
  ],
  "sample_size": 1981
}'

PROPOSED-UPDATE 1 — knowledge/emerging-patterns.md, the Binding-Gate Penalty entry (opened 2026-09-15)

Amendment 2026-09-15 (Stafford, same day the pattern opened) — the pattern's own falsification condition was met, from 2018, and the pattern needs restating rather than closing.

The pattern was opened with the claim that "the cost of a check is not the check, it is the compulsion," and pre-registered two conditions: a third member must be human subjects, and any instrument finding a coercive check beating an advisory one where the reviewer can genuinely verify is a counter-example. Both are satisfied by Powers et al., JAMIA 2018;25(11):1556–1566 — a systematic review of 32 studies of clinical hard-stop alerts, of which four directly compared hard stops to overridable soft stops and three favoured the hard stop on the desired process outcome. Human subjects, peer-reviewed, eight years before the pattern was opened, in the clinical-informatics literature this base first touched on 2026-09-13.

The mechanism survives; the generalization does not. A drug-interaction alert is the paradigm case of a checker that can verify — deterministic against a known interaction table — which is exactly where Agachan et al.'s verify-versus-opine rule predicts the mandatory version should win. Restate the pattern as: a mandatory gate outperforms an advisory one where the gate holds a criterion it can check, and underperforms it where the gate holds only judgement. That formulation survives all four members and stays falsifiable.

Members, with direction: Powers 2018 (mandatory wins, 3 of 4, verifiable criterion) · Strom 2010 (mandatory wins on behaviour, harms patients — the limiting case: verifiable criterion, but the gate could not see urgency, and nobody was measuring outside the gate's own metric) · HealthPlix 09-13 (advisory wins, −8.6% errors) · Agachan 09-15 (advisory wins, −51.5% tokens, no verifiable criterion available).

Status: NOT escalated, and the count should not be read as agreement — the members disagree by design, which is the point. Four disciplines carry forward. Powers reports no excellent-quality study among the 32, heterogeneity preventing meta-analysis, and likely publication bias favouring positive results; four head-to-head studies is a thin base and does not establish the reverse generalization either. Strom's harm result is four events and a stopping decision, not a powered estimate. And the honest framing of the whole pattern has changed: this is not a novel two-source pattern, it is a late entry into a contested question with a thirty-year evidence base.

Standing rule proposed from this episode: before opening a pattern about a mechanism, check whether an older field has already run the comparison. A mature literature does not add a third source — it tells you whether your variable is the right one.

PROPOSED-UPDATE 2 — knowledge/story-threads.md, The supervisor pattern (opened 2026-09-15)

Amended 2026-09-15 (Stafford): the thread's stated highest-value criterion — "a human-subject version varying whether a reviewer holds a verifiable criterion or only judgement" — is partially met on the day it was written, by Powers et al. 2018 (4 head-to-head hard-vs-soft comparisons, 3 favouring the coercive gate) and Strom et al. 2010 (the randomized one, stopped early after 4 treatment delays). Narrow the criterion so it is not met by something already in hand: a human-subject study in which the reviewer's verification capability is itself the manipulated variable — same task, same gate, reviewers who can check against a stated criterion versus reviewers who can only judge. Neither clinical study varies that; both hold verification capability fixed and vary only coerciveness. That study does not exist in this base and would be the single most valuable acquisition for this thread. Next check unchanged at 2026-10-15.

PROPOSED-UPDATE 3 — knowledge/story-threads.md, Does anyone design for verification cost — redirect, not an advance

Note 2026-09-15 (Stafford): criterion (1) — a named organization publishing a verification standard — remains unmet, and across the fourteen instruments in six functions it has checked, the search has been looking in the wrong field. The thread has scanned enterprise, finance, engineering and consulting sources for a published review standard. Clinical informatics publishes them, and has for three decades — alert governance, override review, sampling of flagged orders. This is not scored as meeting criterion (1): those standards govern rule-based decision support, not AI output review, and the distinction is the thread's whole point. But the criterion should carry an explicit note that the nearest existing analogues are clinical, so the next check looks there first and scores the gap precisely rather than recording another null. Proposed criterion (6): any organization publishing a review standard for generative AI output that specifies what the reviewer checks against — which is the clinical convention applied to a setting where the criterion is not deterministic, and is the thing nobody has done.

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

Yale Center for Medical Informatics (Powers, Shiffman, Melnick, Hickner, Sharifi) — first seen 2026-09-15 via the hard-stop systematic review (JAMIA 2018;25(11):1556–1566), which refuted the Binding-Gate Penalty pattern on the day it opened. Signal count 1. Promotion condition, stated tightly because the risk with a thirty-year literature is ingesting all of it: a second find counts only if it measures a verification step against an outcome — not if it describes alert design, taxonomy or user experience. Same condition applies to JAMIA as an outlet candidate (signal count 1). 60-day expiry 2026-11-14.

PROPOSED-UPDATE 5 — process, needs Brandon's decision

The COUNTERARGUMENT mirror rule fails silently when Stafford doesn't run. Two of the last three mirrors were late for this reason (09-07 find mirrored 09-10; 09-13 find mirrored today). The scan cannot write to thesis-challenges.md — that file is Stafford-owned — but its brief could carry a standing MIRROR-DUE line whenever it files a COUNTERARGUMENT-tagged entry, so the debt appears in the scan's own output and any later Stafford run picks it up from the brief rather than by diffing four days of commits. One line in the scan's brief composition spec. Brandon's call, since it is a change to research-evidence.