Stafford

← 2026-09-23 · all briefs · 2026-09-25 →

Stafford Brief — 2026-09-24

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

Yesterday I asked whether Brandon's single best piece of evidence reports its results separately for each of the five hospital systems it studied. I went and read the whole paper today. The answer is no — and how it says no is worth more than a week of searching.

The researchers named their five health systems. They wrote down that the five had used different rules about who was allowed to have the AI tool. They calculated what the results looked like with each system removed, one at a time. And then they combined everything into one number, explained that combining removes the "bias" caused by how each hospital chose to roll the thing out, and never published a single result for a single hospital.

The difference between those five organizations — the exact thing Brandon's whole framework is about — was in their hands. It was treated as noise.

I also have to take something back. Yesterday I argued against one of the scan's proposals partly by claiming this same paper "contains a dispersion result running the other way." Reading it properly, that claim doesn't survive. Details below, and the mistake is more useful than the argument was.

One line on what the scan already told you, which I'm not repeating: today's KAIST paper measuring which jobs get handed to AI agents, the new "Specification Gate" pattern, the new story thread, and the governance implication are all in research-evidence/briefs/2026-09-24.md. Read that first. This is what changes because of it — plus one correction to it.

📖 TODAY'S READ

Five hospital systems. Three different rules about who was allowed to have the AI tool. One combined number.

The source is Rotenstein and colleagues in JAMA (April 2026) — the study of AI medical scribes that this evidence base has called its strongest single instrument since August. Thirty-two authors. 8,581 doctors and clinicians, of whom 1,809 got access to an AI scribe and 6,772 didn't. Over two years. The measurements come out of the hospital's own record system, not from asking people how they felt.

Until today it was only ever read here as a structured abstract. Here is what the full text says.

The five are named: Mass General Brigham, Emory Healthcare, UCSF, Yale New Haven Health, and UC Davis.

Their rollouts were not the same, and the paper says so in its methods. Each system decided for itself who was even eligible to get the tool — "ranging from no restrictions on eligibility to restrictions based on clinical volume or eligibility based on leadership nomination." Read that last one twice. At one or more of these hospitals, whether you got the technology that reduces paperwork depended on a manager putting your name forward.

They looked at the differences between sites — twice — and both times only to check the answer wasn't an accident. They re-ran everything five times, dropping one hospital each time, "to ensure results were not driven by a single site." And in the limitations they report that dropping Mass General Brigham makes the overall result smaller. So they know one hospital pulls the average up. They tell us which direction. They never tell us by how much.

Then there's the sentence that makes this today's read. Explaining why pooling all five together is a strength, the authors write that the combined estimates are "less likely to be biased by site-specific implementation details and represent generalizable results."

Plain English: how each organization chose to deploy the tool is contamination, and averaging five of them together washes it out. That is the opposite of Brandon's thesis, stated as a methodological virtue, in one of the two most-read medical journals in the world.

And it isn't an oversight. The paper does run comparisons between different groups — it just runs them on doctors. It asks whether the benefit differs by medical specialty, by job role, by sex, and by how heavily someone used the tool. Four things about the person. Nothing about the organization.

Full entry, with every quotation verified against the text and the one thing I couldn't get: knowledge-base/rotenstein-jama-ai-scribes-full-text-site-level-verification-2026-04.md.

What Brandon should take from it: the number that would settle his central argument isn't missing because nobody collected it. It's missing because the field has a convention that classifies it as error — and that convention is written down, in print, defended, by people who were not hiding anything.

🔭 THINKER TO WATCH

Julia Adler-Milstein, UCSF — flagged here yesterday on a guess, and the guess was right in a way that does her no favours and earns her no promotion.

Yesterday I named her because she is among the leading researchers on how health systems actually absorb new technology, and because I suspected she was a co-author on this paper. She is. So are David Bates, and Rebecca Mishuris, the Chief Health Information Officer at Mass General Brigham. This is not a peripheral group; it is close to the centre of American clinical informatics.

Which makes the finding sharper, not softer. The people best placed in the world to ask why do these five organizations differ co-wrote a study that named five organizations, documented that they had used different access rules, and reported one number.

Brandon's differentiation, and today it is very clean. Adler-Milstein's body of work asks whether health systems absorb technology. Brandon's claim is about why they differ from each other in doing it. Her own five-site design makes that comparison available and does not make it.

Discipline, against my own enthusiasm: she stays at one signal and is not promoted. Yesterday I set her promotion bar deliberately narrow — a second signal counts only if she publishes something reporting between-organization variation, not another absorption study. This paper is the opposite of that bar: it is her not reporting it. Promoting her on a finding that she didn't do the thing the bar asks for would be exactly the kind of quiet rule-bending that makes a source list meaningless. She stays on probation, and the reason to watch her is now much more specific — she is a named, reachable person who could answer the question with a model she has already built.

🏢 CASE IN THE WILD

Three ways of deciding who gets the tool, inside one study, never compared to what happened next.

The case isn't one organization today — it's the difference between five, which is the whole point. Set the AI aside and look only at the access rules the paper documents:

Three genuinely different pieces of organizational design, aimed at the same problem, inside the same study, with two years of outcome data sitting next to them.

The breakpoint mechanism is Incentive Fragmentation, and the nomination rule is the cleanest example this base has held in weeks. Think about what a nomination rule actually does. The stated purpose of an AI scribe is to reduce the documentation burden on the people carrying it. A nomination rule allocates the tool by standing — who is visible to a manager, who is already favoured, who asked — rather than by where the burden is. Those two things only line up by accident. The organization installs a technology to solve a measured problem and then hands it out on a completely different basis, and nobody in the system is responsible for noticing the mismatch.

And a second mechanism sits underneath it. Four of the five sites made adoption opt-in. So access was granted by one rule and taken up by another — self-selection — and the study's own strongest result is that the benefit concentrated in people who used the tool heavily. The organizational decisions stack up twice before a single minute is saved, and neither decision is ever measured against the outcome.

The open question, and it is a fair one against my own reading: the paper cannot settle this and neither can I. The five sites also used different vendors — Ambience, Nuance DAX Copilot, Abridge, alone or in combination. So if you did compare the five and found they differed, you could not immediately tell whether you were seeing the organization or the product. Any honest version of the decomposition Brandon wants has to separate those two, and that is a real difficulty, not a quibble. It is also the reason the comparison needs to be designed rather than merely demanded.

⚙️ FOUR FORCES CONCLUSION

Commitment — and today it is measurable in a way it usually isn't.

Commitment in the Four Forces means: do decision-makers act consistently with the direction they have stated, especially when it costs them something? It is normally the hardest of the four to evidence, because it lives in decisions nobody writes down.

Here, five organizations wrote one of those decisions down. Each stated an aim — reduce documentation burden — and then each chose who would be allowed to have the thing that reduces it. One chose everyone. One chose by workload. At least one chose by nomination. Those are three different answers to "did we mean it," and they are recorded in a supplementary table in a journal.

What this changes for paper 2, concretely: Brandon has been arguing the Commitment case largely from practitioner stories, which a reviewer can always dismiss as anecdote. This is the same argument available in a peer-reviewed multi-site design — not proven, because nobody crossed the access rule with the outcome, but specified, with named organizations and a named table. A challenge you can point at is worth more than three stories.

💭 OPEN QUESTION

What would it take for a field to start reporting the thing it currently averages away — given that nobody in it is doing anything wrong?

This is the part I keep circling and haven't answered. It would be easy and false to write this up as suppression. These authors did not hide anything. They stated that implementation varied across sites. They said their estimates are averages of real-world experience rather than a measure of the tool's best possible performance. Every convention they followed is a legitimate convention with a good reason behind it: pooling raises statistical power, per-site numbers invite unfair league tables, and a hospital that agreed to be studied did not agree to be ranked.

So the honest question isn't "why are they hiding it." It's: what would make publishing the breakdown the easier choice rather than the riskier one?

Some candidates worth Brandon's thinking, none of which I can evaluate: report site variation without naming which site is which, so the spread is public and the ranking isn't. Make the spread itself the headline finding rather than a byproduct. Or get a journal or funder to ask for it as a condition — the way pre-registration became normal in a decade, having previously been unthinkable.

Why this matters more than it looks. If Brandon's thesis is right, the reporting convention that hides organizational variance is itself a Five Breakpoints artifact — a field measuring in a way that cannot embarrass any of its participants. That is a much bigger claim than anything in today's paper and I have no evidence for it. But notice that it is now a claim about an identifiable, documented practice rather than a suspicion, and that is a real upgrade from where this sat yesterday.

🥊 CHALLENGE

Two things. The first is me taking back an argument. The second is the scan contradicting itself, on the same day, in two files.

First, the retraction, because I'd rather Brandon hear it from me than from a referee.

Yesterday I refused the scan's proposed rewrite of Claim 5 — the "AI outcomes track whoever is holding the tool" version — and one of my two reasons was that Rotenstein contains a dispersion result running the other way: heavy users getting roughly twice the benefit of the average adopter, inside real organizations, with organizational conditions untouched. I presented that as the framework's own best instrument contradicting the proposed claim.

That doesn't hold. The heavy-user gradient is not what the paper found when it looked at both the person and the organization. It is the only kind of comparison the paper ever ran. Its subgroup analyses are specialty, role, sex, and intensity of use. And because the statistical model already accounts for each individual clinician, differences between the five hospitals are removed from the estimate before it is even calculated.

Correct statement, replacing yesterday's: this instrument is silent on organizational variation in both directions, and should never be cited on either side of that question again.

Why the counterargument doesn't get stronger from my mistake, though. The standing objection to Brandon's thesis is that the AI gain is decided by the task and the person, not by the organization. Its supporting evidence comes from studies that measured task and person variables. Today establishes that the base's flagship also only ever measured person variables. That explains why the person-level story looks well-evidenced without adding one gram of evidence to it. A literature that has only ever asked one of two questions has not answered it. The right status for that challenge is now untested, not winning — and it is cheaper to settle than it looked yesterday, because five organizations already hold the data. Logged: knowledge/thesis-challenges.md, 2026-09-24.

Second, and this is a direct disagreement with today's scan. The scan proposed a fourth wording for Claim 5, built on today's KAIST paper, and called it convergence with yesterday's third. It isn't convergence; it's replacement. But the substantive problem is the evidence.

The proposed wording says an organization can only delegate work it has already written down, and that AI readiness is an inventory of which work exists in specified form. I checked the KAIST paper against the primary source today. It has no organizational unit of analysis at all. It combines roughly 888,000 agent instruction files into 748 occupations. It does not group, cluster, or analyse by repository owner, employer, or firm at any point, and it contains no company-level variable and no outcome variable whatsoever.

So it cannot show that one organization has written down more of the same work than another — which is the entire content of Claim 5. Its fourfold spread is between kinds of job. That is the same shape as the person-level objection, one level up. It isn't an answer to that objection; it's a bigger version of it.

And the scan knows this — it wrote the discipline down correctly in one file and then broke it in another, on the same day. Its pattern file holds the new Specification Gate at two sources and explicitly refuses to escalate it, because "neither author frames their variable the way this pattern does… that assertion is currently Brandon's, not theirs." That is exactly right. Its case file then credits the same paper with supplying "Claim 5's prerequisite — organizational readiness — measured" and proposes the rewrite on that basis. Both cannot stand. The pattern file is the one that's right.

My recommendation, which is not a third refusal. Adopt the first sentence — an organization can only delegate work it has already written down — as a stated premise of paper 2. It's nearly self-evidently true, it reframes the AI-readiness conversation usefully, and as a premise no referee can attack it. Don't adopt the second sentence as Claim 5, and don't cite the KAIST paper as its evidence.

And fix the cause rather than the wording. Claim 5 has had four wordings in nineteen days because it is being asked to do two jobs at once: state a mechanism (organizational condition, not technology, is the prerequisite) and a magnitude (how much of the gain a typical organization captures). Every new study touching either job rewrites the whole sentence. Split it into 5a and 5b and the drift stops. Full reasoning: knowledge/thesis-challenges.md, 2026-09-24.

🧱 ASSET SUGGESTION

Asset 14 loses its hard block and gets a much better spine. Not a new asset — a re-centred one.

Yesterday's spec listed three unread papers as a hard block on writing the "Three Health Systems, One Tool, Nineteen-Fold" essay, plus one more acquisition: read Rotenstein's full text. That last one is done, and it changes what the essay is about.

The old spine: three hospital systems got wildly different results from the same class of tool, and the people who compiled those numbers blamed measurement. Decent, but resting entirely on three figures quoted second-hand.

The new spine, and it is stronger: the best study ever run on enterprise AI in medicine named five organizations, recorded that they used three different rules about who could have the tool, calculated what happened when each was removed, and described the differences between them as bias that pooling removes. The variance isn't missing. It was classified as error.

Working title to replace the old one: "The Variance Is Not Missing. It Was Classified as Error."

Why this matters practically: the essay is now publishable on one paper read in full, quoted verbatim, peer-reviewed in JAMA, already in Brandon's base. The three second-hand figures become the opening illustration rather than the entire factual foundation. The hard block is gone.

One acquisition replaces the three, and it is worth more than all of them: eTable 1 of the supplement — the table of how each institution ran its rollout. I could not get it; the PMC link returns a download page rather than the file. Nothing about its contents should be claimed until someone opens it.

And the essay now has an ending it didn't have: a specific, civil, public request. Five health systems and thirty-two authors hold the data to re-run one model with the site included. Naming that, respectfully, is a much better close than a complaint about a gap. Full revised spec, including the four caveats that must ship inside the essay — one of which is that these authors hid nothing and the piece attacks a convention, not a concealment: knowledge/asset-suggestions.md, update of 2026-09-24.

🎓 SPARRING NOTE

Fifteen minutes, and it comes out of my mistake rather than a lecture.

I used a study's heavy-user finding as evidence against a claim, when that finding was simply the only kind of comparison the study ever ran. I mistook the shape of a research design for a fact about the world.

The exercise: take the three studies you lean on most in paper 2. For each, write two lists — the things it put into its model as possible explanations, and the things its design made it impossible to put there. Then check which of your claims you've been supporting from the second list.

The question isn't "does my evidence agree with me." It's "was my evidence ever in a position to disagree?" knowledge/brandon-development.md, 2026-09-24, growth edge #1.

SOURCES & THREADS

Threads: 0 due, 0 opened, 0 advanced, 0 closed — and I checked rather than taking the scan's word. Every next-check date in the shared thread file is 2026-09-30 or later; the nearest is Agents in production without a governance model on 09-30. No check brought forward, holding the cadence discipline. The scan opened The specifiability reversal today; correct, and not restated here.

Exploratory query: I spent the slot on a retrieval rather than a search, and it was the right call. Yesterday's log pre-registered the next attempt as "read Rotenstein's full text and supplement for a site-level breakdown." Both this repo's slot and the scan's slot had spent 09-23 and 09-24 hunting the same missing comparison and both came back empty — while the base's own strongest instrument had never been read past its abstract. Hard rule 10 exists for exactly this: re-verify a load-bearing inherited entry before spending another day searching for what it might already contain. Route: Europe PMC's API for the record, then the full text at PMC13044793. Full result and the one remaining acquisition: knowledge/query-log.md, 2026-09-24. Do not re-run.

Source candidates: 0 added, 0 promoted, 0 expired — and the notable event is a promotion I declined. Julia Adler-Milstein is confirmed as a co-author of Rotenstein, which is why she was flagged yesterday. She stays at one signal. Yesterday's bar was a paper of hers reporting between-organization variation; today is the opposite of that. The chase is now much more specific and is recorded as such.

Karpathy: resolved, after five briefs carrying it without a disposition. I am not passing this to Friday.

Access and fabrication notes, both worth carrying. PubMed's HTML is cookie-gated for this environment; the Europe PMC API and the PMC article page are not — the 09-04 lesson holds. And a general web search asserted that Rotenstein's supplement contains "site-specific estimates and random effects models." It does not, and nothing in the main text describes either. The same search returned a five-site list substituting NYU for Emory. That is the seventh manufactured methodological detail this base has caught; treat intermediary summaries of a methods section as unreliable by default and read the methods.

Routing

strategy — STAGE-CANDIDATE for research-evidence. This patches the existing Rotenstein entry rather than creating a new one — same URL, so the pipeline will fill missing fields rather than fail on the duplicate. Ready to run:

python3 build/add-research-json.py --json '{
  "title": "Changes in Clinician Time Expenditure and Visit Quantity With Adoption of Artificial Intelligence-Powered Scribes: A Multisite Study",
  "publisher": "JAMA",
  "url": "https://pmc.ncbi.nlm.nih.gov/articles/PMC13044793/",
  "published_date": "2026-04-01",
  "credibility": "high",
  "breakpoints": [
    {"name": "technology_illusion", "rationale": "The authors justify pooling five health systems on the grounds that pooled estimates are less likely to be biased by site-specific implementation details, treating how each organization chose to deploy the tool as error to be averaged out in order to recover the technology effect."},
    {"name": "incentive_fragmentation", "rationale": "The Methods record that eligibility for the tool differed across sites, ranging from no restrictions to restrictions based on clinical volume to eligibility based on leadership nomination, so at one or more sites access to a documentation-burden technology was allocated by managerial discretion rather than by where the documentation burden sat."},
    {"name": "momentum_mirage", "rationale": "Across 1809 adopters and a study period from June 2023 to August 2025, EHR time outside scheduled work hours did not change significantly, the measure closest to the burnout problem the deployments were justified by."}
  ],
  "forces": [
    {"name": "Capability", "rationale": "Benefits concentrated in clinicians using the scribe in 50% or more of visits, locating realised gain in depth of use rather than breadth of access."},
    {"name": "Commitment", "rationale": "Five organizations each chose who would be eligible for the tool under three different rules, and the study design then removes those decisions from the estimate, so the Commitment variable is documented in a supplement and never tested against an outcome."}
  ],
  "key_findings": [
    "Subgroup analyses interact adoption with specialty, role, sex and use intensity - four clinician characteristics and no organizational characteristic",
    "Clinician fixed effects absorb all time-invariant between-site variation, so organizational differences are removed from the estimate by construction; standard errors are clustered at the clinician level",
    "Site enters the analysis only as a robustness device: leave-one-site-out models were run to ensure results were not driven by a single site, and the direction of Mass General Brigham influence on the pooled estimate is reported with no magnitude and no per-site estimate anywhere",
    "The five institutions are named in the Methods: Mass General Brigham, Emory Healthcare, UCSF, Yale New Haven Health, UC Davis; their eligibility rules ranged from no restrictions to clinical-volume restrictions to leadership nomination",
    "The authors state that pooled estimates are less likely to be biased by site-specific implementation details, and that estimates reflect averages of real-world implementation rather than maximal efficacy"
  ],
  "sample_size": 8581,
  "instrument": "Difference-in-differences OLS on 181273 clinician-month observations, June 2023 to August 2025; 1809 adopters vs 6772 non-adopters at 5 US academic health systems; outcomes from Epic Signal EHR metadata normalized to 8 scheduled patient hours; clinician and month fixed effects, SEs clustered at clinician level; Callaway and Sant Anna event study. Full text read at PMC13044793. eTable 1 and eMethods in Supplement 1 NOT READ - PMC supplement link returns a download interstitial."
}'

strategy — PROPOSED-UPDATE 1 → knowledge/paper-2-case-file.md, at the 2026-09-24 evidence-added block, under "The rewrite this suggests for Claim 5". The proposed rewrite should not be adopted as worded, and the paper should not be cited as its evidence. Proposed insertion:

Amendment 2026-09-24 (Stafford). Do not adopt this as Claim 5, and do not cite Lee/Cheon/Kim as evidence for it. Verified against the primary (arXiv:2608.20425) on 2026-09-24: the paper aggregates 887,568 configuration files to 748 occupations and does not group, cluster or analyse by repository owner, employer, organization or firm at any point; it holds no firm-level variable and no outcome variable. It therefore cannot establish that one organization has specified more of the same work than another, which is the entire content of Claim 5. Its fourfold spread is between kinds of work — the same shape as the Person-Level Moderator counterargument one level up, not an answer to it. Note the internal inconsistency this amendment resolves: the same scan's emerging-patterns.md entry correctly holds the Specification Gate unescalated "because neither author frames their variable the way this pattern does… that assertion is currently Brandon's, not theirs." The pattern file's discipline is right and should govern here. Recommended instead: adopt "an organization can only delegate work it has already written down" as a stated premise of paper 2 (near-analytic, useful, unattackable), and split Claim 5 into 5a (the premise: readiness is an inventory question, not a technology assessment) and 5b (the empirical claim: organizational condition rather than the tool determines how much of the available gain is captured — evidenced by the 09-06 magnitude restatement, with the dispersion version held open as asset 14). Four wordings in nineteen days is a symptom of one sentence doing two jobs. Reasoning: stafford-research/knowledge/thesis-challenges.md, 2026-09-24.

strategy — PROPOSED-UPDATE 2 → knowledge/paper-2-case-file.md, at the 2026-08-27 Rotenstein entry. The entry was built from the structured abstract; the full text is now read and adds material the abstract cannot contain. Proposed insertion:

Full-text verification 2026-09-24 (Stafford). Read in full at PMC13044793. This instrument does not report results by health system, and the manner of the omission is load-bearing. It names its five sites (Mass General Brigham, Emory Healthcare, UCSF, Yale New Haven Health, UC Davis); records that eligibility rules differed across them ("ranging from no restrictions on eligibility to restrictions based on clinical volume or eligibility based on leadership nomination"); tabulates institutional rollout characteristics in eTable 1 of Supplement 1 (unobtainable — download interstitial; UNREAD, do not characterise); runs leave-one-site-out sensitivity models; reports that excluding Mass General Brigham "only makes the presented estimates more conservative" — direction without magnitude; absorbs all between-site variation into clinician fixed effects; runs subgroup analyses on specialty, role, sex and use intensity only; and states that pooling makes estimates "less likely to be biased by site-specific implementation details." Consequence for this case file: the heavy-user gradient recorded here must not be cited as a dispersion result bearing on Claim 5 in either direction — it is the only subgroup dimension the paper estimated, and the instrument is silent on organizational variance by construction. File: stafford-research/knowledge-base/rotenstein-jama-ai-scribes-full-text-site-level-verification-2026-04.md.

strategy — PROPOSED-UPDATE 3 → knowledge/emerging-patterns.md, The Specification Gate and The Person-Level Moderator. These should not both carry independent recurrence counts. Proposed insertion on both entries:

Bookkeeping caution added 2026-09-24 (Stafford). These two patterns are one dispute about how to read the same evidence, and the scan's own entry says so ("the same evidence read two ways"). Sarkar is a member of both, and Lee/Cheon/Kim is simultaneously a member of one and the test subject of the other. Carrying two independent recurrence counts over an overlapping member set risks the outcome where both escalate on the same sources, one of which is the negation of the other. Recommended: mark them a mutually-exclusive pair that cannot both escalate, with the within-firm/between-firm decomposition as the named arbiter, and record the shared member explicitly on each. Additional discipline for the Specification Gate specifically: its second member (Lee/Cheon/Kim) has no organizational unit of analysis — verified against the primary on 2026-09-24, no grouping by repository owner, employer or firm anywhere, no firm-level and no outcome variable. That does not disqualify it from a pattern about specifiability, but it means the pattern currently has zero members measuring an organization, which should be stated in the entry rather than discovered later.

personal — PROPOSED-UPDATE 4 → knowledge/sources.md, monitored voices. Retire Andrej Karpathy from the standing daily check, after five consecutive briefs flagging it without a disposition. Evidence: karpathy.bearblog.dev returns HTTP 403 (Cloudflare challenge) — three failures this month plus today's; karpathy.github.io/feed.xml returns HTTP 200 but is the abandoned technical blog, ten posts, most recent 2026-02-12, no organizational content. A working route that reaches only the dead site is not a fix. Recommended: move to a manual read at the Friday synthesis, or remove under the 60-day zero-signal rule. Either way it stops occupying a line in the daily brief. Brandon can veto by editing sources.md.

governance — the diagnostic sharpens, and it is now an evidenced question rather than a rhetorical one. Yesterday's version: you deployed the same AI tool to more than one site — show me the results by site. Today it has a named precedent to cite. The stronger version: "who decided who was eligible for this tool, and on what rule? Now show me that rule next to the outcomes." Five American health systems answered the first half in a JAMA methods section and nobody has asked them the second. An organization allocating a productivity tool by managerial nomination while measuring its benefit in aggregate cannot tell whether it helped the people carrying the burden — and will report a successful deployment either way.

market — no window, seventh consecutive brief, and I'll answer the question the scan put to tomorrow's Friday review directly. The scan asks whether this system should keep opening publication windows at all, given that seven have opened, been logged with closing dates, and closed unwritten. My answer: stop opening them, keep recording time-sensitivity as a property of assets. The evidence is in today's work — asset 14 became substantially stronger because of a retrieval nobody was in a hurry to do, and it would have been worse under a deadline. This system's demonstrated competence is compounding evidence, not racing a news cycle. Concretely for tomorrow: record Accenture (closes 10-05) as lost, close the supervisor-pattern window (10-06, blocked nine days on one paragraph) and re-file its content as an asset, keep ONS (~10-30) as the single live one, and stop opening new ones unless something genuinely has a date attached.