Stafford Brief — 2026-10-09 (Friday)

This morning's 12:00 scan found someone who built a 36-dimension instrument for AI leadership and left out the one question Brandon puts first. That is the day's best find and it is not repeated here. Read it in research-evidence/briefs/2026-10-09.md; the Purpose omission, the Grant Thornton board threshold, and the two patterns that moved are all the scan's and all stand.

This brief does three things the scan could not. It ran the one fetch yesterday handed forward and got a result that cuts both ways on Brandon's biggest exposure. It found that the research pipeline's August repair fixed the output and never touched the input, so the contamination it removed is one command away from coming back. And on Friday's audit it has to report that the counterargument ledger now holds twenty-six open challenges and has closed one, ever.

📖 TODAY'S READ

A randomized trial put two kinds of leadership in the same model: the inspiring-visionary kind that every leadership programme on earth measures, and a narrow kind defined entirely by the specific thing being rolled out. The inspiring kind explained nothing. The specific kind explained about four-fifths of the result.

Williams, Ehrhart, Aarons et al., Implementation Science 2024;19:29, published 28 March 2024. This is the mechanism paper from the same trial as yesterday's find — the twenty-odd youth mental health clinics in Idaho, Oregon and Nevada where leadership was improved by coin-flip and the clinics that got it used the same software roughly ten times more. Yesterday's paper showed that it worked. This one asks through what, and answers it.

The chain the authors predicted in advance, and it held. Working on leaders improved a narrow thing called implementation leadership — whether a leader actually knows about the specific practice being introduced, backs it, and keeps going when it gets awkward. That improvement then produced a better working climate for the practice. The better climate then produced higher fidelity: the measure actually got completed and the feedback actually got opened. Each link was tested statistically and each one carried.

The part that matters is what did not carry. The researchers also measured transformational leadership — the broad, charismatic, set-a-vision-and-inspire-people construct. It is the single most measured idea in the leadership literature and it is what almost every executive assessment and leadership-development programme in the market is built on. The intervention barely moved it, and it explained none of the result to a degree you could distinguish from chance. Technical detail: implementation leadership mediated the climate effect with proportion mediated 0.82, p = .004; transformational leadership 0.29, p = .136. LOCI moved transformational leadership only at 4 months (d = 0.31) and not at 8, 12 or 18.

So what it changes, and it is the most useful thing found this week. Yesterday's brief told Brandon his differentiation had narrowed to a single untested claim: that five failure modes are distinguishable and that knowing which one is live changes what you do. Today supplies the claim one level below that — and supplies it at randomization. Being specific about what is broken is not a stylistic preference; in the one trial that tested it, generic organizational quality was inert and specificity did the work. That is the empirical case for having a diagnostic at all, and Brandon did not have it yesterday.

The bound, which has to be said in the same breath every single time. This is one axis with two constructs compared — general versus practice-specific. It is not the five-way discrimination. Brandon's stronger claim is still untested, and an essay that blurs the two will be caught. File: knowledge-base/williams-ehrhart-aarons-loci-mechanism-implementation-leadership-climate-fidelity-2024-03.md.

🔭 THINKER TO WATCH

None today, and manufacturing one would be padding. The scan registered Akben this morning; yesterday's brief registered Aarons and the Williams/Ehrhart cluster around him. Today's paper is the same three people. The honest entry is that yesterday's field-map proposal for Aarons is still sitting unapplied in yesterday's brief, which is the actual action item — see Routing.

🏢 CASE IN THE WILD

The case is this system for the second day running, and that repetition is itself the finding. In August the research pipeline was repaired. The repair fixed the file the dashboards read and never touched the files the pipeline reads, so the contamination it removed is still sitting in the inbox, one routine command away from coming back.

Here is the mechanism in plain terms. Every find enters the evidence base as a small staged file in a waiting room, then gets moved into the main index. In August a cleanup script repaired the main index: it restored reasoning that an old converter had dropped, recovered the Four Forces tags it never parsed, and fixed publication dates that had been contaminated with the date of ingest — the exact failure that damaged 241 entries and is now Hard Rule 1.

The waiting room was never cleaned. It holds 190 files — 189 of them before today's own entry was added. 151 are copies of entries already in the index, and 26 of those copies are the pre-repair versions — no reasoning attached, no Forces, and the publication date set equal to the date it was filed. The repaired versions have real publication dates; some are three weeks or a month earlier. The old, broken copies are still there, and one maintenance script walks that directory directly.

It gets one turn worse. The script that actually does the merging does not read the directory. It reads a separate ledger listing what is waiting. The directory holds 190 files; the ledger lists 32. Nothing reconciles the two. Six files have sat in that waiting room since July in neither list — never merged, never rejected, never seen. They are low-grade sources this base would reject today, so nothing of value was lost. What was lost is the guarantee: a staged find that fails to reach the ledger is invisible forever, and no process would notice.

And I reproduced the failure myself, by accident, within the hour. Staging a correction to yesterday's entry created a duplicate instead of patching it, because the duplicate check only compares against entries already merged and never looks at what is waiting. I removed my duplicate. For the next day, while yesterday's 31-file batch sits waiting, any one of those sources can be staged again and the check will not notice.

The breakpoint is Momentum Mirage, and it is the same one as yesterday with a new face. Yesterday: a correct rule written and never migrated, so 38 files sat stranded. Today: a correct repair applied to the output and never to the input. Both are a decision that was recorded and a thing that did not happen, and both were found by a process looking for something else. Three instances in two days, counting the governance gap the scan found in its own operating manual this morning.

Open question, and it is not rhetorical: who checks the checkers? Each of these was caught by a human-facing process stumbling over it. None was caught by the system. A reconciliation step — does the inbox match the ledger, does the ledger match the index — is perhaps ten lines, and nothing in either operating manual asks for it.

⚙️ FOUR FORCES CONCLUSION

Commitment, and for the first time it is separated from charisma by an experiment rather than by assertion.

The Four Forces define Commitment as decision-makers acting consistently with a stated direction when it costs something — deliberately not as inspiration, vision, or presence. That distinction has always been a definitional choice. Today it is a measured result. The construct that carried the effect is behavioural and specific: does this leader know about this practice, back it, and persist through its obstacles. The construct that failed is inspirational and general.

Why this is worth more than another supportive finding. The scan's own read this morning found Commitment appearing in the Akben Battery as Costly Ethical Enforcement — near word-for-word Brandon's definition, independently derived. Today's paper goes further and shows the alternative construct losing in a head-to-head. Independent rediscovery plus a beaten rival, in one day, on the same Force.

Capability is a null again, and it is the same structural null as yesterday. Every clinic got identical training and identical software by design. Capability in the form organizations actually purchase — licences, training, enablement — was held constant and therefore explains none of the difference. Two papers from one trial, both pointing at it. They are one instrument, and the Routing section says so.

💭 OPEN QUESTION

Is Brandon's diagnostic specific to the failure, or specific to the deployment? Because the trial that just helped him was the second kind, and the five breakpoints are the first.

The specificity that won today was specificity to the thing being rolled out — this practice, this workflow, this platform. A leader scored well by knowing about that. Brandon's five breakpoints are specific in a different direction: they name the kind of breakage — purpose, incentives, process, technology, momentum — independent of what is being deployed.

Those are orthogonal, and the trial only validated one of them. It is entirely possible that deployment-specific leadership is the whole effect and failure-mode discrimination adds nothing on top. It is also possible the diagnostic is doing something the trial could not see: LOCI was applied to clinics with one protocol regardless of what was wrong, so if a clinic's real problem was incentives rather than leadership, the trial has no way to tell us the intervention was mismatched — it would just show up as a smaller effect.

That is a testable question and it is cheaper than the full matched-versus-unmatched trial. Go back to the trial's own clinic-level variation: did LOCI work less well in clinics whose baseline problem looked least like a leadership problem? If the answer is yes, failure-mode discrimination has its first evidence. The data to answer it already exists and belongs to somebody Brandon could email.

🥊 CHALLENGE

Stated at full strength, and it is aimed at this base rather than at the world: the number everyone will want to quote from today's paper is the contaminated one.

This base spent 5 and 6 October on a charge it logged against itself — that it applies a strict one-questionnaire standard to vendor surveys and quietly waives it for peer-reviewed work. Today is the test case, because the paper is peer-reviewed, randomized, and argues for the framework. Applied honestly, the standard demotes the paper's best-looking result.

All three of the middle measures are clinician questionnaires. The same clinicians rated their leader and rated the climate that leader supposedly produced. So the crisp, strongly significant link — leadership to climate — is the same people answering both sides, which is exactly what this base refuses to accept from a vendor. The link that reaches an outcome nobody self-reported, where fidelity comes off the platform's own logs plus reports from caregivers, is the weak one. Technical detail: climate → fidelity, proportion mediated 0.71, 95% CI 0.02–1.00, p = .045, on 17 clinics. A confidence interval running from almost nothing to everything.

In plain English: the part measured by asking the same people twice is tight, and the part that reaches an independently recorded outcome is barely distinguishable from noise. And the paper never raises the issue — it presents clinician raters as a strength, calling them third-party informants.

Disposition: both numbers travel together or neither ships. The 0.82 must never appear alone, in an essay, a deck, or a client conversation. That is the same rule applied to Gravitee and to the Grant Thornton cuts, and applying it here against a find the base wanted to like is the only thing that answers the double-standard charge. Logged to knowledge/thesis-challenges.md.

The second update, which is better news. Yesterday's LOCI prior-art exposure — your cure already exists, it is manualized, it has a registered trial — got deeper and weaker at the same time. Deeper: there is a mechanism paper, a second randomized trial in substance use clinics, and a sustainment paper, so the competing cure is better evidenced than yesterday's entry said. Weaker at the joint that counts: the objection assumes general organizational quality is the right level of description, and the trial's own analysis refutes that. The principle that you must name the specific thing is now a finding of the very study being used against Brandon.

📈 PATTERN BUILDING

Nothing moved, and the deliberate non-move is the entry — because today's paper looks exactly like the third member of a pattern proposed yesterday and must not be counted as one.

Manipulation Beats Measurement, proposed yesterday at two members, holds that for organizational research at achievable sample sizes, causing the variable rather than observing it is the difference between a result and noise. Today's paper randomized an organizational condition at 21 clinics and got a result. It reads like member three.

It is not. It is the same trial, the same clinics, the same research group, the same 2019 cohort as member two. Counting it would be the identical error the scan refused this morning when it declined to let a second cut of Grant Thornton's 950 respondents stand as a second member. One instrument, two papers. Stated here so a future scan does not find two citations and promote the pattern to three.

The genuine third-member candidate exists and is unread: a second LOCI randomized trial in substance use treatment clinics (Journal of Substance Use & Addiction Treatment, 2024). Different sector, different clinics — but the same intervention and the same research group, so whether it counts as independent is a question a read has to settle, not an assumption. Named in the query log as the next slot's target.

🪟 WINDOW FLAG

None. Twentieth consecutive null, and the tally stays one finding rather than twenty. A paper published in March 2024 is argumentative value with no clock on it. Calling it a window would be manufacturing the exact urgency Hard Rule 5 forbids.

🧱 ASSET SUGGESTION

New asset 18: "The Leadership Training You Bought Measures the Wrong Thing."

The pitch in two sentences. A randomized trial compared the inspirational leadership construct that nearly every executive assessment and leadership programme is built on against a narrow one defined by the specific thing being rolled out; the famous one explained nothing and the narrow one explained four-fifths of the result. The organizational capability companies actually buy is measured at the wrong level of description — and that is the empirical case for diagnosing which condition is broken instead of improving leadership in general.

Why it is distinct rather than a slice of something committed. It is the first entry in the ledger whose subject is the level of description rather than a finding about AI, and it is the only one that defends the existence of a five-way diagnostic rather than defending a particular breakpoint. It also pairs naturally with the scan's registered asset from this morning — 36 behavioural dimensions and no item for outcome agreement. Combined, the argument is sharper than either alone: the field is measuring leadership at ever-finer resolution on the wrong axis. Form: a short essay, roughly 1,200 words, one table of two numbers, one concession.

Asset 2 still ships first. The 2 October commitment stands and this does not jump the queue.

And an amendment to asset 3, part 3b — no new number. Yesterday the blocker on 3b became the outcome definition. Today supplies a structural template for half of it: the trial's fidelity measure is two records the deployed system produces anyway — was the step completed, was the output actually opened — crossed with a report from someone other than the person being evaluated. The transferable rule is that at least one of your two records must come from the receiving party, and that is the half practitioners never build. It does not supply the outcome definition itself: fidelity is still not converted value, and the authors say so. Both entries are in knowledge/asset-suggestions.md.

🎓 SPARRING NOTE

Declined — the fifth decline in the last seven days and the right call, with the 10-08 prompt upgraded in place instead.

The cap permits a note: Monday was 5 October and only one has been issued since. But the 10-08 prompt is one day old, on the same growth edge, and unanswered. One unanswered prompt on an edge is a prompt; two is a reading list, and that mechanism is what produced the backlog of five notes Brandon had to release as non-blocking on 2 October.

What changed is that the 10-08 prompt got easier, not harder. It asked for the one-sentence outcome definition that would make a breakpoints-remediation trial about converted value rather than usage — and then what it would be measured from. The second half now has a worked example: two records the system generates anyway, crossed with a report from the receiving party. The first half — the definition itself — is untouched and is still the real work. If anything today sharpens the exercise: it shows the measurement half is an afternoon's work and the definition half is not. Full entry in knowledge/brandon-development.md.

🔍 SOURCES & THREADS

Threads: 0 due, 0 checked, 0 opened, 0 closed, and that is correct rather than idle. The earliest next-check is 11 October, two days out. The scan declined to re-check anything early this morning and cited the 7 September cadence correction for the reason; that judgment is right and is not second-guessed here. No thread was opened for today's find: a mechanism paper from 2024 completing a 2026 paper already in the base is the end of a story, not chapter one, and opening a thread on it would pad the ledger.

Exploratory slot (Friday, standard): spent on the named acquisition target yesterday handed forward, and it hit on the first query. Yesterday's method result instructed the next slot to stop describing the design it wanted and name the specific paper. That instruction worked immediately — Springer required a login, the open-access copy on PMC did not, and the full text was read. Both query logs were checked first for LOCI, Aarons and mediation: the scan's log has never run this; this file's only hit was yesterday's own handoff. Logged with the full yield in knowledge/query-log.md.

Method result for the next slot, and two named targets unspent: naming the specific paper beats describing the design. Targets, in order — the second LOCI randomized trial in substance use clinics (independence must be established by a read before it counts as anything), then the WISDOM sustainment paper.

Candidates: 0 added, 0 promoted, 0 expired. Nothing new surfaced; today's source sits inside a journal and a research programme that yesterday's brief already proposed as candidates, and both proposals are still unapplied.

Rule 6 paid again. The mechanism paper was checked against both repositories by DOI, title and the index before anything was written — clean, genuinely new. And the same check is what caught my own duplicate minutes later, which is how the staging gap in the Case section was found.

Two writes into the evidence layer, both inside the scope Brandon set on 2 October. One new knowledge-base file copied across and staged. And yesterday's uptake entry patched: it had gone into the pipeline with its instrument field empty, which Hard Rule 9 forbids for exactly this class of finding — the base's strongest-identification source had no record of its own design. It now carries the data source, the randomization, the arm sizes, the 33-month window, and a note that it and today's paper are one instrument.

Weekly Synthesis

1. Patterns promoted: none, and one deliberate refusal to inflate. Manipulation Beats Measurement stays at two members for the reason in PATTERN BUILDING — today's paper is the same trial as member two. The scan refused an escalation this morning on its own 18 August condition for the sixth time and refused a new pattern on shared-weakness grounds. Three separate refusals to inflate in one day across two agents is the system's best discipline signal of the week, and worth noting because the opposite failure is the one that quietly ruins an evidence base.

2. Paper-2 evidence status — one claim moved one rung, one claim has not moved in five scans.

Strongest this week: Claim 5's identification request, which the 10-08 split recorded as met for adoption and open for converted value. Today it moves part-way and does not close. Fidelity — did the step get completed, did the output get opened — sits on the quality-of-delivery rung, which this morning's scan independently named as exactly what Claim 5 needs and the AI literature has never operationalized. Two routes, one day, same construct, found from opposite directions and neither able to see the other. That convergence is worth more than either finding.

Weakest, and it is now a five-scan problem: Claim 3's Strategic Disconnection cell is still occupied by Verlinden et al. (2022) alone — pre-generative, abstract-grade, coefficients owed since 5 October, access routes still failing. The scan named it this morning as the base's oldest load-bearing single-source dependency. Nothing has moved it for five consecutive scans. It should take the next exploratory slot that is not already committed, ahead of both LOCI chase targets, because a load-bearing cell with one unreadable occupant is a worse exposure than an unread companion paper.

3. Field map: one addition this week (Akben, by the scan), and yesterday's Aarons proposal unapplied.

4. Staleness: today's find is the useful case of old-but-live. Published March 2024 and never ingested here until today, which is a gap in coverage rather than staleness. The thing it completes is twenty months newer than it is — the 2026 uptake paper came out after the 2024 mechanism paper, so "most recent" would have been the wrong retrieval order and nearly cost the better result.

5. Windows: none. Twentieth consecutive null. Nothing opened, nothing closes within fourteen days.

6. Challenges audit — and this is the Friday finding Brandon should act on. The ledger holds 29 challenges. 26 are OPEN. One has ever been closed.

Ten new open challenges were logged in the twelve days from 28 September. Three are past the 30-day threshold the audit exists to surface, with no response plan recorded: the measurability confound (1 September, 38 days), the organizational gate is removed, not overwhelmed, and the light-touch counterexample — HealthPlix.

The count is the problem, not the ages. A counterargument ledger accumulating about one a day and resolving roughly none is not a working instrument; it is a backlog that makes every individual entry unreadable. And this base has already diagnosed and fixed this exact failure once, somewhere else. The coaching file hit it in September — notes stacking faster than they could be answered until the pile itself became the obstacle — it was named explicitly, and Brandon fixed it on 2 October by triage: declare one live item, release the rest as non-blocking, resume. That worked. The challenge ledger has the identical pathology, four times the size, and the remedy has not been carried across.

That is the third instance this week of a rule being written in one place and not carried to the structurally identical place. The write-access rule with no migration (yesterday, 38 files). The August repair applied to the output and not the input (today, 26 contaminated copies). And now a triage remedy applied to the coaching ledger and not the challenge ledger. The recurring defect in this system is not the rules. It is that nothing asks "where else does this apply?" Recommended, and it is Brandon's call because it changes the operating rules: when a fix lands, name the other place it belongs — or say out loud that there isn't one.

7. Positioning intelligence. The scan's reading stands and is sharpened by today. Three consultancies measured the governance substrate this quarter, all three selling the remedy, none measuring an outcome outside its own questionnaire. Today adds the method they are all missing and shows it is cheap: cross one record the system already produces with one from the receiving party. That is the differentiator — not a better survey, a measurement that does not ask the interested party. It is the sharpest consulting position the base has produced in a week, and it is available now.

Routing

strategy — PROPOSED-UPDATE to knowledge/paper-2-case-file.md, anchor: gap ranking, item 1 / request (b). Add beneath the 10-08 split:

Request (b) moves one rung and stays open. Williams, Ehrhart, Aarons et al. (Implementation Science 2024;19:29, 2024-03-28; same WISDOM trial, NCT04096274) report a fidelity outcome — the product of a platform-recorded completion rate and a platform-recorded feedback-viewing rate, combined with monthly caregiver reports — mediated by implementation climate (proportion mediated 0.71, 95% CI 0.02–1.00, p = .045, 17 clinics). This occupies the quality-of-delivery rung that the 2026-10-09 scan independently identified as what Claim 5 requires and the AI literature has never operationalized. It is not converted value — the authors state the index "does not indicate whether clinicians used the feedback to guide treatment" — so request (b) stays the top item, now with a worked measurement template rather than only a specification. Carry the common-source bound: the strong link in the chain is same-respondent and the link reaching a non-self-reported outcome is the weak one; the 0.82 figure may not be cited alone.

strategy — PROPOSED-UPDATE to knowledge/paper-2-case-file.md, counterarguments section. The scan asked this morning for a fourth entry covering the measurement rival. Add a fifth, and it is the answer rather than another objection: the general-versus-specific result. Generic transformational leadership mediated nothing significant (0.29, p = .136) where practice-specific implementation leadership mediated 0.82 (p = .004), at randomization. This is the framework's first external, randomized support for the premise that the level of description matters — the premise underneath having a diagnostic at all. State the bound in the same paragraph: one axis, two constructs, not the five-way discrimination.

strategy — PROPOSED-UPDATE to knowledge/emerging-patterns.md, anchor: Manipulation Beats Measurement (proposed 2026-10-08, 2 members). Append: "2026-10-09 — a third member was considered and refused. Williams, Ehrhart, Aarons et al. (Implementation Science 19:29) is the SAME trial, clinics, cohort and research group as member two; counting it would be the shared-instrument error. The genuine candidate is the second LOCI RCT in substance use treatment clinics (JSAT 2024, S2949-8759(24)00149-8) — different sector and clinics, same intervention and group — and whether it is independent must be settled by a read before it counts."

governance — the staging pipeline, reported and not fixed, because build/ is outside Stafford's write scope. Three defects, all verified against the files this morning, in descending order of consequence:

The merge script reads a ledger; the directory holds something else. entries/pending/ contains 190 entry files; entries/.staging-index.json lists 32. auto-merge.py iterates the ledger only, so 158 files are invisible to it — including 6 that have sat there since July in neither list, never merged, never rejected (all low-grade sources this base would reject today, so the loss is nil; the missing guarantee is not). Fix: a reconciliation check that fails loudly when the directory and the ledger disagree. Ten lines, and nothing in either operating manual currently asks for it.
The August repair never reached the inbox. 151 of those pending files duplicate entries already merged, and 26 are the pre-repair versions — zero breakpoint reasoning, zero Forces, and published_date set equal to date_added, which is Hard Rule 1's contamination fingerprint exactly. The merged copies carry real dates (one is a month earlier). reconcile-kb.py repaired the index; the staged inputs were never touched, and normalize-pending.py walks that directory directly. Fix: re-run the reconcile over entries/pending/, or retire the already-merged copies. Retirement is destructive and 157 files wide, so it was not executed unattended.
The duplicate gate cannot see the queue. Gate 6 in add-research-json.py compares only against index["research"] — the merged index — so re-staging a source that is staged but not yet merged creates a second entry under a new date-prefixed id. Reproduced accidentally this morning and cleaned up. CLAUDE.md's promise that "duplicates patch missing fields instead of failing" holds only for merged entries. This is live right now: yesterday's 31-entry batch is pending, so any of those sources can be re-staged undetected until it merges. Fix: extend Gate 6 over the pending queue.

governance — the instrument taxonomy has no category for the strongest evidence in the base. add-research-json.py accepts nine data_source_genre values and all nine are observational (resume_panel, postings, survey_self_report, survey_paired, administrative, telemetry, outcome_defined, model_theory, unknown). There is no genre for a randomized trial. Both WISDOM entries are staged as telemetry — truthful about the outcome, silent about the design — with the randomization written into inference_method so nothing is hidden. Proposed: add randomized_trial. The schema currently encodes the base's own blind spot, which is a small, literal instance of the thing the scan's Invisible Cause pattern is about.

governance — the challenge ledger needs the triage Brandon already applied to the coaching ledger. 26 open, one ever closed, three past 30 days with no response plan. The remedy is proven and sitting one file away; see Weekly Synthesis §6. Brandon's call, because it changes how the ledger is operated.

personal — the cheap test of the thing Brandon actually needs, and it belongs to someone he could email. The matched-versus-unmatched trial is the study that settles the five-way claim and it is expensive. A much cheaper partial answer already exists in data that has been collected: whether LOCI worked less well in the clinics whose baseline problem looked least like a leadership problem. The trial measured leadership, climate and fidelity at 21 clinics with baseline values for all of them. If the effect is smaller where the diagnosis was mismatched, failure-mode discrimination has its first evidence — and it cost one analysis of existing data. Williams is at Boise State, Aarons at UC San Diego, and yesterday's brief already proposed both for the field map.

market — yesterday's four PROPOSED-UPDATE blocks are all still unapplied (paper-2 item 1; Manipulation Beats Measurement; Aarons to the field map; two source candidates). That is expected — the shared meta files are Brandon's to apply — but the queue is now nine blocks across two days. Worth knowing that today's proposals assume yesterday's anchors exist; if yesterday's are skipped, the item-1 update above has nothing to attach to.

OPS — research-evidence branch claude/modest-faraday-nsjb4j: 1 knowledge-base file copied, 1 entry staged (2026-10-09-implementation-improving-measurement-based-care-im), 1 prior entry's empty instrument field patched in place, 1 self-inflicted duplicate removed. Pushed to that branch only, never its main, per the division of labour; promote-claude-branches folds it in once scripts/verify.sh passes and opens a pull request if it does not. Yesterday's run merged cleanly — claude/modest-faraday-ux5qfm was promoted, so the 37-file backfill landed. This repo's disposition is on the final line of this run.