Stafford Brief — 2026-10-21

Wednesday falsification sweep. 1 KB file, 1 staged entry, 1 hold resolved, 1 challenge materially enlarged, asset 12 advanced. The sweep did not falsify the claim it went after — and the honest version of that is the day's most useful sentence.

Today's scan brief is excellent and I am not repeating it. Read it for Zhaobei, Tapread, the CAMO network-overlap escalation and the agent-KPI-hacking window. This brief is the delta: what today's work changes in the challenge ledger, in the Tier-2 gate, in the base's own quality rules, and one correction to my own record from yesterday.

One correction to yesterday, stated first because it changed what I was about to write into the shared index. Yesterday's verification hold said the Fu et al. abstract and conclusion both describe the day-level effect as "close to zero." The abstract does not. I re-read it verbatim from the arXiv API today, and its closing result sentence runs the other way: the ordering gain is "outweighed by a much larger negative behavioral response." So the staged key finding was not falsified, as I claimed — it conflated two different things, and that is the correct and narrower charge. Details in OPS.

📖 TODAY'S READ

The remedy that is supposed to make Brandon's framework unnecessary has now been tested twice, in two different industries, on about eighty clinics — and it still only moves whether people use a thing, not whether the thing works.

Aarons, Sklar, Ehrhart, Roesch, Moullin & Carandang, Journal of Substance Use and Addiction Treatment 2024;165:209437. This is the second randomized trial of LOCI — a twelve-month, off-the-shelf programme that coaches an organization's first-level leaders and reshapes the climate around whatever new practice is being introduced. Sixty clinics across nine behavioral-health organizations in California and Arizona were split at random, inside each organization, between LOCI and a plain leadership webinar. 380 front-line providers were surveyed.

Why it is today's read is not the result. It is what the trial registered and then did not report.

Every clinical trial files its intended outcomes in advance, on a public registry, so that nobody can quietly swap the question after seeing the answer. I pulled this trial's filing (ClinicalTrials.gov, NCT03042832) and read it next to the published abstract. Three things do not line up:

1. The filing names twelve months as the moment of truth. The paper reports four months. The study did look at four, eight and twelve — it says so — but the differences it reports are the four-month ones.
2. The filing lists two co-equal headline outcomes. One of them is missing from the results entirely — the generic "inspiring leader" measure (see below).
3. The filing promises to measure whether clinicians actually delivered the practice properly. The paper reports whether they were willing to submit their sessions to be checked. Those are not the same thing. The second is participation in being audited.

Technical detail: between-condition differences reported at 4 months for implementation leadership, implementation climate and implementation citizenship behaviour; MI "reach" by chi-square (1, n = 370) = 5.59, p = .018. Registered primaries — Implementation Leadership and Full Range Leadership — both at 12 months; registered secondary "Fidelity" at 12 months. Results never posted, despite primary completion 2020-04-20.

And then the conclusion claims the programme delivers "greater sustainment" of the new leader behaviours. A durability claim, resting on the earliest of three planned check-ins, from a trial whose results have not been posted in six years. That is Momentum Mirage — the appearance of progress standing in for progress — and it is sitting inside the cure.

What Brandon takes from it: the strongest prior art against his framework has the same disease his framework is named after. He should say so in Asset 12 — gently, with the registration number, and without overclaiming.

The bound, and it matters. The article is paywalled and I did not read it. Everything above comes from the authors' own abstract and the public registry, both read word-for-word. It is not available to say the eight- and twelve-month results were null — only that the paper reports four-month ones against twelve-month filings. Anyone carrying this into print reads the full text first. File: knowledge-base/aarons-sklar-ehrhart-loci-substance-use-clinics-randomized-2024-10.md.

🔭 THINKER TO WATCH

Gregory Aarons, University of California San Diego — and he is here because he is the most serious competitor Brandon has, not because he is adjacent.

Aarons built LOCI and has spent fifteen years doing the thing almost nobody in the change-management market does: manualizing an organizational remedy and then randomizing it. Two registered trials, two sectors, roughly eighty clinics. Brandon has a framework, six papers, and zero randomized anything. On the question does fixing the organization change whether technology gets used, Aarons has the better evidence and it is not close.

Brandon's differentiation, stated honestly rather than defensively. Aarons applies one protocol to every organization regardless of what is wrong with it. Brandon's claim is that there are five distinguishable things that go wrong and that naming which one is live changes what you do. That claim is untested — today's sweep went looking to break it and instead watched the undifferentiated alternative pick up a second industry.

But here is the part worth Brandon's attention, and it comes from Aarons' own work. LOCI measures two kinds of leadership side by side: the generic inspiring-visionary kind that the entire leadership-training industry sells, and the specific kind — does this leader know about, back, and stick with this particular new practice. In the earlier WISDOM trial the generic measure did nothing and the specific one carried the whole effect. In this trial the generic measure was registered as a headline outcome and never reported. Twice now, in two sectors, Aarons' own programme has found that improving the organization in general is the wrong level of description. The principle underneath Brandon's whole diagnostic — you have to name the specific thing — is a finding of the research programme being used against him.

🏢 CASE IN THE WILD

Not a new case — the scan's Tapread write-up stands and I am not duplicating it. One judgment to add, and it is about the base rather than the firm.

The scan closed its Tapread section with a genre caution: four of four CAMO teaching cases contain a contradiction inside their own reported results, now recorded in sources.md as a property of that source genre.

Today's find says that caution is not about teaching cases. A peer-reviewed, registered, randomized controlled trial — the highest-credibility genre this base recognizes — reports a four-month result against a twelve-month filing, substitutes willingness-to-be-audited for the practice outcome it promised, drops a co-headline outcome, and concludes with a durability claim. Same shape. Different genre. Much better reputation.

So the rule should not be filed under CAMO. The thing that caught both is not scepticism about Chinese business schools; it is reading the source's own stated plan next to its reported result. For a teaching case that plan is the structure it says it designed. For a trial it is the registry entry. Proposal in Routing: generalize the caution one level up, out of sources.md's CAMO note and into the base's standing method.

The open question this raises for client work: if the best-identified organizational intervention in existence reports participation-in-measurement as its practice outcome, what exactly is any client's "adoption" number measuring?

💭 OPEN QUESTION

If the strongest competing remedy turns out to move participation in being measured rather than performance — how would Brandon know that his own remedy isn't doing the same thing?

This is today's criticism turned around and pointed at the framework, which is the only honest direction to point it.

LOCI's reach outcome is clinicians agreeing to submit sessions for coding. Brandon's remedies look structurally similar: name the stakeholders whose metrics won't improve and get a written concession (Tier 3.2); ask the three teams whose agreement determines delivery, separately (Tier 3.1). Each of those produces a visible artifact of engagement. A client who signs the concession, attends the session, fills in the rubric — that client has participated. It is exactly the kind of thing that shows up as a success and costs nothing real.

The uncomfortable version: a consulting engagement is a participation-generating machine by construction, and the post-engagement measurement in master-plan Tier 0.2 is the only thing standing between a real instrument and an elaborate attendance record.

So the question has a design answer and Brandon should take it into 0.2 this week: the one agreed outcome per engagement must be something a client could fully participate in and still fail at. If a client can get the outcome by cooperating, the measurement is worthless. That is a one-sentence constraint on a one-page decision, and today's find is the reason to write it down — a trial with randomization, eighty clinics and a registry still ended up measuring cooperation.

🥊 CHALLENGE

The Wednesday sweep went after Brandon's own differentiator at full strength and did not break it. It did something worse: it made the opposition roughly twice as big while leaving the claim exactly as untested as it was.

The target. Five failure modes are distinguishable, and knowing which one is live changes the remedy. The 10-09 entry calls this "the whole of the differentiation." Breaking it requires one undifferentiated protocol working across different settings — which would show that matching the remedy to the diagnosis is unnecessary.

The objection at full strength, as it now stands. LOCI is a named, manualized, twelve-month protocol. It has two separately registered randomized trials — NCT04096274 in youth mental-health clinics, NCT03042832 in substance-use clinics — in two sectors, across roughly eighty clinics, with a tested mechanism behind it. It applies the same protocol every time and reports positive results both times. Brandon's five-way discrimination claim has no trial, no outcome data, and no instrument. The hostile one-liner, which Asset 12 should absorb rather than dodge: the competing cure has eighty randomized clinics and the framework has a taxonomy.

What survives, and it is the premise rather than the conclusion. The sharp form of the objection — why buy a five-way diagnostic when one protocol has trials behind it? — assumes that general organizational quality is the right level of description. Aarons' own programme has now refuted that twice: the generic leadership construct was inert in WISDOM (mediated 0.29, not statistically distinguishable from nothing, against 0.82 for the practice-specific one) and was a registered headline outcome that went unreported here. Improving leadership in general did not do the work; being specific to the thing being implemented did.

Two weaknesses in the prior art that the base did not have before today, both of which belong in Asset 12 as statements of fact rather than as gotchas: the remedy has its own Momentum Mirage (a sustainment claim on four-month data against twelve-month filings, results unposted for six years), and its practice outcome is participation in being measured.

What it would take for this objection to win — unchanged: a trial showing that a remedy matched to a diagnosed failure mode does no better than the same remedy applied unmatched. What would defeat it is the mirror, and it is still the study asset 3b should be built to run — now against a comparator with two sectors behind it, which raises that bar rather than lowering it.

Status: OPEN, and this is now the largest exposure in the ledger by evidence volume. Full entry: knowledge/thesis-challenges.md, update 2026-10-21.

📈 PATTERN BUILDING

Not a new pattern — an extension that takes an existing one out of the genre it was discovered in, which is more valuable than a sixth member would be.

The scan's Borrowed Apparatus has a live finding from today: Tapread built a designed structure and measured against none of it, reporting instead the throughput already sitting in its queue log. The scan's sentence for the paper: measuring conformance requires deciding in advance what the right path was; a throughput number does not. It registered a prediction that the remaining CAMO cases would each do the same.

Today's trial is the same shape, outside the genre, at the top of the credibility ladder. It registered a practice outcome — did clinicians deliver the method properly — and reported a log-resident one: sessions recorded and submitted for coding. Deciding in advance what the right path was is precisely what a trial registration is, and this trial did it, filed it publicly, and then reported the count its monitoring system happened to produce.

Why that strengthens the pattern rather than just adding to it. The scan's cost-gradient explanation was already refuted today (Tapread proved a to-be model is cheap when reorganizing produces one anyway). This shows the substitution survives even when the specification is not merely cheap but mandatory, public and pre-filed. So the mechanism is not cost and not laziness — log-resident outcomes win because they arrive without anyone having to adjudicate them. A conformance measure forces a judgment call about whether the path taken was the right one. A count does not.

Count stays where the scan has it; the proposal is a scope change, not a member. Proposed text in Routing.

🧱 ASSET SUGGESTION

Asset 12 — "What the Five Breakpoints Owe" — just got the anchor its prior-art section needed, and it costs nothing to use.

The argument, two sentences. The framework's most dangerous objection is that its remedy already exists, manualized, with randomized trials behind it; the answer is to concede that completely, by registration number, and then show that the objecting programme's own trials establish the one principle the framework depends on — that organizational improvement in general is the wrong level of description. Conceding eighty randomized clinics is more persuasive than conceding twenty, for the same reason retiring the 70% figure in Paper 3 bought credibility instead of costing it.

Form: the prior-art section of Asset 12, which master-plan Tier 2 names as the gate nothing in the differentiation class ships before. Three paragraphs are now written in all but prose — the concession, the premise refutation, the honest residual — and are spelled out in knowledge/asset-suggestions.md (amendment 2026-10-21).

No new numbered asset. Hard stop 1 holds; this lands in the existing entry.

And the drift worth naming: asset 2 is the only COMMITTED asset, master-plan Tier 1 says finish it, and Stafford has now spent two consecutive days on Tier-2 material instead. The punch-list items Stafford can clear (Sequoia/Block, the Accenture sample and field dates, the Farach half of item 2) are untouched. That is a re-rank by default, which hard stop 4 forbids, so it is logged as drift rather than proposed as an order — tomorrow's slot clears a punch-list item unless Brandon says otherwise. Item 4 is Brandon's alone and is nineteen days old.

🎓 SPARRING NOTE

None, deliberately. The coaching loop has been switched off since 2026-09-25 after nine consecutive declined notes, and master-plan Tier 0.1 makes answering the 2026-09-17 note the thing that unblocks it. A tenth undeclined note would be decoration. Today's material is directly relevant to that fifteen-minute task — the hostile reader who knows Kotter cold now also knows Aarons — and that is the whole of the coaching content until 0.1 is done.

SOURCES & THREADS

Threads: 0 due, 0 advanced by me. The scan reported 0 due with the nearest at 2026-10-28; I re-checked the thread file's forward dates independently and agree — nothing is due today, so nothing was chased. The scan advanced four threads on its own finds.

Exploratory slot — Wednesday falsification sweep, spent on the named unspent target from the 10-09 slot. Both query logs grepped first for S2949-8759 / substance use treatment / normalization from 10-10 forward: zero hits, not a repeat. Query and full yield logged in knowledge/query-log.md.

Method result worth carrying forward, and it is the transferable part of today. The registry is this base's underused instrument. The find came from reading NCT03042832 next to the abstract — neither the paper nor any search would have surfaced a registered-but-unreported co-primary outcome. Standing addition proposed: for any trial-based prior art, read the ClinicalTrials.gov record alongside the abstract before judging the result. Retrieval beat search again, for the fourth time this month; PubMed blocked on cookies and the working routes were the Europe PMC and ClinicalTrials.gov APIs.

Candidates: 0 added, 0 promoted, 0 expired by me. No new outlet — JSAT is Elsevier and the research group is already in the field map. Citation chasing produced one acquisition target, not a candidate: the full text of this trial, which is the highest-value unread page in this lane and needs an author or institutional copy.

Nothing skipped. One source examined, one ingested. The Wednesday slot is a single targeted retrieval by design, not a sweep.

Routing

strategy — Two items, and the first is a one-sentence constraint worth more than the find. (1) Take into Tier 0.2 this week: the one agreed outcome per engagement must be something a client could fully participate in and still fail at. If cooperation can produce the outcome, the measurement is an attendance record. A randomized trial with eighty clinics and a public registry still ended up measuring cooperation — that is the warning. (2) Asset 12's prior-art section is now three paragraphs from written, with registration numbers, and conceding at full scale is the Paper 3 move again on better material.

market — Nothing from me today. The scan's 4–8 week Zhaobei window is the live one and I have no reason to adjust it.

governance — The reach-outcome problem is a client-facing diagnostic question, and it is sharper than the governance item the scan filed: your adoption number — could a team hit it by cooperating? If yes, it measures cooperation.

personal — The drift named in ASSET SUGGESTION is the item that needs a word from Brandon: two days of Stafford budget have gone to Tier-2 material while Tier 1 (asset 2, the only COMMITTED asset) sat still. Hard stop 4 forbids Stafford re-ranking, so it will revert to the punch list tomorrow by default. Say so if you want the Tier-2 run continued instead.

PROPOSED-UPDATE — knowledge/emerging-patterns.md (research-evidence), The Borrowed Apparatus, scope note dated 2026-10-21

Scope extension, 2026-10-21 (Stafford). Not a new member — the pattern leaves the teaching-case genre. Aarons et al. 2024 (J Subst Use Addict Treat 165:209437, NCT03042832) registered a practice outcome — fidelity of motivational-interviewing delivery, filed publicly as a 12-month secondary — and reported a log-resident substitute: providers' willingness to record and submit sessions for coding (chi-square (1, n = 370) = 5.59, p = .018). A trial registration is deciding in advance what the right path was; this trial did it, filed it, and then reported the count its monitoring system produced anyway.

This refutes the remaining cost-based reading of the gradient. Today's Tapread finding already showed a to-be model is cheap when reorganizing yields one as a side-effect. This shows the substitution survives when the specification is not merely cheap but mandatory, public and pre-filed. So the mechanism is neither construction cost nor laziness: log-resident outcomes win because they arrive without anyone having to adjudicate them. A conformance measure forces a judgment about whether the path taken was right; a count does not.

Recurrence count unchanged — this is a scope widening (the pattern holds in peer-reviewed registered trials, not only in teaching cases), not a sixth member.

PROPOSED-UPDATE — knowledge/sources.md (research-evidence), the 2026-10-21 CAMO genre caution

Generalize one level up. Today's caution — four of four CAMO Nano Cases contain a contradiction inside their own reported results — should not be filed as a property of CAMO. The same shape appeared today in a peer-reviewed, registered RCT (NCT03042832): a 4-month result reported against 12-month registered primaries, a registered co-primary (Full Range Leadership) absent from the results, a registered practice outcome replaced by a participation count, and a sustainment claim in the conclusion.

The method that caught both is the same and it is not genre-specific: read the source's own stated plan next to its reported result. For a teaching case the plan is the structure it says it designed; for a trial it is the registry entry. Keep the CAMO count as an instance; record the method as standing practice.

PROPOSED-UPDATE — knowledge/query-log.md (research-evidence), standing method addition

For any trial-based prior art, read the ClinicalTrials.gov record alongside the abstract before judging the result. Verified 2026-10-21: the entire LOCI find came from reading NCT03042832 next to the published abstract. A registered-but-unreported co-primary outcome is invisible to the paper and to any search; only the registry shows it. Registry reads are free, fast (clinicaltrials.gov/api/v2/studies/<NCT>) and resolve pre-specification questions no amount of searching will.

PROPOSED-UPDATE — knowledge/field-map.md (research-evidence), Gregory Aarons entry

Promote from cited-author to tracked competitor, 2026-10-21 (Stafford). Gregory Aarons (UC San Diego) — builder of LOCI, now two separately registered cluster randomized trials in two sectors (NCT04096274 youth mental health, 20 clinics; NCT03042832 substance use treatment, 60 clinics) of a manualized twelve-month organizational remedy. The most serious competitor in the base on the remedy half of Brandon's practice, and on does fixing the organization change whether technology gets used he holds better evidence than the framework does.

Network-level non-independence applies, as with CAMO. Aarons and Ehrhart co-author across both trials and the mechanism paper; the three LOCI items in this base are one programme and must never be cited as independent confirmations of each other. Separate registrations, sectors and randomizations make the trials separate trials — not separate evidence of the programme's validity.

Differentiation, stated as it should be stated in print: LOCI applies one protocol regardless of what is wrong; the framework claims five distinguishable failure modes and that matching matters. That claim is untested. Aarons' own mediation result — generic leadership inert, practice-specific leadership carrying the effect — supports the framework's premise while leaving its central claim unproven.

STAGED — 2026-10-21-journal-of-subs-randomized-trial-of-the-leadership

Staged directly in research-evidence per operating rule 3, gates passed on the second attempt (data_source_genre must be one of the pipeline's enums; survey_self_report with the single administrative outcome described in inference_method). forces: [] deliberately — the article is paywalled and unread in full, and Momentum is withheld rather than guessed (hard rule 2). One breakpoint only, momentum_mirage, rather than double-tagging the measurement substitution as a second breakpoint. instrument populated, including the network-level non-independence flag against the two WISDOM entries.

OPS — disposition on the final line of this run's report; scripts/verify.sh is the gate. Yesterday's run reached main (ee1215a/98a004d), so no OPS flag is carried forward into today.

The 10-20 verification hold is resolved, and resolving it required correcting my own record. The held entry was 2026-10-20-arxiv-preprint-when-better-opportunities-backfire. The hold worked exactly as intended — build/auto-merge.py skipped it, and it did not reach research-index.json at the 12:31 UTC timeout today. Today's scan did not amend it, so deferring a second time would have stranded it indefinitely, which is the failure mode knowledge/pending-proposals.md documents eighty-one times over.

Before applying the corrections I re-read the arXiv abstract verbatim (arXiv API, 2610.25603) — and yesterday's hold misattributed its own evidence. It stated that the abstract and conclusion both describe the aggregate day-level impact as "close to zero." The abstract contains no such statement; its closing result sentence is "A calibrated decomposition analysis further shows that the modest gain from improved lead ordering is outweighed by a much larger negative behavioral response" — which reads the other way. The "close to zero" language belongs to the full text's aggregate estimate (Appendix C), not to the abstract.

So the charge was wrong and the narrower charge is right. The staged key finding was not falsified; it conflated the authors' calibrated channel decomposition (the behavioural channel outweighs the ordering channel) with a measured net day-level effect (statistically indistinguishable from zero). Those are different quantities and both are in the paper. key_findings[4] was rewritten to state them separately, which is accurate under either reading, and the hold is therefore no longer warranted.

Applied to the staged entry (the scan's breakpoint and forces rationales left untouched — they are its authorship): the rewritten key_findings[4]; a new finding giving the within-day pattern behind the day-level zero (−4.5pp at calls 1–3, −2.5pp at 4–6, unchanged mid-day, +3.4pp at end of shift); instrument.observation_window filled in (deployment 2024-08-28, ±60 days with a 10-day donut, roughly July–October 2024, with the warning that the preprint is dated 2026-10-19 and the data is two years older — hard rule 9); instrument.inference_method corrected on compensation (§3.1.3 does describe the scheme and documents it unchanged through the AI expansion — observed-and-constant is the identification, not a limitation); and the real sample counts. Not applied: the Table 6 material and the figure refinements (23.4 / 4.5 / 8.1pp), which could not be re-verified from the abstract in this run and are left as named targets for a full-text re-read. The sidecar now carries this reconciliation in full and its status is resolved_stafford_corrections_applied, so the entry will auto-merge normally.

Writes to research-evidence, all within operating rule 3's direct-write scope: one new knowledge-base/ file (the LOCI trial), one staged entry via build/add-research-json.py, and the amendment plus hold release on the 10-20 pending entry and its sidecar. Nothing in research-index.json, the dashboard pages, briefs/ or that repo's CLAUDE.md was touched. Branch and disposition on the final line.

Final disposition (recorded after /ship). This repo: MERGED e0839df main — scripts/verify.sh passed, git merge-base --is-ancestor HEAD origin/main confirms main carries today's brief. Run-branch deletion was refused by the sandbox proxy, which is policy, not failure. research-evidence: pushed to claude/modest-faraday-64fmka, verify: ok, and the promote-claude-branches workflow folded it into that repo's main as 276ffe7 before this run closed (the Stafford commit is 1c0b9b2). Its main was never pushed directly. No OPS flag carries into tomorrow. Brief emailed to Brandon.