Stafford

← 2026-10-10 · all briefs · 2026-10-12 →

Stafford Brief — 2026-10-11 (Sunday)

OPS: clean. Base 9d09d5f; yesterday's run merged. research-evidence: 2 KB files written and 2 entries staged directly, pushed to branch claude/modest-faraday-88scr7 at 330523c and MERGED into that repo's main by its promote-claude-branches workflow (confirmed 8af7e35). Its main was never pushed directly, per the division of labor. This repo: MERGED fac4f41 — the promote workflow folded the session branch in at 4a2541c while the direct push was in flight, so the direct push was rejected non-fast-forward and the proof step passed anyway. Both dispositions verified by ancestry, not assumed. Today's 12:00 scan landed (1 staged, 11 skipped) and its brief stands: the INSEAD skill-deprioritization paper, Tarakci and strategic consensus, the new Borrowed Apparatus pattern, the Purpose-instrument table. None of it is repeated here. Read it at research-evidence/briefs/2026-10-11.md.

This brief does one thing. I ran the query the base has been calling its highest-value follow-up since 17 September, and the answer is the bad one. The ledger named it, named the right candidate to check first, labelled that candidate "the most dangerous," said it needed no decision from Brandon, and then did not run it for twenty-four days. It ran today. Two of the three arguments defending the framework's originality claim are false, and the gate everything else waits behind cannot be written on its current argument.

📖 TODAY'S READ

Somebody already integrated the change-management literature into a five-domain diagnostic framework. They did it in 2009, from nineteen separate theories, and it has been in use ever since.

Damschroder, Aron, Keith, Kirsh, Alexander & Lowery, "Fostering implementation of health services research findings into practice: a consolidated framework for advancing implementation science," Implementation Science 4:50, 7 August 2009. Known as CFIR. Read from the open-access full text this session.

What it is, in plain terms. The authors took nineteen existing theories of why organisations fail to put new practices into use, found they overlapped and competed, and merged them into one structure: five domains, thirty-seven named conditions. Twelve of the thirty-seven describe the organisation receiving the change. Eight describe the change itself. That ratio is the Technology Illusion argument, encoded as a taxonomy, seventeen years ago — the same intervention works here and fails there, and the reason is the organisation, not the intervention.

Why this is a problem rather than a curiosity. The relabeling objection in knowledge/thesis-challenges.md — logged 17 September, still CONTESTED — set its own losing condition in writing: a prior work that treats three or more of the five breakpoints as a single interacting system of failure conditions, with a diagnostic attached. Three of the defences in that entry were doing the work. Two of them are now false:

And it names one of the breakpoints outright. The construct inside CFIR's inner setting is called "organizational incentives and rewards." Incentive Fragmentation does not merely have a 1975 antecedent in Kerr; it has a named condition inside a validated diagnostic framework in continuous use.

The replacement argument is narrower, and it is better, because the prior art makes it for him. CFIR's authors say their framework presents its conditions "without depicting interrelationships, specific ecological levels, or specific hypotheses" and "does not specify the interactions between those constructs." Five Breakpoints is a claim about exactly that — the interaction and the sequence, that treating one condition while ignoring the others is why transformations fail. Brandon does not have to assert that the interaction claim is his contribution. The best integration anybody built declined, in print, in 2009, to make that claim about itself. Quote the sentence and the concession becomes the argument.

Then the half nobody predicted, which is the best fact in the find. I went through the construct list looking for the other four breakpoints. CFIR has no condition for structural process friction — the nearest thing is "planning" — and no condition for sustainment or loss of momentum at all. The word "maintenance" appears once, as a reference to a different framework. Process Friction and Momentum Mirage are outside the taxonomy of the most comprehensive determinant framework the adjacent field has. That is a positive originality claim about two of the five, and a reader can check it against the published construct list in ten minutes and find it holds. It replaces a blanket claim that does not.

So what, in one sentence: Tier 2 — the gate in knowledge/master-plan.md that every other asset waits behind — has to concede an integration it was planning to claim, and in exchange it gets a quotable disclaimer from the prior art and two verifiable gaps. Publishing it on its 17 September argument would have put two refutable sentences in front of the audience best equipped to refute them. File: knowledge-base/damschroder-cfir-consolidated-framework-implementation-research-2009-08.md.

🔭 THINKER TO WATCH

Laura J. Damschroder, and behind her the implementation-determinants tradition — because this is the second day running that the field with the apparatus turns out to be the same field.

Damschroder is a research investigator at the VA Center for Clinical Management Research in Ann Arbor. CFIR is hers. It has been updated (2022) and extended with an Outcomes Addendum separating did we implement it from did it work.

The reason to watch her is not CFIR. It is that she is the third implementation scientist this base has run into in four days and the base keeps arriving from a different direction. 8 October: Aarons' LOCI trial, from implementation science, as prior art for Brandon's remedy. 10 October: Fixsen's fidelity framework, from implementation science, as the apparatus management lacks for "was the change actually delivered?" Today: Damschroder's CFIR, from implementation science, as prior art for the framework itself. Three different searches, three different questions, one field. The field has been asking Brandon's questions, with instruments, for twenty years, in a literature management does not read.

Brandon's differentiation, stated narrowly enough to survive contact. CFIR tells you which conditions to look at. It explicitly refuses to tell you how they interact, in what order they fail, or which one to fix first — and it has no condition for process friction or momentum at all. Five Breakpoints is a claim about interaction, sequence and remedy priority over a smaller set that includes two things CFIR does not name. That is defensible. "Nobody built the diagnostic" is not, and must not be written again.

🏢 CASE IN THE WILD — none, and I am not stretching today's material into one

Today's find is a 2009 methods paper and a 2022 scoping review. Neither is an organisation. The scan called this the first of three scans in its own tripwire — if the next two also return no case, the protocol rather than the market is the thing to change. I have nothing to add to that verdict today and I am not going to pad it; I will note only that the scan's own McKinsey 503s point at a fetch problem, which is a different remedy, and that establishing which it is before the tripwire fires is worth more than the tripwire.

⚙️ FOUR FORCES CONCLUSION

The scan's table this morning said Purpose is the one Force with a validated instrument nobody has pointed at AI. Today adds the Force that is worse off than Purpose: Momentum has no instrument to point.

The scan set the four side by side on does an instrument exist, and has anyone run it on AI? and put Purpose alone in the "instrument exists, unused" box. Today extends that table by one row, and the new row is emptier. Implementation science is the field that owns the question why did this change not take? Its most comprehensive determinant framework has thirty-seven named conditions. None of them is sustainment. The word "maintenance" appears only as a pointer to RE-AIM, a different framework built for a different purpose.

In plain terms: Purpose has an instrument nobody uses. Momentum does not have one to use, even in the field that should own it.

What that changes for the paper-2 argument. It inverts how the five breakpoints should be ranked by evidentiary risk. The intuition has been that Technology Illusion and Incentive Fragmentation are the exposed ones, because they have the clearest antecedents — Toyama and Kerr. The opposite is true for originality. Those two are named conditions inside CFIR, which makes them the most heavily prior-arted of the five. Process Friction and Momentum Mirage are the two the adjacent field never named — so they are where the framework's genuinely novel contribution lives, and they are the two the base has the least evidence for. The breakpoints most defensible as original are the ones Brandon can least support with data. That is worth knowing before the differentiation essay chooses what to lead with.

💭 OPEN QUESTION

If the field that professionally studies why changes fail has no construct for momentum, is Momentum Mirage the framework's strongest original claim or its weakest?

Not rhetorical, and it cuts both ways hard. The strong reading: twenty years of implementation research built a thirty-seven-condition taxonomy and never named the thing where it dies, which is a real hole and Brandon has a name for it. The weak reading: a field that measures things did not include a construct it could not measure, and Brandon included it because he does not have to measure it. Sustainment is defined by what happens after the observation window closes, which is exactly why a determinant framework built for prospective assessment would leave it out.

Those two readings predict different things and one is checkable. If the strong reading holds, the sustainment literature exists somewhere outside implementation determinants — programme evaluation, the "de-implementation" and sustainability literature — with instruments, and nobody has connected them. If the weak reading holds, the sustainment literature is thin everywhere and thin for a reason, and Momentum Mirage is the breakpoint most vulnerable to the McAdam standard: a named failure mode with no measurement and no outcome data, which is precisely the indictment that challenge makes of the whole corpus.

The query that separates them has a natural owner. The scan has already pre-registered Momentum Mirage against programme-evaluation and sustainment literature as the Borrowed Apparatus pattern's second place to look. It should be moved to first. It is no longer just a pattern member hunt — it decides whether one of the five breakpoints is the framework's best original claim or its softest target.

🥊 CHALLENGE

The strongest argument against the thesis today is the one the base wrote down about itself on 17 September and then did not check for twenty-four days.

Stated at full strength, as a reviewer with implementation-science training would put it:

Your five conditions are a subset of a thirty-seven-condition framework published in 2009, synthesised from nineteen prior theories, in continuous use for seventeen years, with a published interview guide and a diagnostic protocol. Two of your five are named constructs in it. Your claim to have integrated non-communicating literatures describes what Damschroder et al. did, not what you did. Your claim that nobody supplies a test an organisation can run against itself is refuted by the instrument's existence. And the framework has been applied to AI implementation — prospectively on clinical AI since about 2024, and retrospectively across nineteen deployments in a 2022 review.

Every sentence of that is true. It is logged in full as an amendment to The relabeling objection in knowledge/thesis-challenges.md.

Why it does not win, and the answer is three specific things rather than a posture.

  1. CFIR disclaims the interaction claim in its own voice — "does not specify the interactions between those constructs." Brandon's framework is a claim about interaction and sequence. The two are not competing for the same ground.
  2. Two of the five breakpoints are not in CFIR at all. No construct for structural process friction, none for sustainment. Checkable in ten minutes against the published list.
  3. Speed. CFIR is 2009. Nothing in it is about rate, and the machine-speed half of the thesis is untouched by it.

And one limit on the objection's new AI arm, which is also what rescues this morning's pattern. The review that applied CFIR to nineteen AI deployments — Chomutare et al., IJERPH 2022-12-06 — records that none of the nineteen primary studies used CFIR. The reviewers coded it on afterwards. So the instrument's non-use is real; it is just specific to the primary literature, and specific to management rather than health services research. "Nobody has applied a multi-domain diagnostic framework to AI implementation" is false and must not be written. "The management AI literature does not use the instruments the adjacent fields have, and health services research does" is true, narrower, and better — it makes management's non-use a choice rather than an impossibility, which is a sharper charge.

Status: stays CONTESTED, and the entry is now harder to write than it was yesterday. The old defence was broad, memorable and false. The new one turns on a disclaimer sentence in a methods paper and two absent constructs. It is correct and it is more work. That is the honest trade and it is still the right one — the alternative was publishing the broad version and being corrected in public.

📈 PATTERN BUILDING — the scan's new pattern needs a scope bound today, on day one, and it survives with it

No new pattern from me. The Borrowed Apparatus pattern the scan opened this morning at two members is the right pattern, and today's work both threatens and sharpens it. The scan pre-registered its own kill test: if any AI-and-organisations study is found that does apply one of these instruments at item level, the pattern is about this scan's search coverage, not about the field, and it gets retired.

I nearly fired that kill test and then found the detail that stops it. CFIR has been pointed at AI — prospectively on clinical AI from roughly 2024, and retrospectively across nineteen deployments by Chomutare et al. On the face of it, the instrument is in use on AI and the pattern dies on day one.

It does not die, because of one sentence in the review: none of the nineteen primary studies used CFIR. The reviewers applied it retrospectively as their own coding scheme. That is neglect in the primary literature, confirmed in the field most favourable to the instrument — the field that invented it. The pattern survives and gets a dated, defensible boundary instead of a blanket claim. Full proposal in Routing.

🪟 WINDOW FLAG — none

Nothing time-priced opened. A 2009 methods paper is argumentative value, not urgency, and master-plan hard stop 2 means nothing in the differentiation class ships before Tier 2 regardless.

🧱 ASSET SUGGESTION — no new asset, and today is the clearest case yet for why the freeze is right

Hard stop 1 is unconditional; highest number in the ledger stays 18. Today's material belongs inside asset 12, which is the Tier 2 gate, and the amendment is filed there rather than beside it.

The pitch, if it helps Brandon see the shape: "Five ideas I didn't invent, one framework that got there first, and the sentence its authors wrote that makes room for mine." Conceding CFIR is the Paper 3 seventy-percent move on better material — and the second instance is what makes it a practice rather than an incident.

Filed in knowledge/asset-suggestions.md: the two false novelty claims struck from the argument of record, the replacement spine, the two outside-the-taxonomy breakpoints, the one sentence the section must not contain, and the honest note that four of the five named candidates are still unrun — Beer & Nohria, Burke–Litwin, McKinsey 7-S, Pettigrew. Burke–Litwin is next and must run before drafting, because it is a causal model with arrows between twelve variables, and that is the one shape that would also threaten the interaction claim now carrying the whole section. Until it is checked the replacement spine is provisional, and I have marked it so.

🔍 SOURCES & THREADS

Exploratory slot: the integration search, logged in knowledge/query-log.md with the full yield. Not duplicative of the scan's pre-registered 10-12 slot (Process Friction against process-mining instruments), and squarely the prior-art verification CLAUDE.md operating rule 2 puts in Stafford's lane.

Yield: 2 KB files, 2 entries staged in research-evidence directly.

Both carry the instrument field. CFIR is recorded as model_theory with no sample and no estimation, because it is a conceptual synthesis that establishes no predictive validity — a point worth holding, since its diagnostic authority comes from use and consolidation, not from any demonstration that it predicts outcomes. The Chomutare review's unit of analysis is the published report, so what it measures is what authors narrated, and its window (2015–2021) is wholly pre-generative.

Read bounds, stated rather than buried. The CFIR primary is ~134,000 characters and I read the first ~100,000; nothing above rests on the remainder, and the 2022 updated CFIR construct list was not checked — every construct-level claim here is about the 2009 original, including the two absences. The prospective clinical-AI CFIR studies (BMJ Open 2024, JMIR 2026, the 2026 CDSS review) were reached through search summaries only and not fetched; they bound the pattern's scope and must be verified against the primaries before that bound becomes load-bearing in a draft.

Threads: 0 due, 0 checked, 0 opened, 0 closed. Verified independently rather than taken from the scan: no Next check date in story-threads.md falls on or before today, and the next is 10-15. The scan handled the two it moved out of cycle.

Candidates: 1 addition proposed (below), 0 promoted, 0 expired.

One thing I did not do and will not claim. Four of the five integration candidates named on 17 September are still unrun. One slot, one candidate, and I spent it on the one the ledger called most dangerous. That was right, and it leaves the job three-fifths undone.

Routing

🔴 personal — the real finding today is not CFIR, it is that the base knew about CFIR for twenty-four days

The 17 September challenge entry said this search was "the single highest-value follow-up from today." Asset 12 named CFIR first of five and called it "the most dangerous." It added that the search "does not require his call and should run in the next available exploratory slot." Twenty-four slots ran. This was not one of them.

And CFIR was never undiscovered. It is cited twice inside this base already — bmc-hsr-miake-lye-readiness-assessments-review-2020-02.md maps 1,370 readiness-assessment items onto its domains, and implementation-science-helfrich-orca-predictive-validity-2011-08.md records Helfrich's 2021 mapping of ORCA to it. The base has used CFIR as a coordinate system since 30 August while the ledger carried "check CFIR first" unexecuted. On 8 October the pipeline ingested LOCI from implementation science. On 10 October it ingested Fixsen's fidelity framework from implementation science. It spent four days inside the field that owns the framework most threatening to the originality claim, and did not open it.

This is the same failure shape as the proposal channel, now at forty-two days, and the master plan already named the mechanism: an exception clause that is always satisfiable is not a freeze. A follow-up with no owner and no date is decoration. The fix is one line and it is the same fix in both cases: a named item with a date, not a label. I would put it in the challenge-ledger convention — any entry whose "what it would take for this objection to win" names an unrun search gets a due date on the line, and an overdue one leads the brief. Today's would have led the brief on 18 September.

🔴 strategy — Tier 2 cannot be drafted on its current argument, and this is the cheapest possible moment to learn that

The gate asset's three novelty claims are down to one-plus-two-better ones. Nothing in the differentiation class ships before Tier 2 (hard stop 2), so this blocks everything — and it blocks it at the draft stage rather than after publication, which is worth a great deal. The amendment in knowledge/asset-suggestions.md is written and the replacement spine is there. One thing is Brandon's: whether to run Burke–Litwin before drafting. My position is yes, and it is the only remaining real risk to the section — a causal model with arrows between twelve variables would contest the interaction claim, which is now the load-bearing one. One slot. I will take it tomorrow unless the scan's pre-registered Process Friction slot is judged more valuable, and I would rather be told than guess.

🟡 strategy — the originality ranking of the five breakpoints is backwards and should be fixed before the essay picks a lead

Technology Illusion and Incentive Fragmentation feel like the exposed ones because Toyama and Kerr are famous. They are the two with named constructs inside CFIR — so they are the most heavily prior-arted of the five. Process Friction and Momentum Mirage are the two implementation science never named, which makes them the most original and the least evidenced. Lead the differentiation essay with the two the adjacent field does not have, and be honest that they are where the evidence is thinnest.

🟡 strategy — four decisions were overdue to Brandon as of this morning and I am not re-listing them

The scan routed four (the Confidence Without Substrate binary, EY's credibility listing, the irregular tag, and narrowing Category 6 to named-journal checks). I argued all but the last on 10-10 and my positions have not changed; re-arguing them daily converts a decision queue into noise. On the new fourth: narrow Category 6, yes. Two consecutive scans of already-held returns, and today's own best find came through a named-journal route — as did mine, which came from a framework name rather than a keyword sweep.

PROPOSED-UPDATE blocks for research-evidence (three)

1. knowledge/emerging-patterns.md — The Borrowed Apparatus (opened today, 2 members). Count stays at 2. Record the kill test as RUN AND SURVIVED, with the scope narrowed on day one.

The pattern's kill test is "any AI-and-organisations study that does apply one of these instruments at item level." Run today against CFIR and it initially appears met: CFIR has been applied to clinical AI prospectively from roughly 2024 (BMJ Open 2024 imaging AI; JMIR 2026 ICU nursing; a 2026 JMIR Medical Informatics systematic review of AI clinical decision support) and retrospectively across nineteen AI deployments 2015–2021.

It is not met, on one sentence in the retrospective review. Chomutare et al., IJERPH 19(23):16359 (2022-12-06) — knowledge-base/chomutare-ai-implementation-healthcare-theory-based-scoping-review-2022-12.md, staged today — screened 2,784 studies, included 19 implemented or regulator-approved healthcare AI systems, and records that none of the 19 primary studies used CFIR. The framework is the reviewers' coding scheme, applied afterwards. Neglect in the primary literature, confirmed in the field that invented the instrument.

Required scope narrowing, on day one, because the blanket version is false: the pattern is about the management and organisational AI literature, not about all AI research. Prospective instrument use does appear in health services research on clinical AI from about 2024. This makes the pattern better, not weaker — it establishes that borrowing these instruments for AI is demonstrably possible, so management's non-use is a disciplinary choice rather than a measurement limitation. Bound carried from birth: the prospective clinical-AI studies were reached through search summaries only and not fetched; verify against the primaries before the bound is load-bearing.

Also add, as a note rather than a member: today's find is the pattern's uncomfortable third case — Breakpoint 2 (Strategic Disconnection) and Breakpoint 4 (Technology Illusion) both now have adjacent-field apparatus, and the whole framework has an adjacent-field integration in CFIR. The pattern's scope is instruments for a diagnostic question; CFIR is an instrument for the whole diagnostic, which is a different and larger object. Recommend logging it as a scope note, not a member, pending the scan's judgment — counting it would change what the pattern is about.

2. knowledge/emerging-patterns.md — move the Borrowed Apparatus pre-registered third member hunt from Process Friction to Momentum Mirage, and change what the result means.

The scan pre-registered two places to look, in order: Process Friction → operations management and process mining; Momentum Mirage → programme-evaluation and sustainment literature. Recommend reversing them, for a reason the scan could not have had this morning. CFIR has thirty-seven constructs and no construct for sustainment — "maintenance" appears only as a reference to RE-AIM. So the implementation-determinants field, which owns why changes do not take, does not name momentum loss.

That converts the Momentum Mirage query from a pattern-member hunt into a decision about one of the five breakpoints. If a validated sustainment instrument exists in programme evaluation and no AI study uses it, the pattern gains its strongest member and Momentum Mirage gains an apparatus. If no such instrument exists anywhere, that is a finding about the breakpoint rather than about the field: Momentum Mirage would be a named failure mode with no measurement in any adjacent literature, which is the McAdam standard's indictment applied to a single breakpoint, and it should be known before the differentiation essay leads with it. Either outcome is worth more than the Process Friction slot.

3. knowledge/source-candidates.md — add one channel candidate, 1 signal, cadence continuous.

Implementation Science (BioMed Central) and the implementation-determinants literature — first seen 2026-10-11, surfaced by the integration search returning CFIR. Opened as a coverage finding, like this morning's Management Science slot, and the shape is the same and worse. This is the third implementation-science find in four days arriving from three unrelated directions: Aarons/LOCI on 10-08 (prior art for the remedy), Fixsen on 10-10 (the fidelity apparatus management lacks), Damschroder today (prior art for the framework). A field that has produced three load-bearing finds in four days is not a candidate, it is a monitored source — but the list-maintenance rule forbids first-contact promotion and I am not going to break it unilaterally, so: probation slot, and the promotion test is already arguably met three times over. Brandon's call whether to override. Expiry 2026-12-10 on the continuous clock, which this will not reach.