Stafford

← 2026-10-13 · all briefs · 2026-10-15 →

Stafford Brief — 2026-10-14 (Wednesday)

This morning's scan cleared a sentence for you to use in a client room, and the sentence is wrong. The paper it rests on refutes it, two pages past the figure the scan read it from.

The sentence: "the organizational prerequisite is specific to AI — the same management practices that predict whether a firm adopts AI predict nothing about whether it adopts robotics or specialised software." It was registered as a one-sentence asset with "caveats mandatory," against the objection you have been carrying for months — well-run companies are just better at everything.

I spent the Wednesday falsification slot on it and read the appendix. Four problems, all inside the paper. The decisive number is estimated twice with opposite signs. Three of the four nulls sit on technologies almost every firm already has or almost none has. The authors' own alternative explanation is that AI was simply new. And a footnote concedes that their measure cannot see the margin your argument is actually about.

There is a replacement sentence in the same paper and it is better — stronger, in both of the paper's specifications, and the authors' own words rather than an upgrade of them. It is below, and it is the one thing in this brief to take into a room.

One thing got stronger, too, and it is yours. The sub-component that holds up best in the whole paper is targets — how goals are set and communicated — at the 1% level, where the data-tracking measure only clears 10%. That is Strategic Disconnection, and it is the best-evidenced part of the result rather than a secondary note.

I did not re-run the sweep and I am not repeating the scan's brief. Read it first (research-evidence/briefs/2026-10-14.md) — it is a strong one, and the Makridis find is real. Everything here is the delta.

📖 TODAY'S READ

Not a new source — the appendix of this morning's source, which changes what can be claimed from it.

The scan filed Coyle, Nguyen, Poquiz & Riley, More than just plug and play (Bennett School of Public Policy, Cambridge / ESCoE, 7 May 2026) and called the placebo test "the paper's most useful contribution." The file says the scan read the full text including tables 6–8 and the robustness section. It read the placebo from the paper's figure. The numbers behind the figure are in Appendix Table B.5, and they do not say what the figure appears to say.

Problem 1 — the number the whole argument depends on is estimated twice, with opposite answers. The placebo only means something if management quality does not predict buying non-AI technology in general. That is exactly what "better at everything" asserts. The paper tests it directly:

Different samples, no reconciliation anywhere in the paper, and the first of the two is plotted in the very figure the placebo claim comes from. A reviewer who opens the appendix finds both numbers and asks which one you believe.

*Technical detail: Table B.5 reports t-statistics. Management Score 2020 → cloud 0.115 (3.39); robotics −0.018 (−0.28); specialised software 0.008 (0.35); specialised equipment 0.030 (1.44); any non-AI technology 0.285 (7.08); AI controlling for general technology adoption 0.245 (3.14); AI among adopters of other technology 0.265 (3.29). Matching appendix ATETs: any other technology −0.004 (−0.50, n=6,640); specialised software −0.004 (−0.41, n=5,668); specialised equipment −0.005 (−0.34, n=4,012).*

Problem 2 — three of the four nulls have almost nowhere to move. The paper's own footnote 19 gives the adoption rates: robotics 4%, specialised equipment 36%, specialised software 61%, cloud about 70%. You cannot learn much about what distinguishes firms from an outcome that 4% of them do, or one that 61% of them already do. The authors name the ceiling problem for cloud themselves, and note that the UK is the only G7 country outside the top 20 for robot density — so the robotics null is a fact about British industry, not about management.

Problem 3 — the authors' own alternative explanation is that AI was new. From their discussion: the maturity of the other technologies "implies cost reduction, peer learning and vendor simplification, reducing the need for more structured management compared to nascent AI products." On that reading, structured management is what it takes to adopt something immature — and the requirement fades as vendors, consultants and peer experience fill in. That is a diffusion-curve finding with an expiry date, not a standing claim about what AI demands. It is the strongest objection to Claim 5 I have found and I have logged it; see CHALLENGE.

Problem 4 — footnote 20 concedes the margin you actually care about. Their technology measures are yes/no: did the firm buy it. The footnote: for technologies that are industry norms, buying *"might largely be determined by operational necessity rather than management quality. Management practices may nonetheless influence the depth and sophistication of deployment, a margin our measure does not capture."* So the placebo compares whether firms buy a mature technology against whether firms buy a new one. It is silent on whether management governs what either is worth once bought — which is your whole argument.

What to do instead, and this is the take-away. Drop the null. Use the conditional result from the same paper:

Take two firms that already run comparable technology. The better-managed one is still the one that takes up AI.

Technical detail: 0.265 (t=3.29) controlling for other technology adoption; 10.4 percentage points (t=3.55) by matching among firms that had already adopted at least one other technology — larger than the 8.7pp (t=2.79) among firms that had adopted none. The authors' words: "even among firms with comparable technology portfolios, superior management practices continue to predict AI adoption."

Why it is strictly better. It answers "better at everything" by holding the technology portfolio fixed instead of claiming management does nothing elsewhere. It survives both specifications, where the placebo does not. It does not depend on robotics prevalence, the ceiling effect, or which any-tech estimate is right. It is the authors' own sentence. And in a room it is the better move: a null invites "so you're saying management doesn't matter for anything else?" — the conditional version invites the client to picture two firms they know.

And the authors never claim what the sentence claims. Theirs is "these complementarities might differ across individual technologies… The framework should not treat 'technology' or even 'IT' as monolithic." That is careful and defensible. "Specific to AI" is an upgrade you would be making on their behalf.

🔭 THINKER TO WATCH

The scan's entry on Coyle is right and I am not restating it. One correction and one strengthening, both from the tables.

The correction. The scan's differentiation note says "Brandon owns the outcome half and the decomposition." The first half holds — there is no outcome variable anywhere in the paper. The second half is now shakier, because the decomposition is where the paper is weakest in a way the scan did not see: the sub-component results run on 1,292 firms, and the placebo and matching results run on samples between 973 and 6,640. The paper's conclusions are drawn across specifications estimated on samples that differ by a factor of five, and nothing in it tells you which sample is the right one. That is not a flaw you can exploit — it is a reason to cite the paper for one result rather than for a structure.

The strengthening, and it is genuinely good news. The scan reported the sub-component weight as sitting on "targets and KPI count." Correct — and the order matters. Targets is the strongest result in the table and the only one at the 1% level; the data-tracking measure only clears 10%; the other two are null. So the single most robust thing in the best firm-level readiness instrument this base holds is whether the firm sets and communicates clear performance goals.

Technical detail: Table 7, n=1,292 each, t-statistics in parentheses — continuous improvement 0.096 (1.17), null; KPI 0.219 (2.16), p<0.10; targets 0.165 (3.50), p<0.01; employment practices 0.083 (1.19), null.

That is Strategic Disconnection, measured, at firm level, as the load-bearing component — not as the secondary observation the scan's framing makes it. It is a better sentence for you than the placebo ever was: the one management practice that most reliably separates firms that take up AI from firms that do not is whether they can state and communicate what they are trying to achieve. Caveats travel: adoption, not outcome; unrefereed.

🏢 CASE IN THE WILD

Nothing to add, and the scan's HubSpot entry is the best piece of work in its brief. Its core observation — a flattening of exactly the kind résumé-based instruments detect, with the AI explanation denied on the record by the CEO, and no résumé panel able to tell the difference — is a real qualification of the ESCALATED Résumé-Panel Consensus genre and it is the scan's, not mine. Its treatment of the denial as on the record but not disinterested is the right register. No correction, no addition. The thread it opened (12-15, timed past Q4 FY2026) is correctly dated.

⚙️ FOUR FORCES CONCLUSION

The scan's new gap statement is the right one and it is this morning's best piece of thinking. One amendment, in your favour.

Its statement: Purpose and Commitment have each been measured once as a moderator — individual level, self-reported outcomes, unrefereed. Capability five times at firm level against realized productivity. Momentum never. Hold that. It is narrower and harder to attack than the version it replaces, and it names the study nobody has run.

The amendment. The scan credits Purpose with exactly one measurement (Makridis, at individual level). There is a second, at firm level, and it is in the paper the scan filed the same morning — the targets sub-component above, significant at the 1% level against AI adoption. It is not a moderator result, so the gap statement stands as written; it is a determinant result. But it changes the shape of the asymmetry:

Purpose now has two measurements: one as a determinant of whether AI arrives (firm level, 1% level, Coyle et al.) and one as a moderator of whether it pays (individual level, self-reported, Makridis). Capability has five as a moderator. Commitment has one. Momentum has none, on either side.

Why the distinction earns its place rather than complicating a clean line. The determinant/moderator split is the same boundary that limits both of today's finds, and stating it this way makes the missing study specifiable in one sentence: nobody has measured a Purpose variable at firm level against a realized AI outcome. The two papers stand on either side of that hole. Momentum remains unmeasured on either side, which after six weeks of looking is itself worth saying out loud — the Force your own framework treats as the one that decays is the one nobody has instrumented at all.

💭 OPEN QUESTION

If organizational readiness is what it takes to adopt something immature, what exactly is left of Claim 5 in 2029?

This is the maturity objection asked forward rather than defensively, and it is not rhetorical — I do not know the answer and neither does the literature.

The uncomfortable version. Every readiness result in this base is measured in a window where AI was new: Coyle 2020→2023, Pulito 2015→2023/24, Makridis 2023→Q1 2026. If the mechanism is novelty — no vendor playbook, no peer to copy, no consultant who has done it, so the organization must supply the structure itself — then readiness is a transitional requirement and the measured coefficients should shrink as the ecosystem matures. That is a falsifiable prediction and the people who could test it are the ones who produced today's find: the same group, the 2026 MES wave, the same regression. If the coefficient falls, Claim 5 is a claim about earliness.

The version worth building on. Perhaps the maturing ecosystem absorbs the technical cost of adoption and never touches the organizational one — because what vendors can commoditize is implementation, and what they cannot commoditize is deciding which outcome you want and making your managers mean it. On that reading the coefficient on targets should be the one that persists while the capability-flavoured ones decay, and the composite management score should look weaker over time precisely because its technical components stop mattering.

That is a prediction you can date and publish, and it is the shape of argument the master plan says you are short of. Note what it does: it converts the strongest objection to Claim 5 into a test with a named data holder, a named wave, and a sign. The reason to write it down now is that it costs nothing now and cannot be written after the 2026 wave reports.

🥊 CHALLENGE

The maturity objection — "your organizational prerequisite is a finding about novelty, and it expires." New, OPEN, targets Claim 5. Full entry: knowledge/thesis-challenges.md, 2026-10-14.

Stated at full strength: the readiness result everyone in this base has been treating as a fact about what AI demands of an organization is equally consistent with a fact about what being early demands of an organization. The comparison technologies in today's placebo were mature; AI was not. The authors say so themselves — maturity brings "cost reduction, peer learning and vendor simplification, reducing the need for more structured management." If that is the mechanism, Claim 5 is true in 2023, weaker in 2027, and empty at saturation, and a framework sold as a durable account of AI-era organizational design has been built on a diffusion-curve artifact.

It does not need any evidence from outside the paper, which is what makes it expensive. It is assembled entirely from the paper's own discussion section, its footnotes 19 and 20, and the two incompatible estimates in its appendices. That also means it would have been found by any careful reader on the other side of a conversation.

What would make it win outright: either the management–AI adoption coefficient shrinking in the 2026 MES wave (the diffusion-artifact prediction, directly testable by this group), or anyone measuring management practices against the depth of deployment of a mature technology — footnote 20's margin — and finding management predicts it as strongly as it predicts AI adoption.

What would defeat it: a management-practices measure against a realized AI outcome, with the comparison technology measured on the same margin. That is the unoccupied cell the scan's own Four Forces conclusion named this morning, and the amplification thread's criterion (a). Nobody has run it.

One thing the objection does not touch, and this matters for how much it costs you. It is an attack on why readiness predicts adoption. It is not an attack on the conditional result — two firms with comparable technology, the better-managed one adopts AI — which holds whatever the reason, and is not an attack on anything in Papers 1–6. The exposure is to the strong framing, not to the finding. Note also that the base already contains a result against the strong framing: Bloom, Garicano, Sadun & Van Reenen (2009) measured ERP moving plant managers' decision authority one way and data networks moving it the other. ERP plainly made organizational demands. "Other technologies make no organizational demands" contradicts an entry this base already holds.

📈 PATTERN BUILDING

Nothing crossed a threshold. One pattern created this morning needs amending on the day it was created, and the amendment goes to its reason for existing.

The scan opened The Predetermined Moderator (2 members, not escalated): the organizational variable measured before the AI decision — Pulito (skills 2015 → adoption 2023/24) and Coyle (management 2020 → adoption 2023). The construct is sound and the pattern deserves to exist.

But the scan's stated reason for admitting member 2 is the one thing today's verification undermines. Its words: "Pulito removes reverse causality but not 'well-managed firms are better at everything'; Coyle et al. remove that too, with the technology-specificity null. Predetermination plus a placebo test is a materially stronger position than either alone, and no other readiness pattern here has either property."

The placebo does not carry that weight — see TODAY'S READ. What Coyle does carry is better suited to the pattern anyway: the portfolio-conditional estimate. Among firms already running comparable technology, management still predicts AI adoption, in both specifications. That rules out "better at everything" by holding the confound fixed rather than by asserting it is absent, which is the stronger design and is not vulnerable to anything in today's sweep. So the member stays, the pattern's distinguishing claim stays true, and the sentence supporting it has to change. PROPOSED-UPDATE in Routing.

Why I am not proposing a recurrence change. Two members, both predetermined, both with adoption on the left. The pattern's own stated limit — "the left-hand side is ADOPTION, never an outcome; neither member supports Claim 3" — is correct and is the real constraint. Today adds a second limit worth writing into it: both members are measured in a window where AI was new, which is the maturity objection arriving inside the pattern rather than against it.

🪟 WINDOW FLAG

None — twenty-fifth consecutive, and I agree with the scan's reasoning. A May working paper and an April working paper have no closing date. The one thing in today's material that is dated is in OPEN QUESTION, and it is a window on a prediction, not on a publication: the 2026 MES wave will eventually report, and a dated forecast about which coefficients decay is only worth something before it does. That is not a publication window and I am not logging it as one.

🧱 ASSET SUGGESTION

No new asset — a withdrawal and a replacement. Full spec: knowledge/asset-suggestions.md, correction dated 2026-10-14.

The one-sentence asset the scan registered this morning is withdrawn for the reasons in TODAY'S READ. The replacement — take two firms that already run comparable technology, and the better-managed one is still the one that takes up AI — is stronger on five counts and is the authors' own sentence. It goes wherever Claim 5 is argued: asset 2 (Tier 1.1, COMMITTED) if the readiness argument appears there, and paper 2's Claim 5 section.

It also belongs in asset 12, the Tier 2 gate, as a worked example rather than as an argument. The gate exists to say what the framework owes and what it does not yet have, and the master plan's case for it is the Paper 3 seventy-percent move — the concession is the credibility. Here is a 24-hour-old instance: this base read a placebo test from a figure, cleared a sentence for client use, and corrected it from the appendix the next day. That is the discipline the gate is claiming to have, demonstrated rather than asserted, and it costs two sentences.

Sixth asset the scan has registered in twelve days; fifth to resolve into an existing home. Hard stop 1 holds. Nineteen arguments, one committed, none shipped.

🎓 SPARRING NOTE

Not issued — twelfth consecutive. The 09-17 note is still the one live note and still has no written reply, so the cap holds and no new prompt is issued. Twenty-seven days.

Today's material parks against it and adds a fourth axis, which is the first one that is not about priority, method or contemporaneity: durability. Not "isn't this Kotter?", not "was Kotter ever evidenced?", not "a top-five journal derived your breakpoint five months ago" — but "your prerequisite might just be what being early costs, and the vendors are fixing it as you write."

That is the axis where the answer is strongest and the paragraph is shortest, which is why it is worth noting rather than adding to the pile: the response is the prediction in OPEN QUESTION — targets should persist while the technical components decay — and it is a dated, falsifiable claim rather than a rebuttal. A paragraph that answers a hostile reader with a testable forecast is a better Tier 2 opening than one that answers with prior art.

SOURCES & THREADS

Threads: 0 due to me, 0 run, nothing proposed. The queue is 10-15 (nothing — the scan moved both 10-15 threads, the EU AI Act item and the supervisor pattern, to 11-15 after checking them early), then 10-17 amplification theory, 10-20 the sampling-frame thread and Project OT, 10-28 information-routing retirement. No thread's next-check date had arrived that the scan did not cover, so I ran none. The 2026-09-07 cadence rule applies: re-checking before a date arrives produces log entries rather than evidence.

One note for the 10-17 amplification check, which is three days out and which the scan deliberately left on its date: its criterion (a) now has two partials and today's verification weakens the stated advantage of the second. The scan admitted Coyle as the better partial because of the placebo test. The portfolio-conditional estimate is the part that should carry it. Criterion (a) is still unmet for the unchanged reason — adoption, not outcome.

Candidates: none added, none promoted, one chase recorded as declined-for-now. The management-practices literature the falsification sweep surfaced — Bloom, Sadun & Van Reenen's Americans Do IT Better (people-management practices explain the US IT productivity advantage) and the Bloom et al. India management field experiment (computer use rises with management quality) — bears directly on the maturity objection and I deliberately did not ingest either. Both are abstract-grade here, both are 2010–2012, and the decisive evidence turned out to be inside the Coyle paper itself. A tight base beats a large one, and two old papers ingested to decorate a challenge that already stands on a primary is exactly the padding the standing rule forbids. Recorded as a named chase instead: if the maturity objection ever needs external support, Americans Do IT Better is the first fetch, because it measures management against IT returns rather than adoption, which is the margin footnote 20 concedes.

I support the scan's two live method decisions and have nothing to add to either. The CESifo promotion with IZA split off at 1 signal is the right call on the 10-06 Census precedent. The EDGAR method capture is the most operationally useful thing in its brief and the Category 5 replacement should be executed, not re-argued — that is now the second day running it has been recommended.

Exploratory slot: Wednesday falsification sweep, spent on the sentence the scan cleared for use this morning. Yield: 0 KB files, 0 staged entries, 1 new thesis challenge, 1 asset sentence withdrawn and replaced, 1 pattern amendment, 1 strengthening. Logged in full at knowledge/query-log.md. I took this over the 10-13 handoff (the ADAPTOVATE retrieval) deliberately: a sentence already cleared for a client room is a live exposure, a blocked PDF is not.

Method note worth keeping, because it generalizes beyond today. When a brief cites a null from a figure, open the appendix table. The number that decided whether this placebo test meant anything was two pages away and had the opposite sign. Today's reading order — figure → text → appendix tables → footnotes → discussion — found the refutation in the last three.

ADAPTOVATE: chased separately, null, and now diagnosed rather than merely failed. Twenty days owed, blocking asset 15's saturation argument. Both hosts return HTTP 200 at 275KB — it is not blocked, not 403, not paywalled. It serves a JavaScript application: strip the scripts and tags and 30 characters of text remain. No fetch-based tool in this pipeline will ever read it. Disposition: reassign to your browser, like the McKinsey 403, and stop spending slots on it. The distinction is new in this base and worth recording — every previous unreachable source was refused; this one is delivered and unreadable, which looks the same from a status code and needs the opposite response.

Routing

strategy — Stop using the sentence the scan cleared this morning; use this one instead. Take two firms that already run comparable technology — the better-managed one is still the one that takes up AI. Both of the paper's methods agree on it, the effect is larger among firms that already run other technology than among firms that run none, and it is the authors' own words. The placebo version ("predicts nothing for robotics or specialised software") is refuted by the paper's own appendix and should not go into an essay, a deck or a room. Caveats unchanged and mandatory: unrefereed working paper, and the outcome is adoption, not productivity.

strategy — Lead the readiness argument with targets, not with the composite. The most robust single result in the best firm-level readiness instrument this base holds is that firms which set and communicate clear performance goals are the ones that take up AI — at the 1% level, where the data-tracking measure clears only 10% and the other two components are null. That is Strategic Disconnection, measured, and it is the load-bearing component rather than a secondary note. It is also the better sentence for the practice: a client can act on "can you state what you are trying to achieve" in a way they cannot act on a management score.

strategy — The maturity objection is now the strongest standing attack on Claim 5, and the scan told you this morning to lead paper 2 with Claim 5. Those two facts need to meet before the paper does. The objection is assembled entirely from the source's own discussion and footnotes, so any careful opponent finds it. The answer is not a rebuttal, it is a dated prediction (see OPEN QUESTION): if readiness is the cost of novelty, the coefficients decay in the 2026 MES wave; if it is the cost of organizational ambiguity, targets persists while the technical components decay. Writing that down costs an afternoon, it is falsifiable, and it cannot be written after the wave reports. Logged OPEN at knowledge/thesis-challenges.md.

strategy — Momentum has never been measured, as a determinant or as a moderator, by anyone, on either side of today's two papers. Six weeks of this base looking at readiness instruments has produced five Capability moderators, two Purpose measurements, one Commitment measurement and zero for Momentum — the Force your own framework says is the one that decays. At some point that is either the framework's sharpest open claim or a sign the construct is not measurable as stated. It is worth one deliberate decision from you about which, rather than another week of it being the empty cell in a table.

market — Nothing from me today beyond seconding the scan: execute the EDGAR-for-Category-5 swap. Second consecutive day recommended, now with a verified working method rather than a complaint. One line in the scan's CLAUDE.md.

governance — Nothing. Eighteenth consecutive null sweep; the scan's decision to spend the slot elsewhere was correct and produced.

personal — The scan lists five decisions with you. I am adding one and it is cheap, and I am re-ranking by what each unblocks rather than by age. (1) Tier 0.2, the post-engagement measurement — now the terminus of four challenges (the McAdam standard, the sufficiency objection, yesterday's open question, and today's maturity objection, which can only be settled by outcome data). One page, one afternoon, the only time-gated item on the master plan, and every engagement that closes without it is a data point lost permanently. (2) New today, and it is fifteen minutes: approve the dated prediction in OPEN QUESTION — targets persists, technical components decay — or tell me you do not want a dated forecast in print. It converts the strongest objection to Claim 5 into a test, and it expires when the 2026 MES wave reports. (3) The 09-17 paragraph, twenty-seven days, now with a fourth and better-answerable axis. (4) Narrow Category 6 to named journal and preprint-listing checks — the scan reports today is the fourth consecutive scan where the keyword route produced the find and no listing was checked, and calls it a described failure mode of its own process rather than a preference. I agree. Then the scan's remainder: the Confidence Without Substrate escalation (ten scans), the CESifo/Management Science promotions (both reversible, I support both), the Verlinden coefficients (ten scans). And one retirement: the ADAPTOVATE fetch is not a Stafford task any more — it needs your browser, and I have stopped spending slots on it.

PROPOSED-UPDATE → research-evidence knowledge-base/coyle-nguyen-poquiz-riley-organisational-capital-ai-adoption-2026-05.md

Why a proposal and not a direct write. CLAUDE.md operating rule 3 permits me to add files to that repo's knowledge-base/, and its stated reason is "one file per source, so two writers do not collide." Editing the scan's own file on the scan's own source is that collision by definition. Four changes, all verified against the full text this session.

1. Add to ## Summary, after the placebo paragraph:

The placebo test does not carry the weight this entry originally gave it (verified against Appendix Table B.5 and the matching appendix by Stafford, 2026-10-14; the original reading was taken from Figure 2 and the body text). Four qualifications, all from the paper. (a) The decisive quantity is estimated twice with opposite signs and never reconciled. The placebo is informative only if management does not predict adoption of non-AI technology in general. Appendix Table B.5 column (5) estimates that effect at 0.285 (t = 7.08) — larger and tighter than the AI coefficient — while the propensity-matching appendix estimates the same quantity at −0.004 (t = −0.50, n = 6,640). The samples differ and the paper does not say which it stands behind; the LPM estimate is plotted in Figure 2 itself. (b) Three of the four nulls rest on outcomes with little variance: footnote 19 gives 2023 adoption rates of robotics 4%, specialised software 61%, specialised equipment 36%, with cloud at ~70%; the authors name the cloud ceiling effect and the UK's position as the only G7 country outside the top 20 for robot density. (c) The authors' own alternative explanation is technology maturity, not AI specificity — "the maturity of other technologies implies cost reduction, peer learning and vendor simplification, reducing the need for more structured management compared to nascent AI products." (d) Footnote 20 concedes the margin: the indicators are binary extensive-margin only, and "management practices may nonetheless influence the depth and sophistication of deployment, a margin our measure does not capture." What should be cited from this paper in place of the placebo is the portfolio-conditional result: management predicts AI adoption at 0.265 (t = 3.29) controlling for other technology adoption, and by matching at 10.4pp (t = 3.55) among firms that had adopted at least one other technology versus 8.7pp (t = 2.79) among those that had not. That holds the confound fixed rather than asserting its absence, and it appears in both specifications.

2. Replace in ## Five Breakpoints Intersection, the Technology Illusion rationale (currently "The placebo test is the strongest evidence in this base against the 'deploy it and the organization will follow' premise…"):

Technology Illusion: Among firms already running comparable technology, higher baseline management quality still predicts AI adoption — 0.265 (t = 3.29) with other technology adoption controlled, and 10.4pp (t = 3.55) by matching within existing technology adopters, which is larger than the 8.7pp among firms adopting nothing else. This is evidence against the "deploy it and the organization will follow" premise that holds the firm's technology portfolio fixed, so it cannot be answered by "well-managed firms are better at everything." Stated this way rather than via the cross-technology placebo (see the Summary qualification, Stafford 2026-10-14), because the placebo's informativeness depends on an any-technology estimate the paper reports with opposite signs in two appendices, and on nulls measured against outcomes at 4% and 61% prevalence.

3. Amend in ## Five Breakpoints Intersection, the Strategic Disconnection rationale — the sub-component ordering is the opposite of what the entry implies, and it favours this rationale:

Strategic Disconnection (strengthened, Stafford 2026-10-14, from Table 7): The targets sub-component — how performance goals are set, communicated and linked to incentives — is the most robust sub-component result in the paper at 0.165 (t = 3.50, p < 0.01), ahead of KPI tracking at 0.219 (t = 2.16, p < 0.10), with continuous improvement (0.096, t = 1.17) and employment practices (0.083, t = 1.19) both null, all at n = 1,292. The entry's original reading ("the decomposition points at measurement infrastructure specifically") inverts the ordering: the goal-setting dimension is significant at the 1% level and the measurement-infrastructure dimension only at the 10%. This is the strongest firm-level measurement of a Strategic Disconnection proxy in this base.

4. Add to ## Relevance to Brandon's Positioning, replacing the sentence "Brandon can cite one coefficient and one null and the tautology objection is closed":

Cite the coefficient and the conditional, not the null. "Take two firms that already run comparable technology — the better-managed one is still the one that takes up AI" closes the tautology objection in both of the paper's specifications and is the authors' own reading ("even among firms with comparable technology portfolios, superior management practices continue to predict AI adoption"). The null version — "predicts robotics and specialised software at zero" — must not be used; it is refuted by the paper's own appendix (Summary qualification above) and is the version a reviewer will reach for. A durability caveat now travels with this entry: the authors' own maturity reading makes the readiness result consistent with a transitional cost of early adoption rather than a standing organizational requirement. Logged as the maturity objection (stafford-research knowledge/thesis-challenges.md, 2026-10-14, OPEN) and currently the strongest standing attack on Claim 5. It is testable by this group on the 2026 MES wave.

PROPOSED-UPDATE → research-evidence knowledge/emerging-patterns.md, The Predetermined Moderator, member 2

Amendment (Stafford, 2026-10-14) — the member stands; the stated reason for admitting it does not. Member 2 was admitted because "Coyle et al. remove ['well-managed firms are better at everything'] too, with the technology-specificity null." The placebo does not support that — see the qualification proposed for knowledge-base/coyle-nguyen-poquiz-riley-organisational-capital-ai-adoption-2026-05.md: the paper estimates the any-non-AI-technology effect at 0.285 (t = 7.08) by LPM and −0.004 (t = −0.50, n = 6,640) by matching, three of four nulls sit on 4%/36%/61%-prevalence outcomes, and the authors' own alternative reading is maturity. Substitute the portfolio-conditional estimate as the member's distinguishing property: management predicts AI adoption at 0.265 (t = 3.29) controlling for other technology adoption and at 10.4pp (t = 3.55) by matching within existing technology adopters — which rules out the general-capability confound by holding it fixed rather than by asserting it absent, and survives both specifications. The pattern's claim is unchanged and arguably better founded; only its evidence sentence changes. Second limit, to be written in alongside the existing ADOPTION limit: both members are measured in windows where AI was new (2015→2023/24 and 2020→2023). If the mechanism is the absence of a mature vendor and peer ecosystem rather than a standing organizational requirement, the predetermination property is intact and the interpretation is transitional. The discriminating observation is a later wave: a falling management–adoption coefficient as national AI adoption rises. Logged as the maturity objection (stafford-research knowledge/thesis-challenges.md, 2026-10-14).

PROPOSED-UPDATE → research-evidence knowledge/sources.md, operational notes

New failure class, recorded because it is the first of its kind in this base (Stafford, 2026-10-14): delivered-and-unreadable. adaptovate.com/AIReport2026 returns HTTP 200 at ~275KB on both the apex and www. hosts — no block, no 403, no paywall — and contains ~30 characters of extractable text because the page is a client-rendered JavaScript application. A 200 is not evidence a source is retrievable. Every previously unreachable source in this base was refused (403, 503, paywall) and the correct response was a different route; this one is served and the correct response is a browser, so routing effort at it is wasted. Check extracted text length, not the status code, before recording a retrieval as blocked or as owed. Disposition on this specific target: reassigned to Brandon's browser after four attempts across twenty days; removed from Stafford's chase list.

OPS — reported on the first line of the final message. Run branch claude/keen-fermi-z202yb, based on origin/main at 837c45c.

research-evidence: nothing written, no branch pushed, and that is the correct outcome rather than a failure — for the second consecutive day, and for the same structural reason both times. Today's work produced no new source: the falsification sweep's decisive evidence was inside a paper the scan already holds and has already staged, and the two management-practices papers it surfaced were deliberately not ingested (see SOURCES & THREADS). So there was no KB file to copy and no entry to stage. The four corrections above are to the scan's own file on the scan's own source, which is exactly the collision that operating rule 3's one-file-per-source design exists to prevent, so they are proposals. The clone was synced read-only to origin/main (a5e05bf) and today's scan output consumed from it.

A note on the scheduled prompt, recorded so it stops recurring. Today's scheduling prompt states "You have no write access to research-evidence." That is the superseded belief CLAUDE.md operating rule 3 was written to correct — it is false, Stafford does have scoped write access to that repo's knowledge-base/ and staging path, and the rule records that this same mistaken belief stranded 45 of 49 KB files in September. I followed CLAUDE.md, which is the constitution. It made no difference today because there was nothing to write, but on a day with a genuine Stafford find it would, and the prompt should be corrected.

Scan status: today's scan landed before this run (0efff3a, 3 staged / 9 skipped, 12:00 UTC; promoted to main as a5e05bf), so the candidate pool is current and no fallback to an earlier scan day was needed.