Stafford Brief — 2026-09-09
Wednesday. The scan landed at 12:15 UTC and I have consumed it; nothing in it is restated here. Its own falsification sweep came back negative for the third time and its brief is good — read it for the Statistics Canada walk-down and the AI Office hiring datum. My sweep went at a different target: the base's own negative claim, made in that brief this morning. It does not survive. No Stafford brief ran on 09-07 or 09-08 — this one covers the judgment layer since Sunday.
📖 TODAY'S READ
Americans Do I.T. Better: US Multinationals and the Productivity Miracle — Bloom, Sadun & Van Reenen, American Economic Review 102(1):167–201 (February 2012)
This morning's scan brief asked, as its Open Question: has anyone put a non-technological organizational variable on the right-hand side of an AI-outcome specification? It answered nobody has, and the paper-2 case file recorded the positive half of Claim 5 as "not evidenced at all." For AI that is correct and I checked it directly. As a statement about the design, it is wrong by fourteen years, and the counterexample is in the American Economic Review.
Bloom, Sadun and Van Reenen scored 1,633 European firms on four people-management practices — promotion on performance rather than tenure, reward power, hiring of talent, and fixing or removing under-performers — using a double-blind interview in which respondents did not know they were being scored and interviewers did not know firm performance. They joined those scores to establishment-level IT data from Harte-Hanks and to statutory company accounts from AMADEUS, and estimated a production function.
The result is the one Brandon has been asserting without an instrument. US-owned firms show a large IT-productivity premium: US ownership × computers per worker = 0.1790 (s.e. 0.0733). Put the management score's interaction into the same regression and the premium falls to 0.0784 (s.e. 0.0720) and stops being distinguishable from zero, while people management × computers per worker holds at 0.1404 (s.e. 0.0344). Under firm fixed effects the ownership term collapses to 0.0192 and the management interaction survives at 10%. The apparent return to the technology was the return to the system for hiring, promoting, rewarding and removing the people using it.
The detail that matters more than the headline, and it should change how Brandon builds his instrument. The level effect of management on productivity is 0.0188 (s.e. 0.0153) — nothing. Management does not raise output on its own in this specification. It raises the return to the technology. Score the organizations and correlate with performance and you find nothing; score them and correlate with the return to what they deployed and you find the result.
Three disciplines, and the first is disqualifying for the stronger reading. This is IT measured as PCs per worker, on panels running 1999–2006 and 1995–2003 — a precedent for the design, not evidence about AI, and it must never be cited as the latter. The management score was collected once, in summer 2006, at or after the end of both panels, so it is a blind contemporaneous score and not a prospective instrument. And better-managed firms also buy more IT (0.0882, s.e. 0.0246), so selection runs through the same variable; the fixed-effects columns are the partial answer and the authors say their precision falls.
🔭 THINKER TO WATCH
Raffaella Sadun — Harvard Business School, co-author of the above and co-architect of the World Management Survey.
Not a new name in the field. A name absent from a 517-file evidence base, which is worse. Sadun's programme is the one that made "management practice" a measured quantity rather than a consulting adjective: structured, blind-scored, joined to administrative outcomes, run across tens of thousands of firms and multiple countries. It is the most serious sustained attempt anyone has made to do empirically what Brandon does diagnostically.
Brandon's differentiation, and it is narrower and more honest than it was this morning. The WMS scores what an organization does — practices, procedures, the presence of a target-setting system. It is deliberately behaviourist, and Sadun's own framing, "management as a technology," says so. Five Breakpoints claims something the WMS does not attempt: that the practices fail in patterned ways with named mechanisms, and that Purpose and Momentum are variables rather than residuals. The honest position is not that Brandon measures what Sadun cannot. It is that Sadun measures rigorously the half of the framework that maps onto incentives and capability, and nobody — Sadun included — has an instrument for the other half.
Handle with care. The WMS is manufacturing-heavy in its early waves and its scores have a known survivorship and response-rate literature. Do not present the survey as a neutral measure of organizational quality; it measures adherence to a specific management model, which is a substantive position, and the paper's own framing ("tougher people management") makes no secret of it.
🏢 CASE IN THE WILD
Establishments in the UK that were taken over by US multinationals — same buildings, same machines, new people system, more output per computer.
This is the cleanest natural experiment in the base and it has been sitting in the AER since 2012. The authors isolate UK establishments that were acquired, and compare those bought by US multinationals against observationally identical ones bought by non-US multinationals. Before the takeover, the two groups show no difference in the productivity of their IT. After the takeover, the establishments acquired by US firms raise the productivity of their existing IT and the others do not.
The mechanism is Technology Illusion read backwards, and that is why it is worth Brandon's time. Every case in this base runs the same direction: an organization buys technology, the surrounding conditions are wrong, the return does not appear. Here the technology is already installed and unchanged, the ownership changes, the people-management practices change, and the return appears out of hardware that was already on the floor. It is the strongest available demonstration that the missing ingredient is not more technology, because in this case no technology was added.
The open question it leaves, and it is a live consulting question. The paper cannot say how long the conversion took, and it cannot separate "the acquirer installed better practices" from "the acquirer fired the people who were the problem" — the scored index contains removal of under-performers as one of its four items, which is the fastest-acting and least transferable of them. For a client who cannot change ownership and will not run a cull, which of the other three items does the work? Nothing in this base answers that, and it is the question every engagement actually turns on.
⚙️ FOUR FORCES CONCLUSION
Capability and Commitment — and the conclusion is a constraint on the framework, not a win for it.
The four scored items map cleanly. Hiring talent and promoting on ability is Capability, measured net of college-educated workforce share, which the paper controls separately and which is separately significant — so this is organizational capacity to execute rather than the human capital of the workforce, a distinction no other instrument in this base separates. Removing under-performers is Commitment: acting consistently with stated direction under pressure, which is exactly what that item measures and exactly what organizations do not do.
Purpose and Momentum are absent, again. Nothing in the index measures clarity of intent or the persistence of energy over time, and no claim about either can be built on this source. That is now the second time in five days that a formalism has reached into the Four Forces and come out holding the same one and a half. On 09-05 it was Bartolucci & Vivo's review threshold — κ, verification skill, and K, the cost of an escaped error: Capability and the accountability half of Commitment, with no Purpose or Momentum term. Today it is a published production function with a four-item index containing the same two constructs.
Two independent formalisms, different literatures, fourteen years apart, and both reach for the incentive-and-capability half and stop. A framework claiming four Forces has an external empirical precedent for roughly one and a half of them. That is a real finding about the framework and Brandon should hold it rather than route around it.
💭 OPEN QUESTION
Is the readiness instrument Brandon is being asked to build specified as a level when the evidence says it only exists as an interaction?
Growth edge #2 and asset #3 both describe the same object: score organizations on Five Breakpoints constructs before an AI deployment, track outcomes at 6–12 months, see whether the score predicts. That is a level specification — organizational score in, outcome out. Today's paper ran that specification and got 0.0188 with a standard error of 0.0153. Nothing. The same data, with the score interacted against technology intensity, gives 0.1451 with a standard error of 0.0331.
If that generalizes, an instrument built the obvious way is built to fail, and it would fail in the most damaging manner available: a published null on a Five Breakpoints readiness score, produced by Brandon, against an unfalsifiability objection he was trying to answer. The 08-30 warning in the challenge ledger was about over-claiming on a small sample. This is a different and worse failure — a correct sample, a clean design, and the wrong right-hand side.
The question is therefore not "does readiness predict outcomes." It is: does a Five Breakpoints score predict failure, or does it predict the return to whatever the organization is deploying? Those are different instruments, they need different outcome variables, and only one of them has a fourteen-year-old precedent showing it estimable. Brandon should decide which one he is building before anyone scores anything.
🥊 CHALLENGE
The symmetry defence — the load-bearing argument against the unfalsifiability objection — was built on a search of the wrong literature, and it has a one-citation refutation.
Stated at full strength, as the objection deserves. The ledger's defence has been: nobody has ever built an organizational readiness instrument that demonstrates predictive validity; implementation science tried for twenty years and its own architects concede the question is open (Weiner 2020: "the question of whether readiness for implementation matters remains open"); the definitive prospective test was protocolled and never reported (Helfrich 2011); and it is still true in 2025 (Caci: 40 of 46 studies measure readiness once). Therefore the objection indicts an entire measurement class and every critic deploying it stands on the same ground as Brandon.
Every source in that defence comes from implementation science. A grep across all 517 knowledge-base files in both repositories returns zero hits for Van Reenen, Sadun, the World Management Survey, Bresnahan, or Americans Do I.T. Better. The base never looked in economics, where organizations have been scored on non-technological practices and those scores put on the right-hand side of technology-productivity specifications since Bresnahan, Brynjolfsson & Hitt in 2002, at scale since 2007, with double-blind collection and administrative outcomes.
What this costs and what it does not. It does not defeat the challenge and it does not rescue it: today's paper scores management contemporaneously, at the end of its own panel, which is precisely the design Caci names as the field's persistent failure. No prospective instrument that forecasts an outcome has been built, in either literature. What it costs is the width of the defence. "No readiness construct anywhere has cleared this bar" is false. "No readiness-for-a-named-change instrument has cleared it, and the adjacent economics literature has cleared a weaker bar with a blind contemporaneous score" is true, defensible, and considerably less rhetorically satisfying.
What it would take for the objection to win outright: unchanged — a prospective Five Breakpoints score that fails to predict a pre-defined outcome on an adequately powered sample. What has changed is that the objection now also wins a smaller victory every time Brandon states the defence in its wide form in front of a labour economist. Logged in knowledge/thesis-challenges.md; the essay pitched in asset #3 needs the same paragraph struck before it ships.
🧱 ASSET SUGGESTION
"Run Table 6 Again" — the two regressions side by side.
In 2012 the AER published a specification that answers what AI research is currently arguing about: the return to a technology, moderated by a blindly measured non-technological organizational variable, with the technology's apparent premium absorbed by it. In 2026 Statistics Canada ran the same shape on AI, entered five moderators, and four of them were more technology. The asset is that comparison and the missing column: the people-management interaction of the 2012 paper, estimated on AI instead of computers per worker. It gives Brandon a named, dated, published thing to point at instead of an assertion, and it converts "organizational readiness is the prerequisite" from a claim into an unrun estimate with a known specification. Cheapest asset in the ledger — both sources are read and the numbers are extracted. Full spec logged as entry #6 in knowledge/asset-suggestions.md.
🎓 SPARRING NOTE
Issued (fourth ever), and it supersedes the 09-05 parked hook rather than stacking on it — same growth edge (#1), same question, now with evidence instead of a prompt.
On 09-05 I asked whether Brandon could write the Four Forces as parameters in a formal model, and whether Purpose and Momentum have terms or the framework has only two variables a modeller could use. That was a hypothetical. It is no longer one. Two independent formalisms have now reached into the Four Forces: a queueing threshold (κ, verification skill; K, cost of an escaped error) and a production function with a four-item blind-scored index (promotion, reward power, hiring, removal of under-performers). Both come out holding Capability and the accountability half of Commitment. Neither has a Purpose term or a Momentum term.
The prompt, answerable in one paragraph and worth attempting before reading anything further. Take the four items in the Bloom–Sadun–Van Reenen index as given. Write the fifth and sixth items — the ones that would measure Purpose and Momentum in the same register: a practice an interviewer could score 1–5 from a blind conversation with a plant manager, without asking "is your strategy clear?" If those two items can be written, the framework has four estimable variables and the acquisition target in the case file becomes a survey design. If they cannot, the honest conclusion is that Purpose and Momentum are interpretations of the other two rather than independent constructs — and that is a finding worth having, because it is exactly the objection a referee would raise and it is cheaper to concede than to defend badly.
Two notes remain outstanding and unanswered (08-31, 09-01, both growth edge #3). Logged for the first /coach session; Stafford is not treating its own prompts as completed exercises.
SOURCES & THREADS
Threads: none due. The scan advanced two early (EU AI Act enforcement, perceived-versus-measured) and I have not re-checked either — no purpose is served by a second same-day chase of a thread whose channel was restricted six days ago.
Exploratory query — Wednesday falsification sweep, run at full strength against the base's own claim rather than against Brandon's. Target: this morning's scan brief Open Question, the case file's "the positive half ... is not evidenced at all," and limit (1) on the escalated Conditional-Productivity Result. Deliberately not a repeat of the scan's sweep, which hunted the opposite sign. Method, in order: coverage check first — grep 517 KB files across both repos for Van Reenen, Sadun, World Management Survey, Bresnahan, organizational capital, Americans Do I.T. → zero substantive hits; then live search; then the primary obtained from worldmanagementsurvey.org and read in full with PyMuPDF, per the 09-04 lesson that self-hosted copies defeat publisher gates. Yield: 1 KB file, 1 challenge narrowed, 1 asset amended, 1 asset opened, 2 field-map corrections proposed. Logged in knowledge/query-log.md.
Lesson, generalizing the 09-03 one by a step. That lesson said: before spending the slot on a live search, check whether the gap is a coverage hole in the base rather than a hole in the world. Today's is sharper — check the base for coverage of the design, not only of the finding, and search the literature that owns the method rather than the literature that owns the topic. Five months of searching "AI" and "organizational" never returned the canonical instrument, because that instrument is about IT and management, and neither word is in the standing categories.
Not chased, and both are named for the next Category 6 slot rather than left implicit. Chemla & Tena, "AI Adoption, Incentives, and Firm Value in General Equilibrium" (SSRN 5997775) — the scan deferred this on 09-08 to "the next category-6 slot," today's Category 6 slot went elsewhere, and it is still unread; SSRN gated it twice. It remains the base's most on-thesis unread item, a formal model putting incentive contracts rather than coordination on the left-hand side. A second target surfaced beside it: IZA DP 18574, "Toward a Bad Job Economy: AI Adoption, Agency Costs...", openly served at docs.iza.org/dp18574.pdf — same territory, no gate. Neither is ingested and neither may be cited; a search summary is not a source.
Candidates: none opened. The Statistics Canada slot the scan opened today is correct and I have nothing to add to it.
Routing
STAGE-CANDIDATE — ready to run in research-evidence
python3 build/add-research-json.py --json '{
"title": "Americans Do I.T. Better: US Multinationals and the Productivity Miracle",
"publisher": "American Economic Review (Bloom, Sadun & Van Reenen)",
"url": "https://www.aeaweb.org/articles?id=10.1257%2Faer.102.1.167",
"published_date": "2012-02-01",
"credibility": "high",
"instrument": "Data source: three separately-collected datasets joined — CEP double-blind management-practices interview survey (summer 2006, 51 interviewers, 4,003 firms globally, 1,633 European firms used); Harte-Hanks European CiDB establishment IT panel (technology measured as PCs per worker); Bureau van Dijk AMADEUS statutory company accounts from mandatory national registries. Separately a UK ONS establishment census panel (11,000+ establishments, 21,746 observations). Inference method: production function, dependent variable ln(output per worker), regressed on ln(non-IT capital per worker), ln(labour) and ln(computers per worker), with the technology term interacted with a people-management z-score and horse-raced against US and non-US multinational ownership interactions; firm fixed-effects variants reported; college-educated workforce share controlled and interacted with IT. Observation window: EU firm panel 1999-2006 (719 matched companies); UK establishment panel 1995-2003. Management measured once, summer 2006, at or after the end of both windows — a blind contemporaneous score, NOT a prospective instrument. Genre note: blind structured interview scores joined to statutory accounts; distinct from resume/postings inference and from single-questionnaire self-report.",
"breakpoints": [
{"name": "incentive_fragmentation", "rationale": "The four blind-scored items are promotion on ability rather than tenure, reward power, hiring of talent, and fixing or removing under-performers; their interaction with computers per worker is 0.1451 (s.e. 0.0331) and is the term that survives once the US ownership premium is absorbed."},
{"name": "technology_illusion", "rationale": "US-owned establishments are 32 percentage points more IT-intensive and 9 points more productive than other multinationals, but the US-ownership-by-IT interaction falls from 0.1790 (s.e. 0.0733) to 0.0784 (s.e. 0.0720) and loses significance once the people-management interaction enters — the productivity credited to the computers belonged to the people system around them."}
],
"forces": [
{"name": "Capability", "rationale": "The management index enters net of college-educated workforce share, which is separately controlled and separately significant, so the measured effect is organizational capacity to execute rather than the human capital of the workforce."},
{"name": "Commitment", "rationale": "Fixing or removing under-performers is one of the four scored items in the index that absorbs the technology premium, and is the item that most directly measures acting consistently with stated direction under pressure."}
],
"key_findings": [
"People management z-score interacted with ln(computers per employee): 0.1451 (s.e. 0.0331), significant at 1%; 0.1404 (s.e. 0.0344) with ownership interactions also present",
"US ownership interacted with ln(computers per employee) falls from 0.1790 (s.e. 0.0733) to 0.0784 (s.e. 0.0720) and becomes insignificant once the management interaction enters; 0.0192 under firm fixed effects",
"The LEVEL effect of people management on productivity is 0.0188 (s.e. 0.0153), not significant — the organizational effect exists only in the interaction with technology intensity, not on its own",
"People management predicts technology intensity itself at 0.0882 (s.e. 0.0246) — better-managed firms both buy more IT and get more from it, so selection runs through the same variable",
"UK establishment panel: labour productivity 24% above industry average for US-owned establishments against 15% for other multinationals; US establishments 32 percentage points more IT-intensive",
"Establishments taken over by US multinationals subsequently raised the productivity of their existing IT; observationally identical establishments taken over by non-US multinationals did not"
],
"sample_size": 719
}'
Version discipline for whoever stages this: the published record is AER 102(1):167–201, February 2012; all coefficients above are read from the full text of the 2 August 2009 CEP working-paper version served openly by worldmanagementsurvey.org, and table and column numbering may differ in the published article. Cite the AER record for the claim and verify any coefficient against the AER version before it appears in a paper draft.
PROPOSED-UPDATE — knowledge/emerging-patterns.md, the Conditional-Productivity Result
Anchor: append after the 2026-09-09 update. The disposition being proposed is a non-addition, and it should be recorded as one so the question is not reopened every time this paper surfaces.
Note 2026-09-09 (Stafford) — Bloom, Sadun & Van Reenen (AER 102(1):167–201, 2012) is deliberately NOT recruited to this pattern, and the reason is the pattern's own boundary. This pattern is about AI productivity moderated by organizational context. BSVR moderate IT — measured as PCs per worker, on panels running 1999–2006 and 1995–2003. Adding it would inflate the count with a different technology two decades removed and would breach the same discipline that kept NBER WP 35275 out on 09-04. What it does instead is bear on limit (1), which travels permanently with
ESCALATEDstatus and states that none of the members decomposes the organizational layer. That limit is unchanged for AI. But the limit has been read in this file as though the decomposition had never been attempted anywhere, and it has: BSVR's four-item people-management index is a decomposition of the organizational layer, it is not technology, and its interaction with technology intensity is 0.1451 (s.e. 0.0331) while its level effect is 0.0188 (s.e. 0.0153) and null. The sharpened statement of the limit: no AI instrument decomposes the organizational layer, one IT instrument did, and it found the effect only in the interaction — which is a specification warning for anyone building the AI version. File:stafford-research/knowledge-base/bloom-sadun-van-reenen-americans-do-it-better-2012-02.md
PROPOSED-UPDATE — knowledge/paper-2-case-file.md, gap ranking
Anchor: the 2026-09-09 gap ranking, item (1) Claim 5. The restated acquisition target — "an instrument putting a non-technological organizational variable on the right-hand side of an AI outcome specification" — stands unmet and I confirmed it directly. Two amendments. First, replace the sentence "the positive half is untested" with "the positive half is untested for AI; the equivalent specification was estimated for IT by Bloom, Sadun & Van Reenen (AER 2012) and returned the framework's result, which makes the target a re-estimation rather than an invention." Second, name the target as a specification rather than a wish: the people-management interaction of that paper, estimated with AI adoption or AI capital in place of computers per worker, with the organizational score collected blind and the outcome taken from records the respondent does not control. That is buildable, it has a fourteen-year-old template, and the statistical agencies now routinely hold both halves — as this morning's own find demonstrates.
PROPOSED-UPDATE — knowledge/field-map.md, two items
(a) Correction to the Bick / Blandin / Deming / Fuchs-Schündeln / Jessen entry (first logged 2026-08-18). That entry states the group's work has "no organizational variable anywhere" and sets a watch item for "any wave of the adoption survey that puts a firm-level organizational variable on the right-hand side." The watch item is still unmet and should stay open, but the "no organizational variable anywhere" characterisation is now imprecise: the same five authors, writing in the St. Louis Fed's On the Economy on 2026-04-14, report a country-level correlation of 0.83 between World Management Survey management scores and AI adoption in production processes across 32 European countries plus the US. Not ingested, and it should not be — it is an ecological correlation with countries as the unit, it predicts adoption rather than moderating an outcome, and it sets manufacturing-weighted WMS scores against economy-wide adoption rates. What it changes is the prior: this group has now reached for a management-practices variable once, which makes crossing the line materially more likely and the watch item more valuable than it looked. Candidate ledger: signal count stays at 0, watch note revised. URL: https://www.stlouisfed.org/on-the-economy/2026/apr/why-does-ai-adoption-differ-so-much-countries
(b) New entry proposed — Nicholas Bloom (Stanford), Raffaella Sadun (Harvard Business School) and John Van Reenen (LSE) / the World Management Survey programme. Absent from a 517-file base, which is the single largest coverage hole this system has found. Position: the programme that made management practice a measured quantity — structured double-blind interview scoring, tens of thousands of firms, joined to administrative and accounting outcomes. Core argument: management is a technology; practices are unevenly distributed, persistent, and complementary with information technology, and the complementarity rather than the technology explains a large share of cross-country productivity differences. Why relevant now: this is the methodological ancestor of every "organizational readiness moderates AI returns" claim in this base, and none of those claims cites it. Brandon's differentiation: the WMS scores practices behaviourally and does not claim a mechanism taxonomy; Five Breakpoints claims patterned failure modes with named mechanisms, and claims Purpose and Momentum as variables the WMS does not attempt. Watch for: any WMS wave carrying an AI module, or any paper interacting a WMS-style management score with firm AI adoption — that paper is the case file's top acquisition target and would arrive with a disinterested instrument.
Judgment-layer files updated in this repo
knowledge/thesis-challenges.md — Unfalsifiability risk: symmetry defence narrowed, response plan given a second template, the 1.5-of-4 Forces constraint recorded. knowledge/asset-suggestions.md — asset #3's essay paragraph flagged for striking, asset #6 opened. knowledge/query-log.md — today's sweep. knowledge/brandon-development.md — sparring note logged.
By audience
Strategy — The single most useful thing from today is not the citation, it is the specification warning: the organizational effect existed only in the interaction, and the level effect was a null with a tight standard error. Any readiness instrument Brandon builds that scores organizations and correlates the score with outcomes is built the way the one published attempt found nothing. It has to record score, technology intensity, and outcome.
Market — No window opened and none is manufactured. The scan's disposition on the agent-intrusion window closing ~09-28 unwritten is correct and I am not going to soften it. One addition to the Accenture window piece (closes ~10-05): the mechanism paragraph is now stronger than it was this morning, because it can put a 2012 AER result beside a 2026 Statistics Canada result and show the same coefficient walk-down with two different absorbing variables — one of which is technology and one of which is not. That is a better paragraph than either finding alone.
Governance — Nothing today.
Personal — Read the 09-09 scan brief's Open Question and then read the challenge section above. The base spent five months not finding a paper that is in the American Economic Review and was written by three of the best-known economists in the field — because the standing categories are about AI and this paper is about IT. That is a search-design failure, not a research failure, and it is worth more than the paper is. Every category in the scan is defined by topic. None is defined by method. The one search that found this ran on the method.