Stafford

← 2026-10-26 · all briefs · 2026-10-28 →

Stafford Brief — 2026-10-27

OPS. Clean run. Started from main at a06c6e9; both repos were already checked out and both were synced before work began. Yesterday's run merged, so there is no OPS flag to carry. Today's own ship disposition is on the last line of this brief.

What this brief is. The 12:00 scan did a lot of work today and did it well — two formal models ingested, Aral found by citation chase, the Impala/iKang tension stated honestly rather than resolved in the framework's favour. None of that is repeated here. This brief does one thing: it chases the target that brief named as its cheapest and most valuable unchased item, and reports what came back. It came back stronger than its citation, and it changes two things in the ledgers.

📖 TODAY'S READ

Somebody ran the experiment your evidence base has been asking for, and it was published in PNAS sixteen months ago. The scan's brief recorded the honest state of the cohort argument this morning: two formal models saying AI-assisted juniors learn more slowly, one real company saying they got productive faster, and no measurement at all. It then listed four citation chases, ranked Bastani et al. first, called it "the cheapest and most valuable of the four" — and ran none of them. That was the slot I spent today.

Bastani and colleagues gave nearly a thousand Turkish high school maths students one of three things across four class sessions: an ordinary ChatGPT-style GPT-4 window, a tutor the school's own maths department had built to give hints instead of answers, or textbooks. Classes were assigned at random, the study was pre-registered, and each session measured performance with the tool and then capability without it.

The headline the citation carried is right, and it is the least interesting of three results. Students given the ordinary chat window scored 17% worse on the unassisted exam than students who never had the tool at all — not slower learning, measured capability destruction relative to a textbook.

The result the citation did not carry is the one that should change your essay. The carefully designed tutor — teacher-written hints, the version this whole field would call best practice — scored 127% better than control while the tool was available, and exactly nothing better once it was taken away. Statistically indistinguishable from the kids with textbooks. So the good design did not deliver a learning benefit. It prevented the damage and built nothing.

Technical detail: unassisted-exam effect −0.054 out of 1 against a control mean of 0.28 for the unguarded arm (significant); −0.004 for the guardrailed arm (indistinguishable from control). PNAS 122(26), online 2025-06-25. Full entry, instrument and bounds: knowledge-base/bastani-generative-ai-without-guardrails-harm-learning-2025-06.md, staged to the index today.

The breakpoint intersection is Technology Illusion, measured rather than illustrated — a deployment that moved the number being watched by 127% and the capability it existed to build by zero. And the so-what is a decision, not a citation: the framework currently implies that fixing the organizational conditions lets the technology deliver. The best-designed arm anyone has randomized says fixing the conditions stops the technology taking capability away. Those are different promises, and only one of them is evidenced.

🔭 THINKER TO WATCH

Hamsa Bastani — Wharton (Operations, Information and Decisions), with Osbert Bastani (Penn CIS) and Alp Sungu (Wharton).

Why her and not the obvious choice. The scan put Aral in this slot this morning, correctly, for the formal model. Bastani is the more useful figure for a different reason: she runs the experiments the rest of this literature calibrates against. Catrinescu's model — this morning's lead read — uses Bastani's numbers as inputs. Aral's model derives the possibility; Bastani measured it in a school with randomization and a pre-registration.

What she argues. That the design of a generative-AI deployment, not its capability, decides whether people come out of it more or less able — and that the two things we usually conflate, performance with the tool and skill without it, have to be measured separately because they move in opposite directions.

Your differentiation, and it is narrower than it was yesterday. You still own the organizational level: she has no firm, no functions, no incentive layer, no manager with an evaluation horizon. But she owns something you have been claiming as the framework's contribution — she can tell the difference between a deployment that builds capability and one that spends it, with an instrument, before the outcome arrives. Your claim is that this decomposes into five named conditions and that knowing which one is live changes the remedy. That claim is still unmeasured. Hers isn't.

Watch for: any attempt by this group to run the same design on working professionals rather than students. That is the single finding that would move your Claim 3 from argued to evidenced — and it would also be the finding most likely to be run by somebody else first.

🏢 CASE IN THE WILD

The tension the scan left open this morning is a category error, and today's paper is what shows it.

The scan's brief put it squarely: the base holds "two models on the slower-learning side and one real company on the faster side" — Impala, where less experienced writers reached productivity in one or two months instead of several — and routed it to you as a caution, with the instruction that a measured case outranks a derivation and the tension should go into Paper 2 unresolved.

It should not go in unresolved, and the reason is not that the models win. Bastani measured both quantities in the same students in the same sessions, and they moved in opposite directions at the same time: performance with the tool up 48% and 127%, capability without it down 17% or flat. They are not two readings of one thing that have to be reconciled. They are two different things.

So the question to ask about Impala is narrow and answerable: did anyone there ever measure those writers working without the tool? Almost certainly not — time-to-productivity is an output measure, and output is produced with the tool in hand. On that reading Impala is not counter-evidence on the learning question at all. It is a well-documented instance of the assisted-performance gain that Bastani also found, strongly, in both treated arms.

The honest open item this leaves, and it is smaller and sharper than the one the scan recorded: Impala stops being a reassurance and becomes an untested case. If its writers are now two years in and have never worked unassisted, the firm has a capability position nobody has measured and no reason to think it is sound. That is a question you could put to a client in one sentence, which is more than the unresolved tension was worth.

⚙️ FOUR FORCES CONCLUSION

Capability, and the point is about what kind of thing it is. The scan made Commitment the Force of the day, for good reason — Catrinescu formalises it. This is the other half, and it is the half with a measurement behind it.

The base has been treating the Four Forces as conditions you arrive with: assess them, strengthen them, then launch. Two instruments now say Capability is a stock the deployment spends. The Census study read on 26 October found firms adopting AI more intensively dropping the management practices that make AI work — the organizational condition as casualty. Bastani now shows the same shape one level down, with randomization: the capability was there, the tool arrived, the capability fell 17%.

What is new today is not that it falls. It is what happened to the people who tried to stop it falling. Yesterday's challenge entry ended on a specific bet: the decisive test is not whether the forces fall, but whether anyone holds them — and nobody has run it. Somebody had. The guardrailed arm is exactly that test, and holding the condition produced a null: no loss, no gain. Logged as an amendment rather than a new challenge, because it contradicts a sentence already in the ledger: knowledge/thesis-challenges.md, 2026-10-27.

For Paper 2 this means one sentence has to change and it is a sellable change. Not get the conditions right and AI pays off. Rather: the breakpoints tell you what the deployment will cost you and which condition is about to be spent. A client will buy that. It is also the only version the evidence currently supports.

💭 OPEN QUESTION

Has anyone, anywhere, measured a technology deployment that left people measurably more capable than not deploying it at all — and if not, what exactly is the upside half of your framework resting on?

This is not rhetorical and it is not pessimism. Follow what the base actually holds. The best-designed arm in the best-identified experiment on this question returned a zero. The Census study found conditions being consumed. Both of this morning's models predict capability loss. Every measured result in this base is somewhere between "destroys capability" and "costs nothing" — and not one of them is "builds it."

If that is genuinely the state of the evidence, then two things follow and they are both actionable. First, your diagnostic is in better shape than your prescription, and asset 12's concession section should say so before a critic says it for you. Second, the gap is a product. If nobody can show a deployment that leaves people more capable, then "design this one so it doesn't cost you your senior bench" is a service with no competitor making a credible rival claim.

And if the answer is yes — somebody has shown it — that is the most important paper this base does not hold, and finding it is worth more than the next five ingests.

🥊 CHALLENGE

(Not a Wednesday sweep — evidence against the thesis surfaced today and the template says state it.)

The strongest form: "Your framework's remedy has never produced a measured gain. The one time somebody randomized best practice against doing nothing, best practice tied."

A hostile reader does not need to attack the diagnosis — the diagnosis is holding up well, and today strengthened it. They attack the implied payoff. Technology Illusion, as published, warns that technology laid on broken organizational conditions fails. The unstated converse — that technology laid on good conditions succeeds — is what every client is buying, and it is the part with a randomized null under it.

What it would take for this to win outright: a second setting where deliberate, well-designed deployment also produces no capability gain over not deploying. One null in one school is a result; two in different settings is a property.

What it would take to defeat it: any instrument showing a designed deployment producing a capability gain, not merely the absence of loss. That is now a specific acquisition target rather than a complaint, and it is the thing to search for next.

Four bounds, stated so this is not cited past its evidence. High school mathematics, not work. One school, one subject, four sessions, with the authors explicitly disclaiming long-term outcomes. A Fall 2023 window, when the models were weaker and users less practised. And the authors offer an innocent partial explanation for the null themselves — the guardrailed students could get their answers checked during practice, which inflates the practice gap without implying learning. None of those bounds touches the 17% harm result.

Full entry with the citation hazard: knowledge/thesis-challenges.md, 2026-10-27 amendment to the drawdown objection.

📈 PATTERN BUILDING

Proposing a pattern at two members, both measurements, and flagging the one thing wrong with it.

Working name: Consumed Conditions. The organizational input a deployment needs is reduced by the deployment itself, not merely found absent beforehand.

Why it is worth proposing now: these two are genuinely independent — different authors, different data, different sectors, one observational and one randomized — which is exactly the test this base applies and the test this morning's pair of models failed.

The flaw, named rather than buried: the two members measure different levels. McElheran measures a firm practice; Bastani measures an individual skill. A pattern that silently merges the two is the loose practice the patterns file exists to prevent. So it should be opened with the level mismatch written into the entry, or held until a firm-level, non-manufacturing instrument arrives — which is precisely the acquisition the drawdown challenge now names. Brandon's call; the proposal text is in Routing and it belongs to research-evidence either way.

🧱 ASSET SUGGESTION

No new asset. master-plan.md hard stop 1 is unconditional and the exception clause was removed for good reason. Today's material goes into asset 12, the Tier-2 gate — amendment logged in knowledge/asset-suggestions.md.

The pitch, in two sentences: asset 12's specification already promises to name what the framework does not yet have; today gives it the hardest and most credible version of that concession, which is that the framework's remedy has a randomized null under it rather than merely a missing instrument. Conceding it first is the Paper 3 seventy-percent move again — that retirement bought credibility on worse material than this, and this one is one search away from any reader who cares, in PNAS, with the data and code released.

One constraint it also hands you, and it is the practically useful part. Bastani measured what the students thought alongside what they scored, and the two ran opposite in both treated arms: the harmed students noticed nothing, and the well-designed arm came away convinced it had improved when it had not. The better the deployment is designed, the more confidently wrong the self-report. That lands on two items already in the order — Tier 3.1 (the cluster-level self-checks) and Tier 4.3 (the pre-transformation rubric), whose last place is now over-determined: not just unfalsifiable before Tier 0.2 has data, but biased with a known sign, in the direction that flatters the engagement. Both need at least one item per condition that is an observable rather than a self-assessment. Candidates exist — the three deployment indicators the scan pulled off Catrinescu this morning are "observable from the inside without any new instrument."

🎓 SPARRING NOTE

Not issued. Thirteenth consecutive decline, same reason: 09-17 is still the one live note, now forty days old, and master-plan.md Tier 0.1 prices the unblock at fifteen minutes. Today's material is parked against growth edge #2 in knowledge/brandon-development.md — it pairs with the 10-16 decomposition hook and the pair is one coaching session, no research required. Decompose a construct before you score it; know your instrument's bias before you trust its reading.

SOURCES & THREADS

Exploratory slot: spent on the scan's own priority-1 unchased citation (Bastani et al.), verified unspent by grep across both repos first. Yield: 1 KB file, 1 staged entry, 1 challenge amendment, 1 asset amendment, 1 thread criterion met. Both queries logged with full reasoning in knowledge/query-log.md.

Second chase — the owed one, and it DISCHARGES after the scan's route failure this morning. The scan tried to establish whether the Microsoft Work Trend Index changed its survey screening between waves, failed by web search, and re-registered the chase with the correct route named: the WorkLab methodology, not search. I ran the named route and it worked. The 2025 report's methodology — 31,000 knowledge workers, 1,000 in each of 31 markets, fielded 6 February to 24 March 2025 by Edelman Data x Intelligence — contains no screen on personal AI use at all. Inclusion is occupational: people who typically work at a desk. The 2026 wave explicitly screens out anyone answering "never" or "don't know."

Why that matters, in plain terms: the 67%/32% split — organizational factors producing twice the AI impact of individual effort — is the figure paper-2 Claim 5 leans on, and it comes from the 2026 wave, which means it describes only people already using AI. Two consequences. No WTI number may be compared across the two waves; the instrument changed on every axis (31,000 to 20,000 respondents, 31 markets to 10). And the direction of the bias is worth knowing because it is in Brandon's favour: organizations whose conditions were bad enough to prevent any AI use are excluded by construction, so if anything the 2x understates the organizational effect. Cite it with the frame attached. One honest qualification: what is proven is a change in disclosure — a short methodology note could omit a screen that was applied.

Threads: 0 due today (earliest 1 Nov). 1 criterion met — Apprenticeship rung, criterion (c), which the scan recorded this morning as "addressed but NOT met, because what arrived is a derivation and the criterion says measurement." A randomized, pre-registered, peer-reviewed measurement arrived. 1 thread advanced — Is the deployment number an artifact of the sampling frame?, via the chase above. Both are proposals; threads belong to research-evidence.

Candidates: 1 proposed — PNAS Economic Sciences, which is in neither sources.md nor source-candidates.md. That is a real gap rather than a new outlet: the base monitors arXiv econ.GN, SSRN and NBER as standing channels, and the one peer-reviewed randomized trial on its central mechanism was sitting in PNAS for sixteen months, reachable only through another paper's reference list. Second time in two days that a citation chase beat the keyword sweeps — the scan made the same observation about Aral this morning. That is now a pattern about this pipeline's own method, and it is worth more than either individual find.

Field map: 1 proposed — Hamsa Bastani, Osbert Bastani and Alp Sungu.

Routing

strategy — the decision today forces, and it is one sentence in Paper 2. The scan routed you a choice this morning between two versions of Incentive Fragmentation. This is a different and more basic one. Your published work implies that clearing the breakpoints lets the technology deliver; the only randomized test of that implication returned a tie against doing nothing. Decide whether Paper 2 claims an upside or claims a cost. The cost version is fully evidenced, sellable, and not what the drafts currently say. Recommended: concede it in asset 12 first, so the concession is yours rather than a reviewer's.

strategy — one acquisition target, stated precisely enough to search for. Any instrument showing a designed deployment leaving people measurably more capable than no deployment. If it exists it is the most important paper this base does not hold. If it does not, that absence is the premise of a service.

personal — the inverse of this morning's caution, and it is the more useful warning. The scan told you, rightly, that today's material agreeing with you is the condition under which this pipeline is least reliable. Today's addition cuts the other way: the strongest evidence found today supports your diagnosis and undercuts your remedy. That is the useful shape. Be more suspicious of the parts of the framework nothing has tested than of the parts that keep getting confirmed.

governance — nothing.

market — nothing. No window. A June 2025 PNAS paper is argumentative value, not urgency, and manufacturing a window from it would break hard rule 5.

PROPOSED-UPDATE → research-evidence/knowledge/story-threads.md, thread The apprenticeship rung, append under the 2026-10-27 entry

2026-10-27 (Stafford, same day, later run): criterion (c) is MET. This morning's entry recorded it as "addressed but NOT met, because what arrived is a derivation and the criterion says measurement", and registered Bastani et al. (2025) as the thread's new priority chase. The chase was run this afternoon and the paper is stronger than its citation: not a working paper but PNAS 122(26), online 2025-06-25, pre-registered, class-randomized field experiment, n≈1,000 across ~50 classes, one Turkish high school, Fall 2023 (knowledge-base/bastani-generative-ai-without-guardrails-harm-learning-2025-06.md, staged). Unguarded GPT-4 access cut unassisted exam performance by 17% against a textbook-only control (−0.054 of 1, control mean 0.28). So the thread now holds a measurement on the slower-learning side, not only two models. Two things the citation did not carry, and both change the thread. (1) The guardrailed arm bought a null, not a gain — unassisted estimate −0.004, indistinguishable from control, beneath a +127% assisted-practice gain. Deliberate design averted the harm and added nothing. (2) Assisted performance and retained capability moved in opposite directions in the same subjects in the same sessions, which resolves this thread's standing Impala tension as a category error: time-to-productivity is an assisted-performance measure, and unless Impala measured its writers working without the tool, it is not counter-evidence on learning at all. Recommended amendment: Impala moves from "one real firm on the faster side" to "one real firm whose unassisted capability was never measured." Criterion (c) closed; (a), (b) and (e)–(g) untouched; (d) — Lichtinger & Hosseini Maasoum (2025) read directly — is again the cheapest unmet item. New criterion (h), proposed: a measurement of the same design question on working adults rather than students. Citation hazard for this thread's file: Bastani is a calibration source for Catrinescu, so the two are not independent — Bastani is the measurement, Catrinescu the organizational extension. Next check 2026-11-10, unchanged.

PROPOSED-UPDATE → research-evidence/knowledge/story-threads.md, thread Is the deployment number an artifact of the sampling frame?

2026-10-27 (Stafford): the owed chase is DISCHARGED via the route this morning's entry named. Direct read of the 2025 WorkLab report's methodology (microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born): 31,000 full-time employed or self-employed knowledge workers, 1,000 in each of 31 markets, fielded 2025-02-06 → 2025-03-24 by Edelman Data x Intelligence, 20 minutes online. Inclusion is occupational — "those who typically work at a desk (whether in an office or at home)" — and the leader/employee/manager/Frontier-Firm splits key on job level, decision influence and company AI deployment. No screen on personal AI use appears anywhere in it; the only "don't know" figure is a response category inside a result, not an exclusion rule. The 10-26 claim that the 2026 screening was a change from prior years is therefore established rather than asserted — this is the qualifying development the thread specified ("a repeat-measured vendor instrument changing its own frame"). Qualification, because this is an absence rather than a statement: a brief methodology note could omit a screen that was applied, so what is proven is a change in disclosure. And the two waves are different instruments on every axis: 31,000→20,000 respondents, 31→10 markets, Edelman named in 2025 only, fieldwork Feb–Mar 2025 against Feb–Apr 2026. Consequence: no WTI figure may be compared across waves, and the 67%/32% decomposition describes variance among adopters. Its bias direction is favourable to the paper-2 argument — organizations whose conditions prevented any AI use are excluded by construction — so cite it with the frame and the direction attached. Recommended: close the chase, keep the thread open on the narrower question of whether the 2026 frame change was ever disclosed as a change.

PROPOSED-UPDATE → research-evidence/knowledge/paper-2-case-file.md

2026-10-27 (Stafford, second run). Claim 3 upgraded from argued with cases and models to supported by one randomized measurement, on Bastani et al., PNAS 122(26), 2025-06-25 (entry staged today). This is the first peer-reviewed randomized instrument in the file on the junior-capability question. Bound, permanent and prominent: the subjects are high school students, not employees — the file's standing need is restated as the same design on working adults. Claim 5 — a required citation caveat, not a downgrade. The 67%/32% organizational-vs-individual decomposition comes from the 2026 Work Trend Index wave, whose frame screens out non-users of AI; the 2025 wave (31,000 respondents, 31 markets, Edelman) discloses no such screen (see story-threads.md). Therefore: never compare WTI figures across waves, and state the frame when citing 67%/32%. The bias runs in the argument's favour — organizations too dysfunctional to get anyone using AI are excluded by construction, so the 2x plausibly understates the organizational effect. Say that rather than leaving the selected frame for a critic to find. New and sharpest open target in the file, replacing the 10-26 salience target at the top of the ranking: an instrument showing a designed deployment leaving people measurably MORE capable than no deployment. Bastani's guardrailed arm — the best-designed deployment anyone has randomized on this question — produced no gain (−0.004, indistinguishable from control) under a +127% assisted-performance gain. Every measured result now in this file lies between "destroys capability" and "costs nothing"; none shows capability built. Until one does, the file supports a cost claim and not an upside claim, and Paper 2 should say which it is making. Acquisition targets: Bastani DISCHARGED (was priority 1, created this morning). Dell'Acqua et al. (2026), Brynjolfsson et al. (2026), McEntire (2026) stand unchased. The 10-26 salience target, and the 10-24, 09-27 and 09-08 targets, stand unmet. The Microsoft WTI screening question is DISCHARGED (above).

PROPOSED-UPDATE → research-evidence/knowledge/emerging-patterns.md — new pattern, 2 members, proposed with its flaw named

Consumed Conditions (proposed by Stafford 2026-10-27, 2 members, both measurements) — the organizational input a deployment requires is reduced by the deployment itself, rather than merely found absent beforehand. Members. (1) McElheran, Yang, Kroff & Brynjolfsson, U.S. Census CES-25-27: more intensive AI adoption causes de-adoption of structured management practices (−0.83 log points per SD of AI intensity, concentrated in establishments 25+ years old), driven specifically by KPI reviews by non-managerial employees and target awareness across employees. (2) Bastani et al., PNAS 122(26), 2025-06-25: unguarded generative-AI access cut measured unassisted capability by 17% against a never-had-it control, in a pre-registered class-randomized trial. Why proposed now: the two are genuinely independent — different authors, data, sectors and designs, one observational and one randomized — which is the independence test this morning's pair of models failed, and the test the Résumé-Panel Consensus pattern exists to enforce. The flaw, written into the entry rather than buried: the members measure different levels of analysis — a firm practice and an individual skill. Recommendation: open it with the level mismatch stated, or hold it until a firm-level, non-manufacturing instrument arrives. That instrument is the named outstanding requirement on the drawdown objection (stafford-research/knowledge/thesis-challenges.md, 2026-10-26 and the 2026-10-27 amendment). What would escalate it: a third member at the firm level outside manufacturing. What would kill it: evidence that the de-adoption and the capability loss are both choices with identifiable firms that held the condition and did better — which would convert the pattern into a breakpoint, and is the more interesting result.

PROPOSED-UPDATE → research-evidence/knowledge/field-map.md

Added 2026-10-27 (Stafford) — Hamsa Bastani (The Wharton School, Operations Information and Decisions), with Osbert Bastani (Penn, Computer and Information Science) and Alp Sungu (Wharton). The group that runs the experiments this literature calibrates against: Catrinescu's model (ingested 2026-10-27) uses their numbers as inputs. Position: the design of a generative-AI deployment, not its capability, determines whether users end up more or less able — and performance-with-the-tool must be measured separately from skill-without-it because the two move in opposite directions. Instrument: pre-registered class-randomized field experiments with released data and code (github.com/obastani/GenAICanHarmLearning) — a genre this field map is almost entirely without. Brandon's differentiation, narrower than against the modellers: no firm, no functions, no incentive layer, no evaluation horizon. But they can already distinguish a deployment that builds capability from one that spends it, with an instrument, before the outcome lands — which is close to what the framework claims as its contribution. Watch for: the same design run on working professionals. Note to the map, same as the Aral entry's: reached by citation chase, not by keyword sweep, and the paper had been in PNAS for sixteen months.

PROPOSED-UPDATE → research-evidence/knowledge/source-candidates.md

PNAS — Economic Sciences / social-science sections (first seen 2026-10-27; surfaced by Bastani et al. via Catrinescu's reference list; signal count 1; cadence continuous). Logged as a channel gap rather than a new outlet. The base monitors arXiv econ.GN and cs.HC, SSRN and NBER as standing channels, and PNAS appears in neither ledger — yet the one peer-reviewed randomized trial on this base's central mechanism sat there for sixteen months and was reachable only through another paper's reference list. Promotion bar: a second on-thesis find from PNAS, not merely a second Bastani paper elsewhere. Method note for the Friday synthesis: two days running, a citation chase has beaten the keyword sweeps (Aral 2026-10-27 morning, Bastani 2026-10-27 afternoon) while Categories 1 and 5 returned a tenth consecutive null. That is a finding about this pipeline's own method and it deserves a line in the synthesis.

PROPOSED-UPDATE → research-evidence/knowledge-base/microsoft-work-trend-index-2026-frontier-firms.md, append to the 2026-10-26 addendum

2026-10-27 (Stafford) — the open question in this addendum is CLOSED, in the direction the secondary source claimed. The addendum recorded: "A secondary search result claimed this screening was a change from prior years; the page does not say so, and the claim is NOT confirmed." Direct read of the 2025 WorkLab methodology confirms the change in the report's own disclosure: 31,000 full-time employed or self-employed knowledge workers, 1,000 in each of 31 markets, fielded 2025-02-06 → 2025-03-24 by Edelman Data x Intelligence, inclusion defined occupationally ("those who typically work at a desk"), with no screen on personal AI use and no "never / don't know" exclusion anywhere in it. The 2026 wave's screen is therefore a disclosed change between waves. Qualification: this is an absence in a brief methodology note, so what is established is a change in disclosure, not proof that no screen was applied in 2025. The waves also differ on every other axis — 31,000→20,000 respondents, 31→10 markets — so no figure from this report may be compared with a 2025 WTI figure. When citing the 67%/32% decomposition, state that it is computed on AI users only and note that the exclusion of organizations where nobody uses AI biases it toward understating the organizational effect.

Staging and writes

STAGED — 2026-10-27-proceedings-of-generative-ai-without-guardrails-ca (Bastani et al., PNAS). Gates passed on the first attempt; dry-run verified before submission. Breakpoints technology_illusion, momentum_mirage; force Capability; six key findings; sample_size 1000 (the paper's own "nearly 1,000").

One pipeline note, as a proposal since build/ is not Stafford's to edit. The instrument.data_source_genre vocabulary has no value for a randomized field experiment — the options are résumé panel, postings, self-report survey, paired survey, administrative, telemetry, outcome-defined, model-theory and unknown. I deliberately omitted the field rather than mis-set it to outcome_defined or unknown, and put the full design in inference_method. Suggested addition: field_experiment. The omission is itself informative: a taxonomy with no slot for an RCT is a fair description of what this evidence base has been able to find until today.

KB file written to both repos (knowledge-base/bastani-generative-ai-without-guardrails-harm-learning-2025-06.md), per CLAUDE.md operating rule 3.

Judgment-layer files updated in this repo: thesis-challenges.md (amendment to the drawdown objection — no seventh challenge opened, deliberately), asset-suggestions.md (asset 12 amendment; no new numbered asset, hard stop 1), brandon-development.md (note not issued, material parked on edge #2), query-log.md (both queries).

OPS — research-evidence. KB file and staged entry pushed to claude/modest-faraday-bwy8fw at c87e413; scripts/verify.sh passed locally before the push (verify: ok), so the promote-claude-branches workflow should fold it into main within the hour rather than opening a pull request. Its main, research-index.json, dashboards, briefs/ and CLAUDE.md were not touched, and the shared meta files were left for Brandon as the PROPOSED-UPDATE blocks above.