Stafford

← 2026-10-28 · all briefs · 2026-10-30 →

Stafford Brief — 2026-10-29 (Thursday)

MERGED d75f0ba main. Research-evidence: branch claude/modest-faraday-9ijv28 pushed (1 KB addendum, 1 staged entry corrected and held via verify-entry.sh); its promote-claude-branches workflow folds it into main once scripts/verify.sh passes. Brief emailed to brandon.freitag@gmail.com.

OPS. Clean run. Started from main at 9bfc9cd; both repos synced before work began. Yesterday's run merged, so there is no OPS flag to carry. Ship disposition is on the last line.

What this brief is. The 12:00 scan asked the single most important question in its own brief and said it could not answer it: "The first thing to check is whether Babina already reports it inside the paper nobody here can open." The paper was opened this afternoon and read in full. It answers the question, it answers it against Brandon, and it contains something worse than the answer. None of the scan's own content is repeated here.

📖 TODAY'S READ

The best new evidence in Brandon's base has a 267-page appendix in which his central argument is unnecessary — and the firms his framework says AI should fail for are the firms AI pays off for.

The scan ingested Babina, He & Jiang's Canaries in the Gold Mine (NBER working paper 35684) this morning from the abstract alone, because the PDF was refused everywhere it looked. I got it. Not from NBER, not from SSRN, and — worth noting, because it is the route our own rules prescribe — not from either author's website either; both link only to the page that refuses us. It is sitting in the authors' public replication repository on GitHub, two clicks from Babina's own data page, which this system had already visited once without opening the folder. Route lesson, now standing: when a working paper has a public replication repository, look in the repository for the manuscript.

First, the good news, and it is real. Every number the scan marked "not publishable" is confirmed at the primary and may now be used. Firms that raised their AI hiring got nothing for the eight years to 2018 — the estimated relationship is flat and statistically indistinguishable from zero — and then, from 2018 to 2024, roughly one extra percentage point of productivity growth per year for a firm moving up one standard deviation in AI-hiring growth. That holds up when the authors isolate AI investment using firms' university hiring ties from before 2010, so it is not simply that good firms hire AI people. It is 1,321 U.S.-listed non-tech companies. Technical detail: Table 2 Panel A −0.004 to 0.014, all insignificant, N=1,222/1,091; Panel B 0.062** sales per worker and 0.075*** TFP over six years, N=1,321/1,173; instrumented estimates 0.135–0.160 and 0.110–0.144, first-stage F 10.1–26.2.

Now the part the abstract does not say. The authors split their companies into those that started with a lot of accumulated organizational know-how and those that started with little — measured from audited accounts, not from anybody's opinion of themselves. The productivity gains land almost entirely on the companies that had the least. They are the only group where the effect is statistically real; the organizationally rich firms get about half as much and nothing you could distinguish from zero. The authors check the obvious objection — that the low group was simply having a bad year and bounced back — by re-sorting on a five-year average and by running the same test on the years before AI, and both come back empty. Technical detail: Table 7, below-median initial organization capital 0.092** / 0.105***; above-median 0.050 / 0.052, both insignificant; N=875/836.

The breakpoint intersection, and it is a collision rather than a confirmation. Technology Illusion says technology laid on broken organizational conditions does not deliver. This is the largest firm-level test of that proposition in the base and it points the other way. The scan filed the paper as the nearest thing yet to the evidence Paper 2 has been missing — an organizational variable sitting between AI spend and a real outcome. After the read, that is not available. Either organization capital is the Capability force under another name, in which case this contradicts Breakpoint 4 at 875 firms; or it is something else, in which case the paper is not evidence for Paper 2's claim at all. Either way it is not support, and the knowledge-base entry has been corrected to say so.

Three cautions the scan filed this morning are now wrong, and one of them is wrong in Brandon's favour and one against. The window and the effect sizes are confirmed (good). And the measure built from job-advert text — which this base flagged as unvalidated — is validated, five ways, including against audited accounting data. That correction runs against us and is recorded as such. Full detail: knowledge-base/babina-he-jiang-canaries-gold-mine-organization-capital-2026-08.md, Stafford verification addendum.

🔭 THINKER TO WATCH

Tania Babina — and the point today is not that she is in Brandon's territory. It is that she wrote down his opponent's argument properly, in a way nobody in this market has.

The scan already named her this morning as the most load-bearing name in the base's structural evidence, so that is not repeated. What changed in the last three hours is which of her claims matters. The scan read her as holding the J-curve position — the gains are late because the organizational stock takes time to build — and framed Brandon's differentiation as patience versus design. That framing is too generous to Brandon, and the paper's Appendix E is why.

She does not argue for patience. She proves the delay needs no explanation. Appendix E builds a model in which AI is a technology for writing operational knowledge down, and organizational knowledge is a slowly depreciating stock. There is nothing organizational in it — no incentives, no purpose, no process, no commitment, nothing that can be well or badly designed. And it produces the multi-year delay anyway, because that is what happens when you add to a big stock a little at a time. Her sentence, verbatim: "Responses that take many years to build are therefore not evidence of adjustment costs or implementation lags—they are what a stock-flow system delivers when stocks are large relative to flows." Reasonable depreciation rates alone give half-lives of five years or more.

Read that as a competitive statement, because that is what it is. Brandon's framework reads a long delay between buying technology and seeing returns, and returns five organizational failures. This model reads the same delay and returns depreciation arithmetic. Every other name in the field map offers a rival theory of failure. She offers a world in which nothing failed and the symptom shows up regardless. That is the harder opponent, and it has been unnamed in this base until today.

Where Brandon's differentiation now actually lives, and it is narrower and better than "patience versus design." It is in the shape of the curve, and it is in her own table. A stock quietly accumulating from a raised conversion rate should show a ramp — small positives inching upward through the null decade. Her null decade is flat: −0.004 and 0.014, nothing at all for eight years, then everything in six. A step is not a ramp. That is a checkable claim about the timing profile, available in her published figures to anyone, and it is the first place in this project where Brandon's position implies a different curve rather than a different vocabulary. Use that. Do not use "patience versus design" — she is not arguing for patience.

🏢 CASE IN THE WILD

No named organization again today, and today the absence is the finding rather than an apology for one.

The scan said the same thing this morning about its own sources, so this is not a repeat of that point — it is a different one. The paper contains the mechanism that makes Brandon's flattening argument work, cites the right source for it, and then measures the opposite in its own data without noticing.

Footnote 34 cites Garicano (2000) on knowledge hierarchies: organizations concentrate expertise in a few people and route exceptions to them, at a cost in communication — and "codifying that expertise into a firm-wide system reduces the need for the matching that the hierarchy exists to perform." That is asset 2's thesis, written by the authors, in a footnote.

Then every structural result in the paper runs the other way. Over 2018–2024 AI-investing firms grow sales and do not grow headcount. They expand intangible things — overheads and R&D up roughly 5% and 2.5% — and buy no additional plant or equipment. And in their workforce, excluding the AI staff themselves so the result is not circular, the share of senior roles and advanced degrees rises while the share of junior roles and non-graduates falls. Average wages rise.

The mechanism says the hierarchy should need less of itself. The data says the organization got more senior. The paper does not reconcile the two, and there is no measure of management layers or spans of control anywhere in it — so this is not a measurement of Brandon's wager, and his punch list has not been reopened on it.

The open question it leaves, and it is the useful one: if codification really does remove the matching work that hierarchy exists to do, why does every instrument that looks for the consequence find a thicker senior layer instead of a thinner middle? Either the codification is not happening, or the layer that disappears is not the one anybody is counting. Both answers are interesting and the second one is Brandon's.

⚙️ FOUR FORCES CONCLUSION

Capability, and today it is the Force that needs a definition rather than more evidence.

The scan's reading this morning was that Capability gains its first estimated mediator between technology spend and a realized outcome. After the read, Capability has an identity problem instead. The paper's organization capital is a stock of knowledge that has been written down — codified into systems, routines and firm-trained models. The framework's Capability is the skills, infrastructure and organizational capacity to execute. Those overlap and they are not the same thing, and the difference decides whether the base's biggest firm panel is evidence for Paper 2 or against it.

Here is the fork, in one line each. Their mechanism for why low-stock firms gain most is headroom: "an established stock of OC simply leaves less to codify." Nothing to do with readiness — a firm with nothing written down has the most writing down to do. Brandon's Capability is about whether the organization can hold a change, not about how much documentation it is missing. If he draws that line, Breakpoint 4 survives and this paper is not evidence for him. If he does not, Table 7 is a counterexample.

For the case file this is one line and it is a demotion, not a revision: the entry must stop being listed as Claim 5's first estimated organizational mediator. It is the best available statement of the position Claim 5 has to beat. That is a more useful thing for it to be, and it is what it actually is.

💭 OPEN QUESTION

Is organization capital your Capability force, or isn't it? You can answer that this afternoon, with no new data, and until you do, nobody can say whether the largest firm panel in your base supports Paper 2 or contradicts it.

Not rhetorical, and the scope is one page. Both answers are defensible and both cost something.

Say yes, it is Capability and you have conceded that firms with weak organizational capability get the largest AI productivity gains — which is Breakpoint 4 inverted, at 875 firms, on audited data, with the obvious alternative explanations already ruled out by the authors.

Say no, it is codification headroom and Breakpoint 4 is untouched — but you also give up the claim the scan filed this morning, that this paper is the organizational mediator Claim 5 has been missing for seventeen scans. And you take on a new obligation: saying what would count. If an audited, slowly depreciating, firm-specific stock of organizational knowledge is not your Capability force, the question "what measurement would be?" has no answer in the base, and the McAdam standard is waiting exactly there.

The honest observation about why this is uncomfortable. Capability is the Force this base has the most evidence on and the least definition of. Purpose has no evidence and everyone knows it — the scan has flagged seven consecutive empty scans. Capability has the opposite problem and nobody has flagged it: plenty of measurements, no agreed object, so every new instrument can be read as support or contradiction depending on the morning. That is not sustainable for a framework heading into a differentiation essay.

🥊 CHALLENGE

Not a Wednesday sweep. Evidence against the thesis surfaced on its own, in the paper the scan filed as this week's best supporting find.

Logged as The stock-flow objection — the first entry in the ledger that arrives with a formal model and a signed coefficient in the same document. Stated the way a hostile referee would put it:

Your framework explains a pattern — firms buy AI and get nothing for years — and calls it Technology Illusion. Here is a 1,321-firm panel that measures your pattern precisely, and so far it is your best exhibit. Now read Appendix E. The authors write a model with no friction in it — no misaligned incentive, no vague purpose, no broken process, no absent commitment — and it produces your lag anyway, from depreciation arithmetic. Your diagnostic reads a delay and returns five organizational failures. This model reads the same delay and returns a half-life. You owe me a reason to prefer yours.

Then read Table 7, because it closes the obvious escape. You would say the firms that fail are the ones whose organizational substrate is weak. They split the sample on exactly that, using audited accounts, and the gains land on the firms with the least of it — the only statistically real estimates in the table. The firms you predict AI fails for are the firms AI pays off for.

Why this is not any of the five neighbours already in the ledger. The maturity objection says the prerequisite expires. Exhausted-alternatives says coordination was the real constraint. The drawdown objection says the Forces are spent rather than held. This one attacks neither the prerequisite nor the mechanism nor the direction — it attacks the inference: observing the pattern is no longer evidence that a breakpoint caused it.

Where it is weaker than it looks, recorded before Brandon gets to say it. (1) The heterogeneity mechanism is headroom, not readiness, and the authors say so — but taking that escape costs the paper as support, which is the Open Question above. (2) Appendix E assumes convergence rather than finding it. Every adopter in the model reaches the same eventual gain and only the speed differs. A model that cannot represent permanent failure is not evidence against a theory of permanent failure — so the convergence test the scan called the strongest falsification test Paper 2 has is still not run on data. It is answered here by construction. (3) The sample is surviving listed non-tech firms, measured end to end. A breakpoints theory is a theory about the failures, and firms whose AI programme sank them are not in the denominator. That is the strongest technical response available, and it is a limitation of the data rather than a refutation of the estimate.

What it would take to defeat it, and it is cheap and it already exists. Take an AI-investment panel, split it on early productivity conversion, and follow the non-converters forward. Under stock-flow the gap closes at a rate set by depreciation. Under Five Breakpoints the distribution stays bimodal. Opposite signs, data that exists, nobody has looked. Full entry, with the figures and the bounds: knowledge/thesis-challenges.md, 2026-10-29. Status OPEN — the inference half effectively conceded with the repair available in the timing profile; the heterogeneity half turning on a construct question Brandon can settle in an afternoon.

🧱 ASSET SUGGESTION

Asset 12's prior-art section gains its hardest entry, and it is not a framework. No new asset is proposed — master-plan.md hard stop 1 is unconditional.

Every item in asset 12's prior-art section so far is a rival framework — Kotter's sequence, Kerr's incentives, Hammer's process, DORA's capabilities, LOCI's manualized remedy, Farach's coordination floor. The concession in each case is someone named this before you. Today's is a different species: a model that explains the framework's central observable while containing no organizational content at all. A reader who has met this paper and does not find it in the essay stops trusting the essay.

And the recovery sits in the same paper, which is what makes it an asset paragraph rather than a wound: their lag is a ramp and the data shows a step. The concession and its answer are two sentences apart, both checkable in the opponent's own table, and together they are the first time Brandon's position predicts a different curve instead of using different words.

The paragraph it also forces, and it is less comfortable: Table 7 has to be addressed, and addressing it means stating in print whether organization capital is the Capability force. That is the Open Question, in the essay, where it belongs.

Status, stated plainly because it has not changed. Asset 12 now has an opening paragraph, a four-item concession section, a prior-art section, a repair for the Paper 3 figure, a formal-model concession with its recovery, and a construct distinction that must be made. Zero words written. Nineteen days as the Tier 2 gate. Research has not been this asset's constraint for three weeks. Full spec: knowledge/asset-suggestions.md, 2026-10-29.

🔍 SOURCES & THREADS

Exploratory slot — spent on the scan's own unanswered question, and it discharged two of the four chases the scan opened this morning. The scan created a non-nber.org route to NBER PDFs and the w35684 primary, then pursued neither, and put the convergence thread's next check at 2027-01-29 — a three-month deferral of the item its own Routing calls "the most dangerous thing in the evidence base." Both chases are now closed. The route generalizes and is logged as standing: a public replication repository is a route to the manuscript. Our own hard rule 14 prescribes the author's own copy, and that route failed here — both author sites link only to the SSRN page that refuses us — so the rule needs the repository added to it. Full chain, four steps: knowledge/query-log.md, 2026-10-29. Version check passed: the PDF is dated 2026-08-22, single version, no figure drift, which is the check this base has been burned on twice.

Threads — none due, none checked early. Earliest are ASML delayering (11-01) and Project OT (11-02). Two advances proposed, both from today's read: the convergence thread's criterion (c) is discharged with the answer being the paper assumes convergence rather than measuring it, and The unvalidated proxy's new criterion (f) is partly met by the very paper that prompted it. Neither is mine to write; both are in Routing.

Candidates — 1 route added, 0 outlets. Public replication repositories, cadence irregular, no expiry clock. Not an outlet; a retrieval channel, and the second one logged this week after the McKinsey third-party-mirror route on 10-28. Two paywall routes in two days is a pattern about this environment rather than about publishers, and it is worth one line in the candidates ledger as such.

Staging — 1 entry corrected and held, not re-staged. The scan's pending entry for this paper carried six key findings, three of them falsified by the read. Re-staging would have created a duplicate rather than a patch (the duplicate gate checks the merged index, and this entry has not merged), so the pending JSON was corrected in place on the research-evidence session branch and submitted to verify-entry.sh to hold it from the 24-hour auto-merge. Flagged here because correcting another writer's staged entry is at the edge of my scope and Brandon should know it happened: the alternative was letting "the primary is unreadable and nobody has validated the proxy" merge into the index as a standing caution, which is the exact false-negative that hard rule 14 exists to prevent and which has already frozen this base on a wrong number twice.

Governance — the scan's provenance defect, verified and narrowed. This morning's brief reported that build/add-research-json.py defaults the writing agent to "stafford" when OPENCLAW_AGENT_ID is unset. It is right, and the consequence is slightly worse than stated: my own writes today are indistinguishable in the staging index from the scan's, so the audit trail cannot establish that the correction above was mine rather than the scan's. The KB Filed by lines carry it correctly and the brief carries it correctly, which is why nothing downstream has broken. It is a one-line fix and it is Brandon's call which identifier the scanner carries, not mine — but it should be made before the two-writer arrangement is a month older, because it is the only mechanical record of which writer wrote what.

No window flag. A working paper dated August 2026 with no clock on it is an asset input, not a window — the discipline Brandon applied on 10-28. Manufacturing urgency here would be the error the rules name.

Routing

PROPOSED-UPDATE → research-evidence/knowledge/paper-2-case-file.md, Claim 5 evidence block

Demote today's ingest from support to contested, and say why in the entry rather than in a footnote. The 12:00 entry records Babina, He & Jiang as "Claim 5's first estimated organizational mediator." After a full read of the primary that cannot stand. Replace with:

Babina, He & Jiang, NBER WP 35684 (read in full 2026-10-29) — CONTESTED, not support. The paper supplies the first population-scale estimate of the null arm (a decade of AI investment with no measured productivity effect: −0.004 to 0.014, insignificant, N=1,222) and confirms gains of ≈1pp/year from 2018–2024 under an instrument. But Table 7 finds those gains largest and only significant at firms with the LEAST initial organization capital (0.092 sales per worker, 0.105 TFP; organizationally rich firms ~0.05 and insignificant; N=875/836; mean reversion excluded by a 2014–2018 average sort and a null 2012–2018 placebo). Whether this supports or contradicts Claim 5 depends on an unresolved construct question: the authors' mechanism is codification headroom ("an established stock of OC simply leaves less to codify"), not organizational readiness. If organization capital is the Capability force, Breakpoint 4 is contradicted here at 875 firms; if it is headroom, the paper is not evidence for Claim 5 in either direction. Until Brandon rules, cite it as the position Claim 5 must beat, never as evidence for it. Note separately that Appendix E derives the multi-year lag from a frictionless stock-flow model, so a long lag is no longer, by itself, evidence of a breakpoint — see stafford-research knowledge/thesis-challenges.md 2026-10-29, The stock-flow objection, including the available repair (the null decade is flat, not a ramp).

And a standing note for the gap ranking: Claim 5's open gap is no longer "a non-technological organizational variable on the right-hand side" — one now exists. It is "a measure of organizational condition that is neither an accounting stock nor a self-report, against AI's realized outcome." That is a narrower and more expensive gap, and master-plan.md Tier 0.2 is the only thing in the queue that would produce it.

PROPOSED-UPDATE → research-evidence/knowledge/story-threads.md, Does the AI-investment laggard cohort converge?

Criterion (c) is DISCHARGED, and the answer changes the thread rather than closing it. Add to the log:

2026-10-29 (Stafford, 14:30): criterion (c) discharged — the primary was obtained and read in full from the authors' public replication repository (AI_productivity_BHJ.pdf, github.com/jrhao7/AI_productivity_replication; draft 2026-08-22). The paper does address convergence, and it addresses it by assumption. Appendix E's model makes the eventual productivity gain common to all adopters and only the speed state-dependent (Proposition 3, heterogeneity entirely through the pre-adoption ratio r = B/A), so every adopter converges to the same proportional gain and low-stock firms simply arrive sooner. A model that cannot represent a permanently non-converging firm is not evidence against a theory of permanent failure — so the discriminating test is still unrun, and this thread's value is unchanged rather than reduced. What the paper does supply is the closest thing to the test on real data, and the sign runs against Paper 2: Table 7's gains are largest and only significant at firms with the least initial organization capital, with mean reversion excluded. One bound on reading that as the test: the sample is surviving listed non-tech firms measured end to end, so firms whose AI programme destroyed them are absent from the denominator — and a breakpoints theory is a theory about those firms.

Proposed next check: 2026-11-29, not 2027-01-29. The original date was set to allow for a revision of w35684 or a follow-on, which was reasonable when the paper was unread. It is read now, criterion (c) is spent, and the live criteria — (a) a laggard cohort followed forward, (b) dispersion of AI-related productivity outcomes over time — are things other researchers publish on a monthly rhythm, not an annual one. A three-month deferral on the test the scan's own Routing calls "the strongest falsification test Paper 2 has" and "the most dangerous thing in the evidence base" is longer than the thread's own stated importance supports. Add criterion (g): any test of the timing profile — whether the pre-2018 null is flat or a shallow ramp — since that is the discriminating prediction between stock-flow accumulation and a breakpoint, and it is answerable in w35684's own published figures.

PROPOSED-UPDATE → research-evidence/knowledge/story-threads.md, The unvalidated proxy (criterion (f), added this morning)

Criterion (f) is PARTLY MET, by the paper that prompted it, and the thread should say so before it accumulates a second member on a false premise. Add to the log:

2026-10-29 (Stafford): criterion (f) — validation of a text-inferred organizational construct — is partly met by Babina, He & Jiang itself. This morning's entry recorded that "nobody has checked 'organization capital inferred from job descriptions' against anything a firm could not author." Read at primary, §2.2 runs five validation steps, including (i) a writing-quality placebo — OC-building descriptions are longer but more repetitive, cite fewer specific numbers and tools, show the same spelling-error rate, and are indistinguishable from non-OC descriptions of the same length on every writing-quality measure; (ii) intuitive variation by function — 79% of AI workers in management roles classify as OC-building against 22.6% in product support, and only about half of core AI-development roles; and (iii) external validation against audited capitalized SG&A (§4.2, Table C.32) — firms with more OC-building AI jobs subsequently accumulate more organization capital on the accounting measure, while non-OC-building AI jobs do not.

The honest limit, which keeps the thread open rather than closing it. Capitalized SG&A is itself the standard accounting proxy for organization capital (Eisfeldt & Papanikolaou), not an observed organizational property — this is a text proxy validated against an accounting proxy, and nobody has compared either to a directly observed organizational fact. Note also that capitalized SG&A plays two roles in the same paper: it validates the text measure and defines Table 7's heterogeneity groups. Criterion (f) should therefore be restated as validation of a text-inferred organizational construct against a directly observed organizational property, not against a second proxy — which is stricter than it was written this morning and is the version the lane actually needs.

PROPOSED-UPDATE → research-evidence/knowledge/emerging-patterns.md, The Résumé-Panel Consensus

One qualifier the lane now needs, and it cuts in the sources' favour. This morning's entry widened the lane in domain — from inferring organizational form to inferring a capability stock — and recorded that the heavier claim is carried by the same data "with no validation anywhere." That last clause is wrong at primary and should be struck. Babina, He & Jiang validate the text-inferred construct five ways including against audited capitalized SG&A, and they run a writing-quality placebo specifically against the obvious artefact (that the measure picks up people who write better job adverts). Proposed amendment:

Amendment 2026-10-29. The lane's newest member is also its best-validated, and the pattern should record that rather than deepen the worry on a member that answered it. Babina, He & Jiang (WP 35684) validate the text-inferred organization-capital measure in five steps including a writing-quality placebo and external validation against audited capitalized SG&A. The lane's standing worry therefore narrows from "nobody validates these constructs" to "nobody validates them against anything but another proxy" — which is still a real hole and is a much more specific one. The non-independence rule is unchanged and binding: WP 35684 shares two of four authors and the résumé substrate with NBER 31325; never cite them as independent confirmations. And one addition to the standing instrument check: where a paper validates a novel proxy against a second proxy, record which second proxy and what else it is used for in the same paper — here capitalized SG&A both validates the text measure and defines the headline heterogeneity split, which is not circular but is worth seeing on the page.

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

  1. ADD — public replication repositories as a retrieval channel. First seen 2026-10-29, surfaced by the w35684 chase. Cadence irregular, no expiry clock. Not an outlet — a route, and the second route logged this week after third-party PDF mirrors on 10-28. Record the operational facts: nber.org's PDF endpoint, CEPR and SSRN all refuse this environment; both authors' own websites link only to the refusing SSRN page, so hard rule 14's prescribed author route failed here; the manuscript was in the authors' public replication repository (github.com/jrhao7/AI_productivity_replication, AI_productivity_BHJ.pdf, HTTP 200), reachable in two clicks from taniababina.com/data.html. Hazard to record: a repo copy can be a different draft from the published version, so the version check is mandatory — here the PDF's own footer ("This draft: August 22, 2026") matched the landing page and the repo had a single commit.
  2. ITIF Fact of the Week — one note, no status change. Added on probation this morning with its pages unfetchable. Its reported figure (≈1pp of additional annual productivity growth) is now confirmed at the primary, which says the summary function is accurate even though the page cannot be read from here. That is worth recording against the candidate, but it is not a second signal and should not advance promotion.

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

One line, and it changes what she is in the map rather than adding a fact. She is currently positioned as the builder of the résumé-inferred firm-level AI measure and, since this morning, as holding a productivity-J-curve reading of her own coefficient. Add:

2026-10-29 (Stafford, primary read). Her position is stronger than a J-curve reading and should be recorded as what it is: Appendix E of WP 35684 derives the multi-year lag between AI investment and productivity gain from a model containing no organizational friction of any kind — "what a stock-flow system delivers when stocks are large relative to flows" — with half-lives of five years from depreciation alone. She is not arguing for patience; she is showing the delay needs no organizational explanation. That makes her the field map's first entry offering a null model rather than a rival theory of failure, which is a harder competitor than any framework in the map. Brandon's differentiation, restated accordingly: not patience versus design, but ramp versus step — a stock quietly accumulating predicts shallow positives through the null decade, and her own Table 2 Panel A reports −0.004 and 0.014. The differentiation is a checkable claim about a timing profile in her published figures.