Stafford Brief — 2026-10-16 (Friday)

Read the scan's brief first (research-evidence/briefs/2026-10-16.md). It found two genuinely important experiments and I am not repeating them. But it told you to walk into rooms this week leading with a number it had not read the paper for, and that number is the wrong one to lead with. That is the first item below and it is the only one that is urgent.

Four things today.

1. The headline figure is a side result, not the finding. The scan's lead — "told how others reorganized, they earned 1.9 times the revenue" — comes from a paper nobody has read, because the publisher blocked it. I found the study's public pre-registration. Revenue was not one of the three results the authors committed in advance to measuring, and their own plan said to trim the extreme top of the revenue distribution — which is exactly where the 1.9x comes from. There is a better number in the same experiment, it was pre-committed, and it is the one that actually says what you want to say.

2. Chasing that paper's own footnotes found the gap the scan has been failing to fill for twelve straight days. Its weekly synthesis lists "comparing AI-native organizations to conventional ones" as untouched since early October. The answer is a Harvard working paper by one of the same authors, published in June, sitting unblocked on Harvard's own server. I read it end to end.

3. That paper contains the single best sentence this evidence base has ever had for your argument — and it is a null result. Firms whose staff demonstrably use ChatGPT, Copilot and Cursor are not smaller, not flatter, and do not have fewer managers. The firms that are flat got flat a different way. This belongs at the front of The Flattening Bet.

4. The same paper is the strongest attack on your framework's audience that I have logged. It is below under CHALLENGE, at full strength, logged OPEN. Three things limit it and two of the three are the authors' own words.

📖 TODAY'S READ

Giving your people AI tools does not make your organization any flatter. The companies that are flat didn't get there by adopting AI — they got there by selling it. That is a measured result, from the only study that has ever separated those two things, and it is your Technology Illusion with a number attached.

Kim H & Koning R, "AI-Native Firms," Harvard Business School Working Paper 26-090, 9 June 2026. Read in full. knowledge-base/hbs-kim-koning-ai-native-firms-2026-06.md

They take every Y Combinator startup from early 2020 to late 2024, plus every US venture-backed startup first funded in those years, and compare the ones that call themselves AI companies against the ones that don't — matched to the same industry and the same year, so they are comparing like with like. Then they attach each company's actual staff records.

The AI companies are much smaller organizations doing comparable business. About a quarter fewer employees — roughly 13 people against 42. Half a management level less depth. About 15% fewer managers, even after allowing for the size difference. More engineers, fewer junior people, more senior people. And here is the part that stops it being a story about failing companies: they raise the same money and carry the same valuations. Same business value, much less organization.

Then they do the thing nobody else has done, and it is the reason this matters to you. They split "being an AI company" into two completely different things. One: the company's own staff use AI tools at work — measured by whether its job ads name ChatGPT, Copilot, Cursor and thirty others. Two: AI is inside the product the company sells. Then they ask which of the two actually explains the smaller, flatter organization.

It is entirely the second one. Companies that build AI into their product are smaller and flatter by every measure. Companies whose staff merely use AI tools are not — not smaller, not flatter, no fewer managers. The authors write it plainly: "equipping workers with ChatGPT, Copilot, or Cursor does not, on its own, predict smaller firms."

Technical detail: 1,666 firms, 467 self-identified AI. Product channel: about 10 fewer employees, 13% smaller in logs, 19% in Poisson, 0.2 fewer hierarchy layers, all statistically significant. Process channel: nothing significant on size, depth or manager share. Both channels in the same model, with industry-and-cohort controls.

What it changes for you, concretely. The Flattening Bet currently argues the deployment-impact wall from surveys — the 88%-deploying-and-81%-seeing-nothing family, Accenture's falling value line. Those are all arguments about performance. This is about structure, which is what a delayering essay is actually about. Promote it to the lead of the evidence section and let the surveys sit behind it. Amendment filed: knowledge/asset-suggestions.md, 2026-10-16.

And one thing you may not do with it, which I have made binding in the ledger. The staff records come from Revelio Labs — the same résumé-derived data genre as Ewens & Giroud, on substantially the same population of US venture-backed startups, read from a single snapshot in January 2025. Never cite this as independent confirmation of Ewens & Giroud's delayering result. Two results from one data genre on one population is the error this base charges other people with, and a flattening essay holding two flattening results is exactly where it would get committed.

🔭 THINKER TO WATCH

Hyunjin Kim (INSEAD) and Rembrand Koning (Harvard Business School).

The scan added them to the field map this morning for the randomized experiment. That is the lesser of their two papers and the field map should say so. The pair have quietly assembled the only research programme that measures both halves of the question you work on: one paper randomizes telling firms how to reorganize around AI and measures what it pays, the other measures what firms built around AI actually look like inside. Kim is at INSEAD alongside Puranam, which means three of the four people now doing serious empirical work on AI and organizational structure sit in one building.

Why they matter now rather than eventually. They are the ones who decomposed product from process. That decomposition is the move your framework needs and nobody else in this base has made it.

How your position differs, and it is a structural advantage rather than a disagreement. They study startups — explicitly, and for a stated reason: "new firms are less constrained by legacy structures." Every finding they have is about organizations with no legacy incentive system, no entrenched handoffs, and no governance body with a veto. They have deliberately chosen the population in which the Five Breakpoints bind least, and they say so. Their own closing limitation is that they cannot tell whether these shapes survive growth — "whether the demands of coordination reassert themselves, is something we cannot yet address."

That is the whole gap, and you are on the right side of it. They have the measurement apparatus and no access to incumbents. You have access to incumbents and no measurement apparatus. The study that would settle both their open question and yours is one instrument away, and the instrument is Tier 0.2 on your own master plan.

Field-map correction proposed in Routing — the entry exists but is filed under the weaker paper.

🏢 CASE IN THE WILD

Two companies selling the same thing to the same customers. One has 912 employees. The other has seven.

Careerist delivers career courses through third-party tutors. Educato AI delivers exam preparation with AI doing the teaching. The authors found the pair by searching for companies whose own descriptions of their businesses sit closest together in meaning, then looked at the headcounts. 912 against 7. The same exercise produced Legion Health, an AI psychiatry platform with 28 employees, against a conventional mental-health firm with a far larger workforce.

The mechanism is Technology Illusion inverted, and it is worth being precise about it. Careerist's 912 people are not waste. They are the product — recruiting tutors, scheduling them, quality-checking them, handling the failures. That is what the organization is. Educato did not organize that work better, or automate it, or run a transformation. It moved the work out of the organization and into the thing it sells. Which is why the pattern is strongest exactly where you would predict: in services businesses that historically scaled by hiring people to deliver the service, AI firms are roughly 70% smaller with hierarchies nearly a full level shallower.

The open question, and it is the one I would put in a client room this week rather than the 1.9x. Careerist cannot become Educato by giving its 912 people AI tools — that is the null result above, measured. It can only become Educato by changing what it sells. So: how much of your headcount exists to deliver something a product could now deliver directly — and is that a transformation question at all, or a product question you have been routing to the wrong function? That question costs an executive nothing to answer and it is very uncomfortable, which is the combination you want.

⚙️ FOUR FORCES CONCLUSION

Capability, and this paper moves it somewhere the framework has never put it: out of the organization altogether.

Every Capability finding in this base locates capability in people or in practices — skills, training, management quality, infrastructure. Kim and Koning's central claim is that AI lets a firm "shift some of this processing from the organization into the product itself, allowing the firm to deliver knowledge-work capabilities without building them internally through people and hierarchy." And they measure it: comparable valuations on a quarter fewer staff and half a level less hierarchy.

What that does to the paper-2 argument. Capability has been the Force this base can evidence most easily. It is now also the Force with the most awkward evidence, because a firm can be highly capable with almost no organizational capability — the capability is in the product. The Four Forces are defined as properties of an organization. This is the first instrument suggesting one of them can be satisfied outside it. That does not break the framework; it means Capability needs a sentence saying whose capability, and the sentence does not currently exist.

No Purpose, Commitment or Momentum conclusion, and today that absence is worth one line rather than none. The scan established this morning that no firm-level study has moved a real outcome by a Purpose or Commitment variable, after three attempts. Today's paper does not break the run. Momentum is the interesting one: the authors pose its exact question as their own closing limitation — do these flat shapes survive scale, or do coordination demands come back? — and then say they cannot answer it. That is a gap in the literature stated by the literature, not evidence. Tagging Momentum off it would be the loose tagging the rules forbid.

💭 OPEN QUESTION

If the flat organization comes from what the firm sells rather than from how it works, then is "AI-native organizational design" a category error — and is the thing you actually diagnose a product decision wearing an org chart's clothes?

This is not rhetorical and I do not know the answer. The scan's open question this morning asked which ingredient of a bundled AI deployment did the work. Today's paper asks something harder and more awkward for the framework: the one firm-level study that decomposes AI's organizational effect finds the effect does not come from how the organization works at all. It comes from the boundary of the firm — what is inside the product and what is inside the payroll.

Two readings, and they lead to different practices. If the AI-native form is a product-architecture fact, then your five breakpoints are a theory of why incumbents cannot execute a boundary change, which is a sharper and more defensible claim than the one you make now. If instead the product channel is doing the work only because tool adoption is still shallow — and 13.5% of these AI firms naming a tool in a job ad is shallow — then the process channel is undermeasured rather than absent, and the null will not survive better data. The second reading is the comfortable one and I do not think it is the stronger one, because the product-channel effect is large and the process-channel estimate is not merely insignificant, it points the wrong way on levels.

🥊 CHALLENGE

Nobody transformed into an AI-native organization. They were founded as one — and they are eating your clients' lunch while your clients consider a readiness assessment.

Logged knowledge/thesis-challenges.md 2026-10-16, OPEN, as the entry objection. At full strength: the AI-native firms in this paper are a quarter smaller, half a level flatter, carry 15% fewer managers and hold the same valuations — and not one of them fixed anything. They had no vague purpose to clarify, no fragmented incentives to realign, no legacy handoff chain to unblock. The organizational form your framework presents as the hard-won end state of a failure-prone change programme is being reached routinely by the expedient of starting from scratch, by the companies competing with the people you sell to. Educato's seven against Careerist's 912 is the competitive version of that sentence. If the AI-era organizational advantage accrues through entry rather than repair, then a framework addressed to incumbent repair is addressed to precisely the population not realizing the advantage, and the honest product is a warning, not a method.

Three things limit it. Two are the authors' own and I did not have to argue for them.

First, they disclaim the causal inference the objection needs. "Our cross-sectional design cannot pin down causal mechanisms: the patterns we observe could stem from the product architecture, from AI-enabled efficiencies in internal operations, or from selection of different types of founders into AI ventures." Being an AI company is self-declared — a tag the founders chose — and the authors concede it "may include firms that strategically adopt the label." The label is also moving fast: about 15% of early-2020 YC batches called themselves AI companies, against over half of the 2024 batches. Some of this is a renaming wave.

Second, persistence is unmeasured, and this is the limit that matters most. "Our data covers firms for at most a few years; whether these organizational forms persist as firms scale, or whether the demands of coordination reassert themselves, is something we cannot yet address." The objection is a claim about a steady state drawn from one snapshot of very young companies. And every one of the four flat-organization reversions in the exhausted-alternatives objection — Haier, Spotify, Zappos, Valve — happened at scale, which is the scale this data cannot reach. Those two challenges are the same challenge from opposite ends and they constrain each other: if coordination reasserts itself, the entry objection collapses into your thesis on a delay. Answer them in one paragraph, not two.

Third, and this is the hinge: the mechanism is the product, not the organization. These are not better-organized adopters of AI — the null result says tool adoption predicts nothing structural. They are firms that sell the work instead of staffing it. That is a product-market substitution threat to incumbents and not an organizational design an incumbent could copy. Which bounds what the objection can claim, and supplies the response.

What would make it win: evidence that incumbents who did repair organizational conditions still failed to close the gap against AI-native entrants — the prerequisite satisfied and the advantage not following. What would defeat it: the same firms tracked over time, showing either that AI-native firms re-layer as they scale, or that incumbents who fixed conditions first captured the gain. Those are one study, and you can enrol in half of it. Tier 0.2 is the only item on your master plan that would produce an incumbent-side data point at all. The scan opened a thread this morning predicting that the mapping effect attenuates in incumbents and that "the attenuation is the Five Breakpoints measured" — good prediction, untestable until somebody measures an incumbent, and you are the one with access to incumbents.

My recommended response concedes the addressee and keeps the mechanism. Do not contest the finding; it is the cleanest Claim 3 evidence available and it is on your side of Technology Illusion. Concede that the AI-native advantage is a product-architecture advantage rather than an organizational-practice one — so an incumbent can neither buy it with tools nor reach it by delayering. Then the sentence that falls out is better than anything currently in the draft: an incumbent that delayers without moving work into a product has cut the layers and kept the work. That is falsifiable, it is your prediction stated as a mechanism, and it is simultaneously the discriminating answer the exhausted-alternatives objection has needed since 10 October.

📈 PATTERN BUILDING

Two proposals, and the second one prevents a false pattern member rather than adding a real one.

1. The Résumé-Panel Consensus gains a fourth member, and the fourth is the reason the pattern exists. Kim & Koning use Revelio Labs workforce records on US venture-backed startups. That is the same data genre as Babina et al. and Ewens & Giroud, on an overlapping population, from a single January 2025 cross-section. The pattern's warning is that this base's firm-level structural literature is largely one data genre wearing several bylines. Today it nearly caught us in the most expensive possible place: a committed essay about flattening, whose Ewens & Giroud verification was just discharged, acquiring a second flattening result that looks independent and is not. Proposal in Routing, with the usage note binding.

2. The Randomized Organizational Lever must not go to three on a bookkeeping accident. The Büchsenschuss/Puranam field experiment circulates under two different titles — as an INSEAD working paper ("The Impact of Generative AI Adoption on Organizational Networks") and on SSRN as "GenAI Adoption Increases the Density of Knowledge and Collaboration Networks" (SSRN 6028034), same four authors, same experiment. The staging pipeline de-duplicates on title and URL, so the SSRN version will stage as a separate entry and push this morning's new two-source pattern to three. It is one study. Flagging it before it happens rather than correcting it after. The pattern's own kill test is a randomized null in an incumbent — not a third citation.

🧱 ASSET SUGGESTION

No new asset. Today's find belongs inside the committed one and it changes its spine, which is worth more than a nineteenth entry.

Master-plan hard stop 1 is unconditional and asset 2 is Tier 1.1, in flight. The pitch, for the amendment rather than a new asset: The Flattening Bet acquires its first structural evidence (everything in it until now argued performance), its strongest single sentence (the process-channel null), and its missing discriminating claim — an incumbent that delayers without moving work into a product has cut the layers and kept the work. It also acquires one binding constraint on how Ewens & Giroud and this paper may be cited together.

Punch-list status: unchanged where it counts. Items 1, 3 and 5 stay discharged. Item 4 is yours alone and is now twenty-three days old — whether the two-year prediction and the stated losing conditions go into print dated. The 10-10 prose gap (the exhausted-alternatives paragraph) got cheaper today because the CHALLENGE section above supplies its discriminating sentence. The research half of this asset has now gained its best evidence and its best sentence on the same day. What is left is one decision and two paragraphs.

Full note: knowledge/asset-suggestions.md, 2026-10-16.

🎓 SPARRING NOTE

Not issued — tenth consecutive decline, and I am not going to explain the reason again at length. Today's material exercises growth edge #2 (measurement design) squarely: the product/process decomposition is a measurement move, and it is the kind of move the readiness instrument needs. The hook is parked in knowledge/brandon-development.md.

The loop has been switched off since 25 September and the unblock has not changed. Master-plan Tier 0.1: answer the 09-17 note, fifteen minutes, one paragraph. It is also the first draft of the Tier 2 gate argument. It is four weeks old.

SOURCES & THREADS

Threads: nothing due that the scan did not cover. It advanced the amplification thread (criterion met, a day early, and corrected its own 09-17 conclusion rather than dropping it — right call), opened the incumbent thread, and cross-referenced two more. Queue is 10-17, 10-20, 10-28. I opened nothing and chased nothing, because there was nothing uncovered.

Exploratory slot: spent on verification of the scan's own headline, and it paid four times. Logged in full in knowledge/query-log.md. Yield: the pre-registration catch, one KB file on the scan's twelve-day-old Claim 3 gap, the two-title de-dup hazard, and one method result.

The method result, because it is reusable and it cost the scan its best find today. The scan filed Mapping AI into Production at abstract-grade after SSRN and both HBS endpoints returned 403. It then wrote a weekly synthesis saying Claim 3 was untouched for twelve scans. The answer to that gap is cited in the blocked paper's own introduction and published openly by the same author. A 403 does not just cost you the figures — it costs you the references section, and CLAUDE.md §Source Discovery's first rule (chase citations to origin) cannot be discharged on an abstract. Hard rule 14 already prescribes the fix and it was not applied today to either blocked paper: go to the author's own copy. hbs.edu/ris/Publication Files/ served a 1.2MB PDF with no block at all. That is now three recent headline ingests filed abstract-grade behind a 403 (Yuan 10-13, Bick today, Kim/Kim/Koning today) where an institutional route was not tried.

One correction to the scan's own figures, small but client-facing. Its brief gives the Büchsenschuss experiment as "316 employees" in one section and "333 employees" in another, and the staged entry's arms sum to 318. The figure is 316. Use the paper, not the brief.

Candidates: 0 added, 0 promoted, 0 expired by me — and one failed route recorded so it is not retried blindly. The scan added the INSEAD working-paper series today with the promotion condition "find a dated index route." insead.edu/faculty-research/publications/working-papers returns HTTP 404. The condition is not discharged. Named-slug and doc.cfm?did= routes exist for individual papers, so INSEAD papers can be read but the series still cannot be checked.

Weekly Synthesis

1. Patterns. The scan's two moves are right and I am not re-litigating them (The Unreliable Witness to ESCALATED with its genre caveat; The Randomized Organizational Lever opened at 2 and deliberately not escalated). My two additions are above: a fourth member for The Résumé-Panel Consensus, and a prevented false third member for The Randomized Organizational Lever. Net: the week's pattern work was unusually disciplined — two refusals, one escalation on a pre-registered condition, and one pattern that declined to escalate itself. That is the behaviour the ledger has been asking for.

2. Paper evidence. The scan's synthesis says Claim 5's standing vulnerability is closed. Overstated, by one step. The objection it closes — "well-run firms are just better at everything" — is genuinely answered by random assignment, and that is real. But it is answered by the discovery result, which was pre-committed, and not by the revenue ratio the brief leads with, which was not. Claim 5's answer is one citation and the citation is "44% more use cases," not "1.9x revenue." Separately, Claim 3 is no longer untouched — twelve scans became one citation chase, and the result splits: the process-channel null arms Technology Illusion at firm level for the first time, while the entry objection attacks the framework's addressee. Remaining gaps, re-ranked: (i) the Purpose/Commitment firm-level moderator — now a fourth consecutive confirmed absence and correctly reclassified as a stated paper gap rather than a daily query; (ii) the incumbent-side measurement, which has been promoted by today's work from a gap in the literature to the single item that four separate challenges now terminate in, and which no amount of searching will fill because nobody is running it; (iii) within-firm dispersion, unchanged, method captured 10-16 but instrument unbuilt.

3. Field map. The scan added four entries including Kim, Kim & Koning. The entry is filed under the weaker of their two papers — correction proposed in Routing. No stale entries surfaced by me this week. The Block/Sequoia double-entry defect flagged on 10-10 is still unfixed.

4. Windows. None opened. Twenty-seventh consecutive null window flag. The scan has correctly stopped treating this as a tally and started treating it as a finding; I agree, and the finding is that this base does not generate time-priced opportunities, it generates argument-priced ones. Nothing in the window file needs Brandon this week.

5. Staleness. The scan flagged Bick ("as of late 2024") and Büchsenschuss (fieldwork Sept–Dec 2024, published 2026). Both correct. I am adding a third and it is the one most likely to be cited carelessly: today's Kim & Koning is a single January 2025 workforce snapshot, which is twenty-one months old and sits entirely before the 2025–26 agentic-tooling wave. The process-channel null is measured in a period when internal AI tooling was shallow — 13.5% of AI firms naming any tool in a job ad. That is the null's largest vulnerability and it must be dated in every citation: "as of early 2025." Refresh query queued.

6. Challenges audit — this is the item I would put in front of you, and it is the seventh consecutive Friday it has been true. The ledger now stands at 40 OPEN, 9 CONTESTED, zero ABSORBED, zero DEFEATED. On 10 October the master plan recorded 43 live entries and zero absorbed, and named that as "the clearest single sign that the ledger accumulates and does not discharge." Six days later it is 49 live entries and still zero absorbed. I added one today, which makes me part of the problem rather than an observer of it. Nine OPEN challenges are more than thirty days old with no response plan, and six of them are the single question the master plan identified as one question — the Tier 2 gate. The quarter's success condition is "six challenges moved from OPEN to ABSORBED with the revision named." Day seven of that quarter, the number is zero and the denominator grew by six.

7. Positioning intelligence. The scan's two reversible promotions (Management Science 10-13, CESifo 10-14) are now four and three scans deep respectively and the Management Science listing obligation is genuinely discharged by substitute — that resolves toward keeping them and I would stop carrying it as open. But the INSEAD promotion recorded this morning should not be read as sound: its own stated condition is undischarged and the route I tried 404s. On the scan's own Category 6 method question — narrow the exploratory slot from keyword sweeps to named channel listings — I now recommend adopting it with one amendment, and the amendment is today's whole lesson. Add the author's own institutional page as a mandatory route whenever a publisher blocks a paper. Both of today's inaccessible papers were openly served by their authors' institutions, and one of them contained the answer to a gap the scan has reported as unfillable for twelve days. That is not a search problem. It is a routing problem, and hard rule 14 already wrote the rule down.

Routing

strategy — change the sentence before you use it, and the replacement is better. Do not lead with "told how others reorganized, they earned 1.9 times the revenue." The study's public pre-registration (AEARCTR-0016746) names three results the authors committed in advance to measuring: the number of AI use cases found, an index of venture progress, and six-month survival. Revenue is not among them — it is one of four yes/no components inside the progress index, and a secondary outcome. Worse for the sentence: the authors' own plan said to trim the extreme top of the revenue distribution before analysing it, and the paper's abstract concedes the revenue gains are "largest at the 90th percentile and above." So the headline number comes from exactly the tail the pre-analysis plan was written to discount, and a hostile reader with ten minutes finds that. Lead with this instead: "In a randomized experiment, startups that were simply shown how other firms had reorganized production around AI found 44% more ways to use it — and that was the result the researchers committed to measuring in advance." It is pre-committed, it is the discovery effect, and it is the better fit for your argument anyway, because your claim is about finding where the work needs rearranging, not about revenue. Two further corrections to carry: the treatment was an eight-week programme (9 Sept – 17 Oct 2025), not three months; and the population was ventures enrolled in an AI founder sprint, scored for traction on a scale starting at zero — "515 high-growth startups" is a gloss the paper does not support.

strategy — the Flattening Bet's evidence section should be rebuilt around a null result, and there is a binding citation constraint. Both in knowledge/asset-suggestions.md (2026-10-16). Short version: promote the process-channel null to the lead, write in "an incumbent that delayers without moving work into a product has cut the layers and kept the work," and never present Kim & Koning and Ewens & Giroud as two independent flattening results — same data genre, same population, one January 2025 snapshot.

governance — the client question I would use this week in place of the 1.9x. How much of your headcount exists to deliver something a product could now deliver directly? Behind it: two companies selling the same service, 912 employees against 7, and a measured finding that giving the 912 AI tools does not change the shape of the organization. It is a product question that arrives disguised as an org-design question, which is why it lands.

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

Fourth member, and it arrives in the most expensive possible place. Kim H & Koning R, AI-Native Firms (HBS WP 26-090, 2026-06-09) measures firm size, functional mix, seniority and hierarchy depth from Revelio Labs position-level records — résumé/workforce inference — on US venture-backed startups first financed 2020–2024, read from a single January 2025 cross-section. That is the same data genre as Babina et al. and Ewens & Giroud and an overlapping population. Binding usage note: never cite Kim & Koning's "half a seniority level flatter" as independent confirmation of Ewens & Giroud's delayering result; cite them as one instrument genre with two applications, or cite one. This matters now because stafford-research asset 2 (The Flattening Bet, COMMITTED) is a delayering essay that holds both results and whose Ewens & Giroud verification was recently discharged. File: knowledge-base/hbs-kim-koning-ai-native-firms-2026-06.md.

PROPOSED-UPDATE → research-evidence knowledge/emerging-patterns.md, The Randomized Organizational Lever (a de-dup guard, not a count change):

Do not let this pattern reach 3 on a title change. The Büchsenschuss, Koch-Bayram, Biemann & Puranam field experiment circulates under two titles: the INSEAD working paper "The Impact of Generative AI Adoption on Organizational Networks: Evidence From A Field Experiment" and SSRN 6028034 "GenAI Adoption Increases the Density of Knowledge and Collaboration Networks: Evidence from a Field Experiment." Same four authors, same single-firm cluster-randomized experiment, same 42 teams. The staging pipeline de-duplicates on title and URL and will stage the SSRN version as a separate entry. Record both titles against the one member now. Also: the headcount is 316 — today's brief gives 333 in its CASE section and the staged entry's arms sum to 318.

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

(1) Claim 3 is no longer untouched, and it resolves in two directions. Attach knowledge-base/hbs-kim-koning-ai-native-firms-2026-06.md. For the claim: the process-channel null (1,666 firms; internal AI tool adoption does not significantly predict headcount, hierarchy depth or manager share once the product channel is controlled; only the product channel does) is the base's first firm-level structural evidence for Technology Illusion. Against the claim: the AI-native structural advantage runs through product architecture, not organizational practice, so it is not reachable by an incumbent — logged as the entry objection, stafford-research knowledge/thesis-challenges.md 2026-10-16. Usage bound: date every citation "as of early 2025" — one January 2025 snapshot, before the agentic-tooling wave, with only 13.5% of AI firms naming any internal tool. (2) Claim 5's citation should be the discovery result, not the revenue ratio. Kim, Kim & Koning's pre-registration (AEARCTR-0016746) lists AI use cases, a venture-progress index and six-month survival as the pre-committed outcomes; revenue is a secondary outcome and a binary component of the index, and the pre-analysis plan specifies winsorizing financial metrics at the 90th/95th percentiles while the abstract states revenue gains are largest above the 90th percentile. Cite "44% more AI use cases" as the Claim 5 anchor and carry revenue as supporting with the tail caveat. Also correct the population description: ventures enrolled in an eight-week AI founder sprint (2025-09-09 → 2025-10-17; 585 enrolled, ~500 expected after attrition), stratified by continent × traction score (0–3) × baseline AI use cases — not "515 high-growth startups."

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

Re-file the Kim/Koning entry under the stronger paper. Hyunjin Kim (INSEAD) and Rembrand Koning (HBS) are tracked as of 2026-10-16 for Mapping AI into Production; their more consequential work for this base is AI-Native Firms (HBS WP 26-090, 2026-06-09), the only study that decomposes AI's organizational effect into a product channel and a process channel. Why relevant: they have both halves of the measurement apparatus this base lacks. Brandon's differentiation: they study startups by explicit design — "new firms are less constrained by legacy structures" — i.e. the population where the Five Breakpoints bind least, and their own stated limitation is that they cannot tell whether these forms survive scale. Watch for: any panel or second cross-section of the same firms; that is the study that resolves both the entry objection and the exhausted-alternatives objection. Note: Kim is at INSEAD alongside Puranam.

PROPOSED-UPDATE → research-evidence knowledge/source-candidates.md, INSEAD working-paper series:

Promotion condition recorded this morning is NOT discharged. The condition was "find a dated index route." https://www.insead.edu/faculty-research/publications/working-papers returns HTTP 404 (checked 2026-10-16, Stafford). Individual papers are reachable by named slug and by sites.insead.edu/facultyresearch/research/doc.cfm?did=<n>, so the series can be read but still not checked. Keep at 1 signal, cadence continuous, condition open. Do not retry the /faculty-research/publications/working-papers path.

PROPOSED-UPDATE → research-evidence knowledge/story-threads.md, the incumbent thread opened this morning:

The thread's prediction (the discovery effect largely survives, the outcome effect attenuates sharply, and the attenuation is the Five Breakpoints measured) is good and currently untestable — no incumbent-side instrument exists and nobody is building one. Note on the thread that the only live route to its evidence is Brandon's own post-engagement measurement (stafford-research knowledge/master-plan.md Tier 0.2), not a search. Four challenges now terminate in that instrument. A thread whose evidence can only be produced by the person reading the thread should say so, or it will be re-checked monthly forever.

personal — the one number that matters this Friday, and I am implicated in it. The challenge ledger is at 49 live entries and zero absorbed, six days after the master plan named exactly that ratio as the clearest sign the system accumulates and does not discharge. I added the fiftieth today and it was the right call on the evidence, which is precisely the problem: nothing in my job stops the denominator growing. Six of those challenges are one question, the Tier 2 gate answers it, and Tier 0.1 — fifteen minutes, one paragraph, four weeks old — is its first draft and the unblock for the coaching loop. I have no new recommendation. The old one has not been acted on and it is still the highest-leverage item on the list.

personal — coaching loop, tenth consecutive decline, no new note. Reason unchanged and parked in knowledge/brandon-development.md. Today's hook (growth edge #2, measurement design: the product/process decomposition as a measurement move) is held there rather than issued.

OPS — three items.

1. Yesterday's run merged cleanly. origin/main carries 623a5b6 (promote of claude/keen-fermi-5256bd, folding the 10-15 brief). No carry-over flag owed and no UNMERGED from 10-15.
2. Scope conflict between the scheduled prompt and CLAUDE.md, resolved in favour of CLAUDE.md for the second consecutive day, and flagged again because it is your call. The run prompt says Stafford has no write access to research-evidence. CLAUDE.md operating rule 3, set by you on 2026-10-02, says the opposite for KB files and staging, and records that the read-only belief stranded 45 of 49 KB files in September. I followed CLAUDE.md: one KB file copied into research-evidence and one entry staged there, on my session branch only. Shared meta files remain proposals, as above. This is the second day I have flagged it; if the prompt is current rather than stale, say so and I will stop writing there.
3. research-evidence branch claude/modest-faraday-rqjyjq — carries knowledge-base/hbs-kim-koning-ai-native-firms-2026-06.md and the staged entry 2026-10-16-harvard-busines-ai-native-firms (pending, published_date: 2026-06-09, instrument.shared_rhs_with set to the Ewens & Giroud September 2026 revision so the hard-rule-9 constraint travels with the entry rather than only with the brief). Pushed to that branch and not to its main; its promote-claude-branches workflow folds it in once scripts/verify.sh passes. Disposition in this run's final line.