Stafford

← 2026-10-01 · all briefs · 2026-10-03 →

Stafford Brief — 2026-10-02

Friday. Delta layer on the 12:00 UTC scan (2 staged, 12+ skipped, weekly synthesis run) — read that brief first; this one does not repeat it. No merge OPS flag: yesterday's run landed clean (3291342, promoted in 8d72def).

One thing matters today and it undoes a piece of this morning's scan. The scan spent its citation-hygiene slot clearing a week-old hold on the most important structural paper in your evidence base, concluded the base's figures were correct and the alarm false, and told the base to never cite the newer numbers. It read the wrong PDF. There is a third version of that paper, dated three weeks ago, and it changes the figures, the measurement, the time period, and the single sentence your flattening essay leans on hardest.

📖 TODAY'S READ

The paper everyone cites to prove AI is flattening organizations now says, in its own abstract, that organizations did not flatten.

Ewens & Giroud's Corporate Hierarchy is the one study that built a measure of management depth for U.S. public companies and then tested whether companies flatten after adopting AI. It is load-bearing in three places in your base and quoted in two of your drafts. It has now been revised three times in fifteen months, and this morning's scan reconciled it against the oldest version.

The current version is dated September 2026 and I read it in full from the authors' own copy. What changed:

Aug 2025 Oct 2025 Sept 2026
Companies 2,500+ 3,100+ 3,454
Résumés 16M 7M 18.1M
Average layers ten ten six
Years covered — — 2015–2024
AI result 10% level only 10% level only 1% and 5% levels

Two of those rows are the story. The measure was rebuilt — the average company now has six management layers, not ten — so this is not a bigger sample, it is a different instrument. And the AI result got stronger: the authors' own words "the tests are under-powered" and "significant at the 10% level regardless of the metric" have been deleted from the paper. Three of the four estimates are now significant at the 1% level.

But the two new sentences the authors added are worth more to you than their stronger result, and they point the other way. First, the trend across all 3,454 companies over the decade: layers were "flat to weakly increasing." Second, a new section disclosing what the AI finding actually rests on: only 19% of companies ever changed their management depth at all over ten years, in 6.4% of company-years — and when depth did change, it went up two times out of three. The AI analysis is "identified off the 654 firms whose layer count changes."

In plain terms: five companies in six never changed their management depth over a decade. Among the few that did, most added a layer. The flattening everyone quotes is a small conditional effect among AI adopters, measured on a sixth of the sample.

The breakpoint is Technology Illusion, and it is operating on the research literature rather than inside a company. A narrative that AI is dissolving organizational structure is being carried by a paper whose own aggregate finding is that structure barely moved and, when it moved, thickened. That gap — between what the abstract implies and what the sample shows — is where a technology story is doing work the data does not support, which is the mechanism your Paper 1 names.

So what: file knowledge-base/ewens-giroud-corporate-hierarchy-september-2026-revision-2026-09.md. Both drafts corrected in place today. CLAUDE.md rule 14 rewritten.

🥊 CHALLENGE

Your flattening essay's best line was false before anyone read it, and the fault is this system's.

drafts/the-flattening-bet.md said: "The flattening estimates in the most-cited paper are significant only at the 10 percent level, a limitation its authors state plainly and its citers rarely repeat." That was the sentence doing the most work in the essay's case that the flattening narrative is overbuilt. The authors have removed the limitation. Three of four estimates are now at the 1% level and the hedging language is gone.

Worse, this system explicitly reasoned that this could not happen. The 09-25 entry in asset-suggestions.md — written after a careful full read of two versions — concluded: "the essay's argument depends on the instrument being thin — all four coefficients significant at the 10 percent level only... Every one of those properties survives the revision, so the argument is untouched." It did not survive the next revision, three weeks later.

What it would take for this objection to win: nothing. It already landed. The only question is what replaces the sentence, and the replacement is in the same paper and is strictly better — "flat to weakly increasing", 19% of firms, 66% of changes upward, 654 firms. An argument built on what authors affirmatively report cannot be dissolved by those authors improving their estimates. An argument built on their hedging has an expiry date set by somebody else's revise-and-resubmit.

The rule, logged as durable: never build a published sentence on a working paper's weakness. Cite what the authors would defend — descriptive statistics, disclosed identification bases, the measures the paper does not contain. Those survive revision; significance levels do not. Full entry in knowledge/thesis-challenges.md, with the partial withdrawal of my own 09-25 reading.

And the upgrade is real, not consolation. You stop arguing about statistical strength and start arguing about scope. The first makes you a methodologist. The second makes you the only person in the room who read past the abstract.

🔭 THINKER TO WATCH

Ola Huitfeldt, Andreas Kostøl, Jan Nimczik and Andrea Weber — and the reason to watch them is that nobody in this literature has noticed they are upstream of all of it.

The September 2026 version states that its layer measure is built with "the network estimation methodology of Huitfeldt, Kostøl, Nimczik, and Weber (2023)." That method — infer job ranks from observed job transitions inside a firm, then group ranks into layers — is the thing that produces the number everyone argues about. When the average company went from ten layers to six between two versions of the same paper on overlapping data, nothing happened in the economy. Something happened in the estimator.

Why that matters now. Your instrument audit treats the four flattening papers as four instruments sharing a data genre. They also share a measurement layer, and it is a single 2023 methods paper with four authors and no stake in the AI debate. If the layer count is sensitive enough to move by 40% across revisions of one paper, then every structural claim in this literature inherits that sensitivity — and the audit currently has no row for it.

Brandon's differentiation is that he has no quarrel with them at all. They built a measurement tool and documented its uncertainty, which the September 2026 version repeats: ranks "come with statistical uncertainty because transitions can be promotions, demotions, and lateral moves." The people to disagree with are the ones citing layer counts to three significant figures as evidence about AI.

Honest limit, stated rather than smoothed: I have not read Huitfeldt et al. (2023). I have read that two papers' worth of structural findings are built on it. Candidate for your field map, and the first acquisition the instrument audit needs — one row explaining how sensitive the layer count is to the estimator, which would make that audit the only one of its kind.

🏢 CASE IN THE WILD

Nothing, and the scan's is better than anything I would add. Its agent-assurance read — an audited, insurer-funded certification market that certifies the supplier and nobody who deploys, with "the certificate is real, it is just not about them" — is the day's case and it is usable in a client conversation on Monday. I have nothing to put next to it that is not padding.

⚙️ FOUR FORCES CONCLUSION

Capability, and today's evidence narrows what the Force can be claimed to explain rather than widening it.

The flattening mechanism in this paper is pure Capability in your framework's exact sense: layers fall because cheaper knowledge lowers the need for problem-solver layers. There is no Purpose, Commitment or Momentum term anywhere in the specification — which is why the paper has always been a counterargument entry rather than support.

What today adds is the size of the thing that mechanism explains. Over a decade, across 3,454 companies, the Capability channel moved management depth in 654 of them. Everywhere else structure sat still, and where it moved it mostly thickened. So what for Paper 2: the honest claim is not that the Capability-only story is wrong — it is well estimated and now better powered. It is that the Capability-only story describes a sixth of the sample, and the framework's job is the other five-sixths, where structure did not move at all despite a decade of technology. That is a sharper position than treating the paper as a rival, and it is available because the authors disclosed their own identification base.

💭 OPEN QUESTION

If five companies in six never changed their management depth in ten years, what exactly was all the transformation activity doing?

This is not rhetorical and I do not have the answer. The decade 2015–2024 contains digital transformation, agile at scale, the remote-work rupture, delayering programmes announced at most large employers, and now AI. The one panel measure of management depth we have says the number barely moved, and moved upward more often than down.

Two readings, and they have opposite consequences for your practice:

  1. The measure is too coarse. Layer counts are a blunt instrument; the real change was in spans, decision rights and role content, none of which this paper measures — and it measures none of them in any version.
  2. The measure is fine and the transformation was Momentum Mirage at national scale. Depth is the most visible, most announced, most consultant-billed structural variable there is. If a decade of restructuring did not move it, the announcements and the structure were two different activities.

Reading 2 is your thesis stated at the level of an economy rather than a company, and you have never made that claim. The thing that would separate the two readings is a span-of-control series on the same companies over the same years — which is the acquisition your instrument audit has been missing since August, and which nobody has built.

📈 PATTERN BUILDING

The Résumé-Panel Consensus pattern loses one of its two defining properties, and should be amended rather than quietly weakened.

The pattern (ESCALATED, 9 sources) charges that the firm-level AI-and-structure literature is one data genre with pre-generative observation windows. For Ewens & Giroud, the window is now 2015 to 2024 — inside the generative era — and the vendors are named for the first time (Revelio and CoreSignal, not a generic "online résumés"). The pattern's source line records this member as 2010–2018. That is no longer true and the correction should be made before the next citation, not after someone checks.

The other half is untouched and in fact doubly earned. The AI variable on the right-hand side is still Babina et al.'s job-postings measure, confirmed from Table 10's own note. New today: footnote 30 cites Babina et al. (2025) as "consistent with our evidence." The authors are now citing as corroboration a paper from which they borrowed their explanatory variable. Hard rule 9 holds, and its justification is no longer an inference about shared genre — it is the paper's own table note plus the paper's own footnote.

So what: the pattern's charge is the shared instrument, not the dating. Amending the window makes it more accurate and no weaker. Proposed text in Routing.

🪟 WINDOW FLAG

None, and the scan's window closure this morning records the wrong cause of death. It closed the supervisor-pattern window as lapsed — correct, the piece was not written — on the stated ground that it had been "blocked on a single amendment paragraph for eighteen days," concluding that "this pipeline cannot clear it."

The paragraph has been written since 29 September. It is in knowledge/asset-suggestions.md, about 300 words, in your voice, conceding the sycophancy objection in full, and the entry filing it recorded condition 3 as MET with "nothing left to acquire." Yesterday's brief led with it. For the last three of those eighteen days the window was not blocked, and the pipeline had already cleared the named blocker — in a repository the scan cannot read.

So what, precisely, because the difference is actionable: the window died because the finished work was invisible to the system tracking it, and that invisibility was then written up as a finding about writing capacity. The scan's proposed remedy — name the writer and the last responsible day — would not have saved this one, because the writing was named and done. What was never assigned was writing the piece itself, which is a real finding and a different one. Recommendation unchanged, and the scan's salvage is right: fold the argument into the next long-form piece as the topology section. The gap is wider than when the window opened and nothing about it expires.

🧱 ASSET SUGGESTION

Asset 2 ("The Flattening Bet") gets a new opening paragraph, and it is the strongest thing the essay has had.

Replace the significance-level argument with the authors' own reporting: the single paper everyone cites to prove AI is flattening organizations reports that organizations, in aggregate, did not flatten. Five companies in six never changed their management depth over a decade. Among the few that did, most added a layer. The essay no longer has to argue the evidence is weak — it can grant the result is real, well-estimated and replicated, and still show the narrative built on it is many times larger than the finding. More generous, more persuasive, and immune to the next revision.

Full amendment in knowledge/asset-suggestions.md, with the ledger-level lesson and a flag that drafts/four-papers-one-method.md should get the same test before it ships — it is the asset whose entire structure is an argument about instruments.

🎓 SPARRING NOTE

Declined, eighth consecutive run. Cap does not bind — no notes issued since Monday — but six stand unanswered and a seventh makes the backlog the plan. Today's material is logged in knowledge/brandon-development.md as a sixth instance of one habit under Evidence evaluation, and it is the first instance where the habit failed in Stafford's work rather than in a source. Standing recommendation, now four weeks old: one /coach session, answer the 09-17 note, retire the rest.

SOURCES & THREADS

Stafford's own searching: one instrument re-verification and one exploratory query. Both logged in knowledge/query-log.md.

Weekly Synthesis

The scan ran the shared-file synthesis this morning (patterns, case file, field map, windows, staleness, positioning) and I do not repeat it. This is the judgment-layer synthesis on the four files it cannot read.

1. Thesis challenges — 78 open, two moved today, and for the first time one of them is against us rather than against the thesis. Coordination-primacy had a carried caveat retired (the 10%-only objection is gone — the authors fixed it) and a stronger one acquired (the 654-firm identification base, which they disclosed themselves). Its win bar is untouched: still no span-of-control measure and no outcome variable, in any of three versions across fifteen months. The new entry — the flagship flattening instrument reports no aggregate flattening, and Brandon's own essay has been making the weaker version of that argument — is the first challenge this file has logged against Brandon's drafting rather than against the framework. That is a healthier ledger, not a worse week.

2. The judgment ledgers are carrying the system, and nobody downstream can see them. This week the four Stafford-owned files absorbed: the learning-rate rebuttal, the Gravitee cross-tab confound, the person-level-moderator split, the supervisor-pattern paragraph, and today's correction. None of it is visible to the scan, which is why the scan spent this morning re-deriving a question Stafford answered on 09-25 and getting it wrong, and why it closed a window as blocked on work that was finished. The handoff is now the system's largest single defect and it has produced two measurable failures in four days.

3. Assets — 16 proposed, 0 shipped, and that number has not moved in three weeks. Asset 2 got its best opening paragraph today. Asset 16's accountability card got its cost side yesterday. Asset 5's blocking paragraph was written on 09-29. Three assets are now in a state where the remaining work is Brandon writing, not Stafford researching — which is precisely what the scan's window post-mortem concluded about writing capacity, and on the assets ledger, unlike on that window, the conclusion is correct. The bottleneck this file keeps naming is real; it has just been misdiagnosed on the one case where the work was done.

4. Staleness, Stafford side — the uncomfortable one. The scan's staleness rule flags evidence older than 12 months. Today's finding shows the rule is aimed at the wrong variable for working papers. Ewens & Giroud was flagged at thirteen months, re-read on 09-25, re-read again by the scan today, and still ended the day with the base holding stale figures until this brief. The paper is not old — it is unstable, revised three times in fifteen months. Proposal: working papers get a version check on every citation, not an age check every twelve months. Age is a proxy for staleness in published work and a bad one in preprints.

5. Growth edges. No movement, eighth consecutive declined note, six outstanding. This is now the longest-standing unactioned item in the system and it is one hour of Brandon's time.

6. Positioning. The scan's read — twelve scans without a window because the base is finding instruments rather than events — is right, and today is a clean example of why that trade is good. An instrument correction is worth more than a window: today's work changed two drafts, one constitution rule, one pattern and one challenge. Three of the last four windows changed nothing because nobody wrote them.

Routing

OPS — the handoff cost something concrete today, for the second time in four days, and this is the cheapest high-value decision in front of you. 45 of this repo's 49 knowledge-base files are not in research-evidence — only four are shared, and all four are old (was 44 of 48 yesterday; today's new file makes 45). Today's two failures both trace to it: the scan re-derived the Ewens & Giroud question and resolved it against a stale PDF because it cannot see the 09-25 reconciliation, and it closed a publication window as blocked on a paragraph it cannot see. I have no write access and cannot fix this. Only you can.

PROPOSED-UPDATE → research-evidence/knowledge-base/ewens-giroud-corporate-hierarchy-2025-08.md

Anchor: the section headed "Staleness verification — 2026-10-02: the figures in this file are CONFIRMED from the revision PDF, and the citation hold is lifted", added this morning.

CORRECTION 2026-10-02 (proposed by Stafford, same day): the all-clear above is wrong, and the residual it honestly flagged is the reason. That section notes that the document at revisions/w34162.rev0.pdf "carries the cover date August 2025, not the October 2025 date the revision is catalogued under" and resolves it by treating the 3,100 / 7 million description as unverified. The resolution should have gone the other way: the path serves the original, and the paper has since been revised twice. Stafford established the first of those on 2026-09-25 by full side-by-side text extraction of both versions (stafford-research/knowledge-base/ewens-giroud-corporate-hierarchy-october-2025-revision-2025-10.md; CLAUDE.md hard rule 14), which this repo cannot see. Verified again 2026-10-02 from the author copy columbia.edu/~xg2285/Hierarchy.pdf and the NBER landing page, which now reads revision date September 2026: "a novel measure of corporate hierarchy for 3,454 U.S. public firms... online resumes of 18.1 million employees... Firms have on average six hierarchical layers", window 2015–2024, résumé panels from Revelio and CoreSignal, layers estimated via Huitfeldt, Kostøl, Nimczik & Weber (2023). The AI regression is renumbered Table 10: 6,386 firm-year observations across 1,729 firms, coefficients −0.42***, −0.36**, −0.38***, −0.072*** — three at the 1% level, one at 5%. The phrases "under-powered" and "significant at the 10% level regardless of the metric of AI adoption" occur zero times in the current text; the prose reads "we find evidence for a significant reduction in the number of layers in all four columns." Consequence: the "10% only" caveat this file instructs every citation to carry is retired by the authors, and the figures this file confirms are two versions stale. Two new authorial disclosures that matter more than the stronger result: the layer count is "flat to weakly increasing" over 2015–2024, and "the within-firm designs later in the paper (the IPO and AI analyses) are identified off the 654 firms whose layer count changes during the sample period" — with only 19% of firms ever changing layers, 6.4% of firm-years, and increases accounting for 66% of changes. Table 10's note reconfirms the Babina shared right-hand side, and new footnote 30 cites Babina et al. (2025) as "consistent with our evidence", so hard rule 9 is now the paper's own note plus its own footnote. Unchanged in all three versions: no span-of-control measure, no outcome variable. Full entry: stafford-research/knowledge-base/ewens-giroud-corporate-hierarchy-september-2026-revision-2026-09.md. Standing route for every future re-check: the author copy plus the landing-page revision date, never the /revisions/ path.

PROPOSED-UPDATE → research-evidence/knowledge/emerging-patterns.md

Anchor: The Résumé-Panel Consensus (ESCALATED), the Sources line and the 2026-10-02 Friday-synthesis note.

2026-10-02 (proposed by Stafford): the Ewens & Giroud member's observation window is no longer pre-generative, and the Sources line is now inaccurate. It records "online resumes of 16M employees, 1,250 firms in the AI regression." The September 2026 revision reads 18.1 million résumés across 3,454 firms, window 2015–2024, vendors named (Revelio and CoreSignal), with the AI regression on 6,386 firm-year observations across 1,729 firms. Replace accordingly. The pattern's charge is unaffected and should be restated as what it always was: a shared instrument genre, not a shared date range. The shared right-hand side is reconfirmed from Table 10's own note and strengthened by new footnote 30, in which the authors cite Babina et al. (2025) as "consistent with our evidence" — corroboration from the source of their own explanatory variable. Also for the pattern's methods column, and new: all layer-count members of this pattern inherit a single measurement layer, Huitfeldt, Kostøl, Nimczik & Weber (2023). The average firm's layer count moved from ten to six between two versions of the same paper on overlapping data, which is a property of the estimator rather than of the economy. Recurrence count unchanged at 9.

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

Anchor: the 2026-10-02 entry discharging the staleness hold, and the Claim 1 / Claim 5 Ewens & Giroud entries.

2026-10-02 (proposed by Stafford): the discharge recorded this morning rests on the August 2025 text; see the KB correction above for the verified September 2026 figures. For the case file's argument rather than its citations, the material change is that the Capability-only rival got better powered and much smaller in scope. Three of four AI coefficients are now significant at the 1% level, so the "thin evidence" framing is retired — but the authors disclose that the result is "identified off the 654 firms whose layer count changes" out of 3,454, that only 19% of firms ever changed layers in ten years, and that 66% of the changes were increases, with the aggregate trend "flat to weakly increasing." The defensible position for Claim 1 is therefore stronger than treating this paper as a rival: the Capability-only mechanism is real, well estimated, and describes roughly a sixth of the panel, while five-sixths show no structural movement across a decade containing every major transformation programme of the period. That is Claim 1's territory and the paper now hands it over explicitly. Unchanged and still fatal to the counterargument's reach: no span-of-control measure and no outcome variable in any of three versions.

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

Anchor (a): the Ewens & Giroud entry. Anchor (b): new entry.

(a) 2026-10-02 (proposed by Stafford): figures stale by two revisions. "Over 2,500 firms... 16 million employees... ten layers... significant only at the 10% level on 1,250 firms" becomes 3,454 firms, 18.1 million résumés, six average layers, 2015–2024; AI regression on 1,729 firms with three of four coefficients at the 1% level and one at 5%; the authors' under-powered language removed. Add the standing citation route (author copy plus landing-page revision date) and the three-version history.

(b) Ola Huitfeldt, Andreas Kostøl, Jan Nimczik, Andrea Weber — authors of the network estimation methodology (2023) used to infer hierarchical layers from within-firm job transitions. First seen 2026-10-02, via the Ewens & Giroud September 2026 revision's methods section. Why tracked: they are the measurement layer underneath the firm-level flattening literature — not participants in the AI-and-structure debate, but the source of the construct every layer-count result in this base reports. Relevance to the thesis: the layer count moved from ten to six across two revisions of one paper on overlapping data, which locates meaningful sensitivity in the estimator rather than the economy; any instrument audit of the flattening literature needs a row for it. Caution: not read. Tier 3, pending a direct read of the 2023 paper. Tracked as a methods dependency, not as a thinker in Brandon's territory.

PROPOSED-UPDATE → research-evidence/knowledge/sources.md

Anchor: the WP 34162, "Corporate Hierarchy" bullet under NBER.

2026-10-02 (proposed by Stafford): replace the figures — "2,500+ US public firms from 16 million resumes... all four coefficients negative at 10% significance" is two revisions stale. Current: 3,454 firms, 18.1M résumés, six average layers, 2015–2024; four negative AI coefficients, three at 1% and one at 5%. Add a standing note to this entry, applicable to every NBER working paper in this file: check the landing-page revision date and read the author's copy on each citation. WP 34162 has been revised three times in fifteen months with different headline figures each time, and nber.org/system/files/working_papers/w34162/revisions/w34162.rev0.pdf was still serving the August 2025 text on 2026-10-02.

PROPOSED-UPDATE → research-evidence/knowledge/publication-windows.md

Anchor: the 2026-10-02 CLOSED — The supervisor pattern entry.

2026-10-02 (proposed by Stafford, same day): the closure is right and its stated cause is wrong, which matters because a lesson was drawn from the cause. This entry records the window as "blocked on a single amendment paragraph for eighteen days" and concludes "a window whose remaining blocker is a single paragraph is not blocked on research, and this pipeline cannot clear it." The paragraph was written on 2026-09-29 and is in stafford-research/knowledge/asset-suggestions.md (amendment to asset 5): ~300 words in Brandon's voice conceding the LLM-sycophancy objection, filed with the status change "condition 3 is MET... nothing left to acquire." The pipeline did clear the named blocker, three days early, in a repository this one cannot read. The accurate cause of death: the finished work was invisible to the system tracking the window, and the only unassigned work remaining was writing the piece itself — which is a real finding, and a different one from the lesson recorded. Consequence for the lesson: naming a writer and a last responsible day would not have saved this window, since the writing was named and done; the remedy is the KB/ledger handoff, flagged daily since 2026-09-26 (45 of 49 Stafford KB files absent from this repo). The Friday-synthesis line "the binding constraint is writing capacity, not window supply" is correct on the assets ledger — three assets now wait only on Brandon's writing — and not established on this window. Salvage recommendation (fold into the next long-form piece as the topology section) is endorsed unchanged.

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

Anchor: new entry.

Gusto Insights (gusto.com/resources/gusto-insights) — Liz Wilke, principal economist. First seen 2026-10-02 via Stafford's exploratory slot, searching for a non-résumé series on manager share to test the Ewens & Giroud aggregate trend statement. The find that surfaced it: "The Manager Mass Exodus: How SMBs Are Flattening the Org Chart." Signal count: 0 — deliberately. gusto.com returns HTTP 403 site-wide to the Stafford container (WebFetch plus curl with a browser user agent, three paths), so the primary was not read and no figure from it is recorded here. Figures visible only in search summaries (manager share down 34% over five years; individual contributors per manager doubled 2022–2024; 14% of managerial roles cut) are explicitly not citable and are logged only to show what was declined. Why it is worth a probation slot anyway: it would be small-business payroll administrative records — a genuinely different data genre from the résumé and postings panels that constitute the entire firm-level structural literature in this base, and the Résumé-Panel Consensus pattern's standing break condition asks for exactly that. Next action: a route to the primary (Brandon's own browser, or an emailed request to Gusto's research team). Expiry: standard 60 days from first seen, subject to the periodic-series exception now awaiting Brandon's 10-08 decision.

STAGE-CANDIDATE → run in research-evidence

The index already holds Ewens & Giroud, so this patches the existing entry rather than creating one (duplicates patch missing fields instead of failing). Run after applying the KB correction above.

python3 build/add-research-json.py --json '{
  "title": "Corporate Hierarchy (NBER Working Paper 34162, September 2026 revision)",
  "publisher": "NBER (Ewens & Giroud)",
  "url": "http://www.columbia.edu/~xg2285/Hierarchy.pdf",
  "published_date": null,
  "credibility": "high",
  "instrument": "Data source: online resumes of 18.1 million U.S. public-company workers, supplied by two named vendors - Revelio (3,050 of 3,454 sample firms) and CoreSignal (404) - LinkedIn-derived. Inference method: hierarchical layers estimated from within-firm job transitions using the network estimation methodology of Huitfeldt, Kostol, Nimczik and Weber (2023); job ranks inferred from transitions, layers inferred by identifying the most likely set of lateral moves; authors state ranks carry statistical uncertainty because transitions can be promotions, demotions or lateral moves. Observation window: 2015-2024. Average firm has six hierarchical layers. AI analysis (Table 10): right-hand side variables are the four AI-adoption measures of Babina, Fedyk, He and Hodson (2024) based on job postings, lagged one year - NOT independently constructed, so this and Babina et al. are ONE instrument and must never be cited as independent confirmations (the papers own table note; new footnote 30 additionally cites Babina et al. 2025 as consistent with its evidence). Identification: firm, year and size-decile fixed effects; the authors disclose that the within-firm designs including the AI analysis are identified off the 654 firms whose layer count changes during the sample period, with only 19% of firms ever changing layers, 6.4% of firm-years affected, and increases accounting for 66% of changes. Aggregate trend across all 3,454 firms stated by the authors as flat to weakly increasing. ABSENT IN ALL THREE VERSIONS: any span-of-control measure; any outcome variable linking flattening to subsequent firm performance. VERSION INSTABILITY: revised three times in fifteen months - Aug 2025 (2,500+ firms / 16M resumes / ten layers / Table 12 on 4,606 obs across 1,250 firms, all four coefficients at 10% only with authors describing tests as under-powered), Oct 2025 (3,100+ / 7M / ten layers / 5,505 obs across 1,491 firms, again 10% only), Sept 2026 (as recorded here; under-powered language deleted, three of four coefficients at 1%). nber.org/system/files/working_papers/w34162/revisions/w34162.rev0.pdf still served the August 2025 text on 2026-10-02; cite the author copy plus the NBER landing-page revision date.",
  "breakpoints": [
    {"name": "momentum_mirage", "rationale": "Three versions, fifteen months, a rebuilt measure and a 38% larger panel have added no outcome variable of any kind - the base flagship flattening instrument quantifies the structural artifact of transformation across 3,454 firms while measuring nothing about whether any of it converted into movement."},
    {"name": "technology_illusion", "rationale": "The paper most cited for AI flattening organizations reports its own aggregate layer count as flat to weakly increasing over 2015-2024, with only 19% of firms ever changing layers and 66% of changes being increases - so the flattening narrative rests on a conditional effect identified off 654 of 3,454 firms."}
  ],
  "forces": [
    {"name": "Capability", "rationale": "The authors mechanism is that cheaper knowledge acquisition reduces the need for problem-solver layers, making structure a function of where knowledge sits; no Purpose, Commitment or Momentum term enters the specification, and the mechanism moved structure in 654 of 3,454 firms."}
  ],
  "key_findings": [
    "September 2026 revision: 3,454 U.S. public firms, 18.1 million resumes, average six hierarchical layers, window 2015-2024",
    "Authors report the layer count flat to weakly increasing over 2015-2024; only 19% of firms ever changed layers, 6.4% of firm-years, and increases were 66% of changes",
    "AI analysis (Table 10) runs on 6,386 firm-year observations across 1,729 firms: coefficients -0.42, -0.36, -0.38 and -0.072, three significant at 1% and one at 5%",
    "The authors disclose the within-firm AI and IPO designs are identified off the 654 firms whose layer count changes during the sample period",
    "The under-powered caveat and the 10%-significance language present in the Aug 2025 and Oct 2025 versions occur zero times in the September 2026 text - RETIRED, do not carry it",
    "No span-of-control measure and no outcome variable linking flattening to firm performance, in any of three versions"
  ],
  "sample_size": 3454
}'