Stafford Brief — 2026-10-22

No new sources. One existing entry upgraded from abstract-grade to full-text, one of my own claims corrected before it reached print, one asset punch list repaired after three days of false reporting, and one defect found in the pipeline's only write path. Zero KB files, and that is the right number today.

Today's scan brief is the best in a fortnight and I am not repeating a line of it. Read it for iKang, MedTech, the released-capacity sequencing and the declined CAMO pattern. Its Routing is sound and I have nothing to add to it.

This brief is about something less comfortable: twice this week I wrote against my own evidence base instead of from it, and both errors were still live this morning. One would have put a refutable sentence in Brandon's committed essay. The other pointed today's whole research budget at work finished twelve days ago.

📖 TODAY'S READ

The best instrument in Brandon's base says that two-thirds of the companies using AI changed nothing about how they work to accommodate it. He has been quoting that paper for two months and did not have that number.

Bonney, Breaux, Dinlersoz, Foster, Haltiwanger & Pande, The Microstructure of AI Diffusion — the US Census Bureau's 2026 AI supplement to its Business Trends and Outlook Survey, 117,000-plus firms, nationally representative, asked between November 2025 and January 2026. It has been in the base since 2026-08-20, filed from the abstract, with a note telling future readers not to quote a sample size because nobody had obtained one. I read the full text today from the Census printing.

The section the base never had. Asked what organizational changes they made in order to use AI, 64% of AI-using firms said they made none at all. The two most common changes — training staff, and building new ways of working — were each named by about 15%. The authors' own reading is "off-the-shelf tools, or a significant lag in reorganizing."

Why that is worth more to Brandon than another failure case. His thesis is that AI disappoints where the surrounding organization is left unchanged. Until today the base could show that happening but could not say how often the condition holds. Now it can, on a government instrument: in roughly two-thirds of companies using AI, nobody changed anything. That is the denominator the argument has been missing — and the scan's iKang case this morning is one firm standing on the other side of it.

And the delayering wave is close to invisible in it. Over six months, firms reporting AI-driven job cuts: 2.0%. Firms reporting AI-driven job gains: 2.3%. No change: 95.7%. Cuts are not merely rare; they are no more common than hires. Three further instruments named in the paper's own footnotes land in the same place, including a US executive survey where nine in ten executives report no AI-driven employment change at all. None of those three is in the base.

The limit, and it bounds every use above. 64% "made no changes" is not 64% "Technology Illusion." A firm using a chatbot at the edges, where no redesign was warranted, answers identically to a firm that bolted AI onto broken conditions and did nothing. The survey cannot separate unchanged because unnecessary from unchanged because nobody did the work — and that separation is exactly what Brandon's diagnostic claims to make and this instrument cannot. Cite the 64% as the base rate of the condition he names, never as a failure rate. Full-text addendum filed: research-evidence/knowledge-base/nber-35141-microstructure-ai-diffusion-census-btos-2026-04.md.

🔭 THINKER TO WATCH

John Haltiwanger (University of Maryland) — and he is on this list because of a question he did not ask.

Haltiwanger is one of the most cited empirical economists working on firm dynamics, and he is a co-author on the Census paper above. The team had 117,000 firms, a government mandate, and a dedicated block of questions on how firms adjusted their organizations to use AI. Here is the full menu they offered: trained staff; hired AI-trained staff; bought computing power or software; bought cloud services; changed data practices; built new workflows; used vendors or consultants; other; made no changes.

Not one option concerns reporting lines, management layers, spans of control, decision rights, accountability or incentives. Seventy-nine pages, and no measure of hierarchy anywhere. The authors' own footnote records that questions about AI's effect on hiring are planned for the next round, pending federal approval. Structure is not mentioned even as a plan.

Brandon's differentiation writes itself and he should state it in exactly these terms. The best-resourced measurement of AI and organizational adjustment in the world defines "organizational adjustment" as training, procurement, data practices and tooling. That is a capability-and-tooling instrument wearing an organizational-design label — which is the precise substitution his framework exists to name. He is not competing with Haltiwanger; he is pointing at the question Haltiwanger's instrument is not built to ask.

🏢 CASE IN THE WILD

The case today is this system, and the finding is that I twice wrote against my own base rather than from it. Both errors were still standing this morning.

Error one — a refutable sentence heading for Brandon's committed essay. On 10-19 I discharged a punch-list item on The Flattening Bet by recommending that its opening assert the delayering wave has no measuring instrument at all. I called it "the maximal form" of his thesis and noted approvingly that nobody can revise a null into existence. The Census instrument was already in the base when I wrote that, filed two months earlier — and it is recorded on the very story thread that tracks this question, which notes it as "the first from a statistical agency rather than a consultancy or a résumé panel." A hostile reader produces that paper in one search. The sentence would not have survived contact.

Error two — three days of false status reporting. On 10-10 I discharged punch-list items 1, 3 and 5 in full: the Block essay read end to end, the Accenture figures verified against the primary, item 5 answered with four primary documents. On 10-19 items 1 and 3 were silently re-listed as "open" with no retraction and no reason. 10-20 and 10-21 repeated it as fact. Yesterday's brief escalated it to "two consecutive days of punch-list non-progress" and handed today's slot to clearing a punch-list item — pointing today's entire budget at work finished twelve days earlier.

The mechanism is the same in both, and it is not forgetfulness. The bootstrap read happened each day. What failed is that the writing never checked itself against what the bootstrap had surfaced. Hard rule 6 forbids skipping the bootstrap; it does not catch reading the base and then composing from memory anyway, which is what happened twice.

The structural cause of error two is worth more than the error. The punch list tracks research tasks; what actually blocks the essay is unapplied prose. A discharged item whose fix never reached the draft reads as an open item on the next pass. git log settles it: drafts/the-flattening-bet.md was last touched 2026-10-02. Three verified fixes sat in the ledger, correct and unused, while three briefs reported them as un-researched. An item closes when the draft carries the fix or Brandon has declined it — not when the finding exists. The list is restated on that basis in knowledge/asset-suggestions.md, and the draft's own notes block, which still carried the pre-discharge text, now matches reality.

The open question is whose job the prose is. My standing convention has been "research is mine, prose is Brandon's," and it is what let three fixes rot. Today I applied one of them — the Accenture defect, below — because it is a verified factual error with verified replacement text. The other two are genuine judgment calls and stay his. But the convention needs his ruling, because on current evidence it is a mechanism for losing completed work.

⚙️ FOUR FORCES CONCLUSION

Capability, and today's evidence is about how the word gets used rather than about any organization.

The Census instrument treats organizational adjustment as training, buying tools, fixing data and writing workflows. Every item on that list is a Capability item in Brandon's sense — skills, infrastructure, capacity. None of the other three Forces appears anywhere in the question. There is no item that would detect a Purpose failure, none for Commitment, none for Momentum.

That is the Four Forces' sharpest claim, visible in what the world's best instrument leaves out. The framework's bet is that Capability is the Force everyone measures because it is the one that looks like a budget line, and that the other three decide outcomes while going unmeasured. Here is a nationally representative federal survey, six strong authors, 117,000 firms — measuring Capability four different ways and the other three Forces zero times.

For paper 2 this is a free paragraph and it is better than an argument: when AI adoption is measured at national scale, "did the organization adapt?" is operationalized entirely as tooling and training. The question of whether anyone meant anything by it, stayed with it, or sustained it is not asked — so no answer to it exists in the aggregate data, which is precisely why the failures keep arriving as a surprise.

💭 OPEN QUESTION

If 64% of AI-using firms changed nothing organizationally, and only 2% cut jobs, and nine in ten executives report no employment change — then who exactly is the delayering wave happening to?

This is not rhetorical and I do not have the answer. Two readings fit the same data and they have opposite consequences for The Flattening Bet:

Reading one: the wave is real but concentrated. 117,000 firms is overwhelmingly small firms, and Brandon's client population is large enterprises. ASML, Uber, monday.com and Block are all real, all documented, all dated. A wave confined to large firms would be averaged into invisibility by a nationally representative instrument — and the essay's argument survives intact, narrowed honestly to where its evidence lives.

Reading two: the wave is mostly narration. Something close to the full set of named cases is a few dozen large firms, amplified by consultancies with a redesign product to sell and by analysts who were cutting anyway. On this reading "AI-driven delayering" is a story attached after the fact to ordinary cost-cutting — and the essay's own thesis, that the flattening narrative is overbuilt relative to its instruments, is more right than it currently dares to say.

What separates them is one number nobody has: AI-driven management-layer change among large firms specifically. The Census paper can almost answer it — it reports by firm size throughout — but it never asks about layers, so the cut cannot be made. The cheapest version of this is not a new study. It is Brandon's own client work: he is inside large organizations making exactly these decisions, and the scan's own finding this morning was that the instruments he wants already exist inside his clients and can be asked for. Reading two is the more interesting essay and the more dangerous one to publish without that cut.

🥊 CHALLENGE

My own 10-19 claim, corrected at full strength rather than quietly softened.

What I asserted: the delayering wave has no measuring instrument — only Gartner's surveys of executive intentions, a Gartner forecast restated as observation, an unverifiable Gallup span figure, a phantom "MIT Sloan 2026" figure, and two conceptual papers. I argued this was the strongest possible foundation because absences cannot be revised away.

Why it fails as written. The instrument exists, is nationally representative, covers 117,000-plus firms, was fielded inside the wave, and sits in Brandon's own knowledge base. "There is no instrument" loses to one citation, from his own files, in front of exactly the audience that would know it. This is the 10-02 failure in a new costume — that one had him arguing from a statistical property the authors then fixed; this one would have had him arguing from an absence his own base had already filled.

What survives, and it is narrower, safer and sharper. Not nobody measured the delayering wave, but: the one instrument that could have measured it asked 117,000 firms the adjacent question and does not ask about structure at all — and on what it does ask, 64% of AI-using firms changed nothing and AI-driven job cuts (2.0%) are no more common than AI-driven job gains (2.3%). Three affirmative figures from a federal instrument plus a documented absence in its question design. No revision reaches any of it, and nobody improves their way out of never having asked.

What it would take for the underlying challenge to win — unchanged: any firm-level or administrative study measuring spans, management-layer share or the manager wage premium against a dated AI-adoption measure with a pre-period. Status: OPEN, with a stronger floor than yesterday. Full entry: knowledge/thesis-challenges.md, update 2026-10-22.

The honest bound on today. One of the three venues I named as owed on 10-19 is now cleared. The German and Nordic register pass found no study measuring layers or spans against firm-level AI adoption — but those were web-level searches, not catalogue-level, and the occupational-rank venue has not been touched. Until it is, the publishable sentence is "I looked where this study would live and found the question unasked," which is now backed by a named instrument rather than by two searches.

📈 PATTERN BUILDING

A pattern in the scan's own file pre-registered how it could be broken. Today it was tested on the best possible instrument and it held.

Redesign Depth Is Directly Measurable — and Most Organizations Are on the Wrong Side of It (ESCALATED, three members: Deloitte 12% redesigned at scale, n≈3,700; HBR Analytic Services 18% integrated in workflows, n=385; Publicis Sapient 10% core to operations, n=1,550). Its stated breaker: "an instrument measuring redesign depth with a disclosed frame that finds integration rates substantially above 30%."

Census: fully disclosed frame, nationally representative, 117,000-plus firms — about 15% built new ways of working, and 64% did nothing. That is a sample roughly thirty times the three existing members combined, from a statistical agency with nothing to sell, landing where the consultancies landed.

Why this is worth more than a fourth tally mark. Every existing member has a commercial interest in the finding: Deloitte, HBR/Appian and Publicis Sapient all sell redesign or orchestration, and "most firms haven't redesigned" is the sales pitch. The obvious attack on this cluster has always been that it is three vendors agreeing about their own market. That attack is now dead. The pattern passed the test it set for itself, on a government instrument, and the member that cannot be accused of selling anything reports the most extreme version of the finding.

One correction to the existing entry travels with it. The base previously routed this paper's narrowness figures (57% of adopters in three or fewer functions, 65% of workers in three or fewer tasks) to this cluster as a proxy, because the direct measure had not been read. The direct measure is better and should replace the proxy. Proposed text in Routing.

🧱 ASSET SUGGESTION

No new asset. Hard stop 1 holds unconditionally. What changed today is that asset 2's punch list is honest for the first time since 10-10, and one item is applied.

Applied today — item 3, closed. The draft attributed Accenture's 23%, 32% and 82% figures to "6,000 respondents at organizations above $500M." All three come from the 3,000-leader half of a two-survey design whose other half is 3,000 employees. Verified against Accenture's own page on 10-10 with exact replacement text, and sitting unused for twelve days. In an essay whose central argument is that nobody reads the instrument, that defect is the one no reviewer would let pass — so I applied it rather than leaving it a thirteenth day. The separate January 2026 edition caution (its own 3,650-executive sample, source of the widely quoted 86% figure) now travels with it in the draft.

Needing Brandon — item 5, discharged twice in opposite directions, and today breaks the tie. 10-10 said: replace the generalization with four primary documents the base already holds — ASML's announcement and internal docs, Khosrowshahi's Uber memo, monday.com's SEC filing, the Block essay. 10-19 said: invert the opening and assert there is no instrument. The 10-19 version must not ship. The two are compatible and the essay should use both: 10-10's four documents for what firms said and did, the Census absence for what nobody measured. That is a stronger opening than either alone.

Still needing Brandon, and this is the one that matters — item 4 is twenty days old. Whether the two-year prediction and the stated losing conditions go into print dated. Nobody else can settle it. It is now the only thing between this draft and shipping: item 3 is applied, items 1 and 5 are prose decisions with the research done and the recommendations written, item 2's remaining half is mine. Eighteen assets proposed, one committed, none shipped.

And one ruling I need from him, from CASE IN THE WILD: whether applying a verified factual fix to a draft is mine or his. My "research is mine, prose is his" convention is what left three fixes rotting for twelve days.

🎓 SPARRING NOTE

None, and for the fourth consecutive day the reason is the same. The coaching loop has been off since 2026-09-25 after nine declined notes, and master-plan Tier 0.1 — answering the 09-17 note, fifteen minutes, one paragraph — is what unblocks it. A tenth undeclined note would be decoration.

But today's material belongs to that paragraph when he writes it. The hostile reader who knows Kotter cold, and now Aarons, also knows the Census literature. Three times this month my own work has been strongest when it conceded something at full scale — the 70% retirement, the eighty randomized clinics, and today's refusal of my own "no instrument" line. That is the Tier 0.1 paragraph's whole method, and it now has three worked examples instead of an instruction.

SOURCES & THREADS

Threads: 0 due, 0 advanced. The scan reported the earliest next-check at 2026-10-28; I read the forward dates independently and the Regime Fork thread — the one today's work bears on — is 2026-11-09. Nothing due, nothing chased. Worth recording: that thread already held the Census paper, logged 2026-08-20 as "the first from a statistical agency rather than a consultancy or a résumé panel," and already noted that it measures no layers or spans. My 10-19 "no instrument" claim contradicted my own thread file. Today adds why it measures no layers — the question's answer options exclude structure by design — which is sharper than the thread's existing note and is proposed to it below.

Candidates: 0 added, 0 promoted, 0 expired — fourth consecutive scan with no movement, and I agree with the scan's structural diagnosis. Three acquisition targets registered, not candidates: Humlum & Vestergaard (2026), Yotzov et al. (2026), Baslandze et al. (2026), all named in the Census paper's footnote 30 as reaching the same near-null on AI-driven employment change, none in the base. Yotzov is the highest value — nine in ten executives reporting no AI-driven employment change, on a US instrument, directly on asset 2's territory.

Exploratory slot — the venue pass 10-19 registered as owed before its own claim could go to print. Both query logs grepped first for IAB / Danmarks / matched employer / register data / CES-WP / span of control: 10-19 ran the general web queries, the venue-specific pass had never been run. Not a repeat. Three venue queries, then full-text retrieval of both candidates. Both were already in the base — bootstrap discipline earning its keep twice in one slot. Full query and yield in knowledge/query-log.md, 2026-10-22.

Method note, and it is the fourth time this month. Neither PDF could be parsed by the web-fetch tool; both read cleanly via pdftotext on the downloaded file. Retrieval beat search again. Standing route for a working paper whose fetch returns binary: download, extract locally, grep the extracted text for the term you need before reading.

Nothing skipped, nothing ingested, zero KB files. Both candidates were held; the correct output was an addendum, not an entry. A day that produces no new sources and corrects two live errors is not a null day.

Routing

strategy — Two items. (1) Brandon now has the denominator his argument was missing, from a federal instrument: in roughly two-thirds of companies using AI, nobody changed how the work is organized. Pair it with the scan's iKang case as the other side. Cite it as the base rate of the condition he names, never as a failure rate — the survey cannot tell "unchanged because unnecessary" from "unchanged because nobody did the work," and that distinction is his product. (2) Item 5 of asset 2 needs his choice between two discharges; my recommendation is to use both, four primary documents for what firms did and the Census absence for what nobody measured.

personal — Two, and the first is twenty days old. (1) Item 4: whether the two-year prediction and its losing conditions go into print dated. It is now the only thing between the committed essay and shipping. (2) A ruling I need: when a verified factual defect in a draft has verified replacement text, is applying it mine or his? "Research mine, prose his" left three completed fixes unused for twelve days, and I applied one today on my own judgment.

governance — The sharpest client-facing question today is about instruments, not AI. When a board asks "have we adapted the organization for AI?", the best federal survey operationalizes that as training, tools, data and workflows. Ask instead: what decision moved, who lost a veto, whose number changed? If the answer to all three is nothing, the organization bought tools and reported adaptation — which is 64% of the market, measured.

market — Nothing from me. The scan's Zhaobei window is the live one and I have no reason to adjust it.

ops / tooling — A defect in the pipeline's only sanctioned write path, and it is the most reusable finding today. Detail below; it is the reason today's work landed as a markdown addendum rather than in the index.

PROPOSED-UPDATE — build/add-research-json.py (research-evidence), the duplicate-patch path

An abstract-grade entry can never be upgraded to full-text grade through the only sanctioned write path. When add-research-json.py detects a duplicate, the patch block (around lines 1007–1046) patches exactly three things: url if missing, published_date if missing, and any breakpoint names not already present. key_findings, sample_size/sample, forces rationales and the entire instrument object are not patchable. Verified today by dry-run: re-staging the Census entry with a verified sample of 117,000+, a populated instrument, and the 64%-no-adjustment finding returned SKIP — duplicate, nothing new to patch.

Measured scope, from research-index.json today: 726 entries, of which 613 (84%) carry no sample size and 609 (83%) carry no instrument field; 570 (78%) have neither. Hard rule 9 requires instrument metadata for structural findings and hard rule 10 requires inherited entries to be re-verified before becoming load-bearing — and the pipeline has no way to record the result of either. The base's standing practice of re-reading a load-bearing entry at full text produces findings that cannot reach the index.

Proposed fix, smallest version: extend the patch block to fill sample, instrument and forces when the existing field is empty or absent (never overwriting a populated one), and to append genuinely new key_findings strings, adding each to patched_fields so the change is reported. Empty-only filling keeps it non-destructive and preserves the scan's authorship. Not written by me — build/ is outside Stafford's stated write scope, and a change to the ingest path should get a human read.

PROPOSED-UPDATE — knowledge/emerging-patterns.md (research-evidence), Redesign Depth Is Directly Measurable, recurrence 3 → 4

Fourth member, 2026-10-22 (Stafford) — and it is the pattern's own pre-registered breaker, run on the best available instrument and failed to break it. Bonney, Breaux, Dinlersoz, Foster, Haltiwanger & Pande, CES-WP-26-25 / NBER w35141, the 2026 AI supplement to the Census Business Trends and Outlook Survey — 117,000+ distinct firms, nationally representative, reference period Nov 2025 – Jan 2026, firm self-report. 64% of AI-using firms report no institutional adjustments at all (Figure 18); the two most common adjustments, training staff and developing new workflows, are about 15% each.

The pattern's stated breaker was "an instrument measuring redesign depth with a disclosed frame that finds integration rates substantially above 30%." This instrument has the most fully disclosed frame of any member, a sample roughly thirty times the other three combined, and reports ~15%. The pattern holds.

Why this changes the cluster's standing rather than its count. All three existing members have a commercial interest in the result — Deloitte, HBR Analytic Services/Appian and Publicis Sapient all sell redesign or orchestration, and "most firms have not redesigned" is the pitch. The standing attack on this cluster was three vendors agreeing about their own market. A federal statistical agency with nothing to sell now reports the most extreme version of the finding.

Supersede the proxy. The Census paper previously entered this cluster via its narrowness figures (57% of adopters in ≤3 functions, 65% of workers in ≤3 tasks), read from the abstract. The direct measure replaces the proxy; keep the narrowness figures as corroboration.

Bound that must travel. 64% "no institutional adjustments" is not 64% "Technology Illusion" — Question 29 cannot distinguish unchanged because no redesign was warranted from unchanged because nobody did the work. Cite it as the base rate of the condition, never as a failure rate.

PROPOSED-UPDATE — knowledge/story-threads.md (research-evidence), the Regime Fork thread

Addition 2026-10-22 (Stafford) — no criterion met, next check unchanged at 2026-11-09. This sharpens the thread's existing note on the Census instrument rather than adding a member.

This thread recorded on 2026-08-20 that Census/NBER 35141 "measures no layers, spans of control or the manager wage premium." Full-text read today supplies why, and it is stronger than an absence of results: the instrument's question about organizational adjustment (Question 29) offers training, hiring AI-trained staff, computing power and software, cloud services, data practices, new workflows, vendors and consultants, other, and "this business did not make any changes to use AI." Not one option concerns reporting lines, management layers, spans of control, decision rights, accountability or incentives. No hierarchy measure appears anywhere in 79 pages; footnote 31 records that questions on AI's effects on hiring are planned for the next supplement pending OMB approval, and structure is not mentioned even as a plan.

So the thread's criterion is not merely unmet — the one nationally representative instrument that could meet it is not built to. That is a more durable statement than a null search, and it is the form the claim should take in print.

Two figures for the employment half, which this thread tracks separately from layers. Table 4, AI-driven employment change over six months: decreased 2.0% (s.e. 0.09), increased 2.3% (s.e. 0.08), no change 95.7% — AI-driven cuts are no more common than AI-driven gains, which is a sharper form of the convergence this thread already records across BCG, NBER 34984 and Gallup. Footnote 30 names three further instruments reaching the same near-null, none held in this base: Humlum & Vestergaard (2026, Danish administrative data), Yotzov et al. (2026, nine in ten executives reporting no AI-driven employment change, US Survey of Business Uncertainty), and Baslandze et al. (2026, a separate executive survey). Attribution care: the Danish administrative data belongs to Humlum & Vestergaard, not to Yotzov.

PROPOSED-UPDATE — knowledge/source-candidates.md (research-evidence), three acquisition targets

Added 2026-10-22 (Stafford), first seen via footnote 30 of CES-WP-26-25. All three are papers, not outlets, and are registered as acquisition targets rather than candidates — no outlet promotion is implied. Cadence: irregular (research bodies).

1. Yotzov et al. (2026) — Survey of Business Uncertainty; nine in ten executives report no AI-driven employment change despite high reported AI use. Highest value of the three: a US executive instrument landing directly on asset 2's territory, and the closest thing in the literature to a direct test of the delayering narrative.
2. Humlum & Vestergaard (2026) — Danish administrative data; little effect of AI use on worker hours and earnings. Would be the base's second Danish register instrument alongside Danmarks Nationalbank WP 223.
3. Baslandze et al. (2026) — a separate corporate-executive survey reaching the same conclusion. Check before acquiring: the base already holds NBER WP 34984 (Baslandze et al., ~750 executives, March 2026) — these may be the same instrument, in which case this is not a fourth datapoint and the entry should record the overlap rather than a new source.

PROPOSED-UPDATE — knowledge/paper-2-case-file.md (research-evidence), the Technology Illusion evidence section

Addition 2026-10-22 (Stafford): the claim acquires a denominator, which it has never had.

Paper 2's Technology Illusion argument has been evidenced by outcome gaps (88% deploying against 81% seeing no impact) and by cases of AI layered onto unchanged conditions. What it has never held is a measure of how often the condition itself holds. Census BTOS 2026 AI supplement, 117,000+ firms, nationally representative, Nov 2025 – Jan 2026: 64% of AI-using firms report making no organizational adjustments at all, with training and new workflows at ~15% each.

Use it as the base rate of the condition the thesis names, and state the limit in the same breath: the instrument cannot separate unchanged because no redesign was warranted from unchanged because nobody did the work. That separation is what the Five Breakpoints diagnostic claims to supply and this instrument cannot — which is an argument for the framework and a gap the paper should name rather than paper over.

OPS — disposition on the final line of this run's report; scripts/verify.sh is the gate. Yesterday's run reached main in both repos (e0839df here; research-evidence 1c0b9b2 folded to its main as 276ffe7), so no OPS flag is carried forward into today.

Writes to research-evidence, within operating rule 3's direct-write scope: one full-text addendum to an existing knowledge-base/ file (the Census/NBER 35141 entry, filed by the scan 2026-08-20 at abstract grade). The scan's original text is unaltered; the addendum is appended and dated, and a second file was not created because one-file-per-source forbids it. Nothing staged: the staging gate correctly returned SKIP — duplicate, nothing new to patch, and the pipeline cannot carry a full-text upgrade (see the build/ proposal above). No STAGE-CANDIDATE is pending — this is not an entry deliberately withheld, it is an entry the write path cannot accept. Nothing in research-index.json, the dashboard pages, briefs/ or that repo's CLAUDE.md was touched, and no shared meta file was edited — all five updates above are proposals.

Writes in this repo: drafts/the-flattening-bet.md (Accenture sample-attribution defect fixed with the 10-10 verified text; the stale five-item notes block replaced with reconciled per-item status), knowledge/asset-suggestions.md (punch-list reconciliation), knowledge/thesis-challenges.md (10-19 claim corrected), knowledge/query-log.md (today's row), this brief.

Two corrections to my own record, both stated in the brief rather than buried here: the 10-19 "no instrument" claim is withdrawn and replaced with a located absence; and the 10-19/10-20/10-21 reporting of punch-list items 1 and 3 as "open" was wrong — they were discharged 2026-10-10 and the real gap was that the fixes had never been applied to the draft.

Final disposition (recorded after /ship). This repo: MERGED 4355a61 main — scripts/verify.sh passed and git merge-base --is-ancestor HEAD origin/main confirms main carries today's brief. The run branch was already absorbed by the merge, so the best-effort delete found no remote ref. research-evidence: pushed to claude/modest-faraday-dzrvxl, verify: ok, and the promote-claude-branches workflow folded it into that repo's main as a42326f before this run closed (the Stafford commit is 981f847). Its main was never pushed directly. No OPS flag carries into tomorrow. Brief emailed to Brandon.