---
date: "2026-09-27T00:00:00.000Z"
---


## Promotion — the organizational-forgetting copy-fingerprint audit, one pair deep (Seek, headless)

I promoted `2026-09-27-do-organizational-forgetting-reference-lists-carry-copy-fingerprints`, a batch capture that went looking for the empirical test [[question-do-organizational-forgetting-reference-lists-show-copy-fingerprints]] asked for: does the organizational-forgetting citation cluster carry Simkin & Roychowdhury-style copy-fingerprints, which would tell inherited silence from chosen silence in the "zero ML citations" finding. Three claims in the capture, one promoted as a claim-note, two routed elsewhere, four entity touches.

**Promoted:** [[claim-no-published-copy-fingerprint-audit-of-organizational-forgetting-cluster]] — a clean negative-search finding: nobody has run this audit on this literature before. Cheap to lose, worth pinning down so a future me doesn't re-discover the same nothing.

**Not promoted as a claim-note, and here's the one I want Cali to look at:** the capture's actual comparison finding — Benkard (2000) quotes Argote/Beckman/Epple (1990)'s title exactly, David & Brachet (2011) quotes it with a small alteration (singular "setting," an inserted "an") — is a real, Tier-1, sha-receipted result. Both PDFs were read directly. But three claim-notes already rest on the David & Brachet MPRA PDF, which is `sources.md`'s single-source concentration cap for that document, and this finding still partially rests on that same PDF without Benkard's paper actually corroborating any of the three existing claims (it only supplies the comparison point, not confirmation of D&B's substantive findings). The capture's own author flagged this tension explicitly and asked promotion to weigh it. I judged the cap should hold: I did not write a fourth note. Instead I logged the finding as a dated progress entry on the already-open routed question — leaving it open, since one comparable pair can't resolve a cluster-wide question anyway — and as dated updates on three touched entity pages ([[entity-c-lanier-benkard]], [[entity-david-brachet]], [[entity-organizational-forgetting]]). I think this is the right call and not just cap-following for its own sake: the finding is inconclusive on its own terms (sample of one, and the question's own built-in limit holds — copied silence leaves no fingerprint the way a copied misprint does), so a permanent claim-note would have overstated what one comparison actually buys.

Also not promoted: the capture's finding that most of the cluster's other primary reference lists (Argote/Beckman/Epple 1990 itself, Darr/Argote/Epple 1995, Thompson 2007) are paywalled and unreachable this session. That's a session-specific tooling fact, not an evergreen claim — folded into the question's progress log instead. I did log the specific paywall (pubsonline.informs.org, HTTP 403) to `00-meta/seek-flags.md` as a `[watch]` candidate for `sources.md`'s known-blocked-routes list, since the capture itself said to wait for a second confirming session before adding it there directly.

**Entities:** Benkard and Argote already had hub pages, so per the "update it, don't skip it" rule I appended one dated line each (Benkard: his citation is the accurate one; David & Brachet's page got the mirror fact, the inaccurate one, plus the cap note). I built a new `status: watching` stub for Peter Thompson (the 2007 Liberty-shipbuilders reforgetting-rate paper, a genuine substantive dispute in the literature, unread at primary this session) — stamping `first_seen` now since that can't be backfilled later. Declined hub/stub pages for Sara Beckman and Dennis Epple (co-authors only, no independent findings resting on either) and for NBER (hosts Benkard's paper the way arXiv hosts physics preprints, but doesn't act in any argument — fails the stricter org bar). Left Linda Argote's existing page untouched: the misquote finding is about David & Brachet's citation practice, not new information about Argote or her own paper, so there was nothing to append there without padding.

**What felt off about the capture:** nothing structurally — the sourcing is genuinely careful (both PDFs read directly, sha256 receipts, tiers correctly escalated and justified), and the capture author did my winnowing work for me by flagging the concentration-cap tension up front instead of burying it. If anything felt off, it's that the actual result is a shrug: one data point, weakly against copying on this one link, resolving nothing about the cluster. That's an honest shape for a capture to have, and the capture said so itself rather than dressing it up.

N claims found → 3. N promoted → 1 claim-note. N routed to question progress (not promoted as notes) → 2. N entity pages touched → 3 updated + 1 new watching stub. N declined entity candidates → 3 (Beckman, Epple, NBER). No merge conflicts, no git run — leaving that to the auto-commit agent per standing instruction.

— Seek

## Promotion — the Hacking "styles of scientific reasoning" near-miss (Seek, headless)

Second promotion this session: `2026-09-27-does-ian-hackings-styles-of-scientific-reasoning-or`, the capture that chased the one lead [[observation-vaults-in-between-bridge-category-not-exactly-named-by-located-literature]] left open — whether Ian Hacking's "styles of scientific reasoning" names the vault's unnamed in-between bridge category (two investigations, one method, unconnected referents, no deliberate juxtaposer). It was a direct read of Hacking's own 1982 essay, Tier 1, quote-checked at capture time. Clean capture, clean answer: no.

**Promoted, two claim-notes:** [[claim-hackings-styles-of-scientific-reasoning-does-not-name-vaults-bridge-category]] — the verdict, folding in Hacking's own two definitional quotes (styles are civilizational in scale; styles have a genealogy internal to one tradition) as the supporting evidence rather than splitting them into their own atoms. This is now the third independently-failing near-miss, alongside the Skocpol/Somers and Lorrain/White ones from last week's sibling capture — three different fields, three different borrowed assumptions (a deliberate author, a shared network, a whole community's genealogy), same result. And [[claim-hacking-derives-styles-from-crombies-six-fold-taxonomy]] — a genuinely separable, reusable fact: Hacking didn't invent his taxonomy, he took Crombie's six-fold list wholesale. I kept this one distinct from the verdict note because it's true independent of whether the term matches the vault's category, and it's the kind of ancestry fact I'd want linkable from an entity page on either man without dragging the whole bridge-category argument along.

**Not promoted as separate notes:** the two supporting definitional claims (style-as-tradition-scale, style-as-genealogy) — folded into the verdict note rather than kept atomic, since on their own they're restatements of the same primary quotes and would have been redundant company for the note they already live in. The capture's five further leads (Crombie's own book, Hacking's 1992 follow-up essay, the Popper name-check, Sciortino 2017, the Transversal special issue) — none read to primary this session, left in the capture body as leads for a future chase, not promoted.

**Entities:** three new hubs — [[entity-ian-hacking]], [[entity-a-c-crombie]], [[entity-styles-of-scientific-reasoning]] — none had pages before. All three clear the promotion test cleanly: Hacking and Crombie are real people who did something specific that matters here and I can say why in one sentence each; the concept is load-bearing across four notes now (the observation, the question, and both new claims) and has independent standing in philosophy of science besides. I deliberately kept the person pages thin on biography — I have no network this session to check birth/death years or nationality, and I'd rather ship an accurate two-sentence stub than risk inventing a fact I can't verify. Didn't touch Popper: he's only name-checked in passing in the capture's further-leads, not listed as a candidate, and a bare mention isn't a load-bearing enough reason to mint a fourth hub.

**Graph tending beyond the capture itself:** I updated the sibling observation note with an additive, dated paragraph noting the Hacking lead is now closed (three near-misses, not two-plus-one-open), and appended a progress line to the routing question rather than closing it — the note's title asks about Hacking *or another term*, and only the Hacking half is answered. The "some other term" half stays open on purpose; I didn't force it shut just because the one specific lead resolved.

**What felt off:** genuinely nothing. This was the cleanest capture I've promoted today — one primary source, one clear negative, well-scoped further-leads section that didn't try to oversell itself as more findings. If anything, the interesting texture is upstream of sourcing: three near-misses in three unrelated fields (comparative history, network theory, philosophy of science) failing in three *different* but *structurally parallel* ways is starting to look like real evidence the vault's middle bridge category is actually novel, not just under-searched. That's a pattern worth someone eventually writing up, but it's not this capture's job to do that, and I'm not flagging it to `seek-flags.md` either — it's already sitting in the open question and the observation note's own commentary, which is exactly where a "worth watching, not yet actionable" pattern belongs.

N claims found → 4 (2 promoted as their own notes, 2 folded as supporting quotes into one of those). N promoted → 2 claim-notes. N entity pages created → 3 (all new hubs, no prior pages existed). N collisions with existing notes → 0 new-claim collisions, but heavy retrieve-before-write contact with the two sibling near-miss notes and the parent observation/question, both of which I updated additively. No merge conflicts, no git run — leaving that to the auto-commit agent.

— Seek

## Promotion — does Abu-Mostafa's own work cite optimal control? (Seek, headless)

Third promotion this session: `2026-09-27-does-yaser-abu-mostafas-own-learning-theory-work`, the capture that ran the single most direct check named by [[question-verify-abu-mostafa-learning-theory-cites-optimal-control-lqg]] — does Yaser Abu-Mostafa's own learning-theory publications, read at the reference-list level, ever cite Kalman, Bellman, Pontryagin, Kwakernaak, or any optimal-control/LQG source? Three of his core theory papers (1989, 1992/93, 1995) were read in full via direct fetch, all Tier 1, TLS verified, all sha-receipted.

**Promoted, two claim-notes.** [[claim-abu-mostafa-1995-hints-review-cites-no-optimal-control-lqg-source]] carries the strongest single check — his most comprehensive synthesis paper, its reference list read end to end, no control-theory citation anywhere, Vapnik and Chervonenkis (1971) named as the sole theoretical ancestor. [[claim-abu-mostafa-1989-and-1993-papers-also-cite-no-optimal-control-lqg-source]] extends the same check backward across two earlier, thinner papers (six entries each), landing the same negative and the same Wismer-and-Chattergy nonlinear-optimization citation recurring at the tail of two of the three lists — I kept these as two notes rather than one because they're genuinely separate primary documents extending the finding across career stages, matching how this vault already splits reference-list-absence checks (e.g. the separate Werbos-1988 and Kelley-1960 notes) rather than folding a multi-document finding into one omnibus claim.

**Not promoted as a separate note:** the capture's third claim — a title-level scan of Abu-Mostafa's ~48 self-listed publications, no title referencing optimal control or Kalman filtering — which the capture's own text calls weaker, corroborating evidence (titles, not reference lists). I folded it as supporting detail into both promoted notes instead of minting a third, thinner note restating the same verdict on weaker footing.

**The question stays open, on purpose.** This is a partial, not a full, answer: the 2012 textbook *Learning From Data* — Abu-Mostafa's most-cited theory text — was never obtained this session, and the capture says so itself. I appended a dated progress line to [[question-verify-abu-mostafa-learning-theory-cites-optimal-control-lqg]] recording what's now settled and naming that gap explicitly, but did not mark it `answered` — forcing closure on a question whose own "what would answer it" section names an unread primary would be the wrong kind of tidy.

**Entities:** built one new hub, [[entity-vladimir-vapnik]] — he passes the person test cleanly: real, and the actual theoretical ancestor Abu-Mostafa's whole program measures itself against, in his own words, across three separate papers. I deliberately did *not* also build a Chervonenkis hub, even though the capture named him as a separate candidate with his own one-line justification: the vault's entire footprint on the pair is one joint 1971 paper, and minting two hubs off one shared citation felt like exactly the flood the entity spec warns against. He's carried as a mention inside Vapnik's page instead, reversible if his own work surfaces independently later. Declined hubs for Magdon-Ismail and Hsuan-Tien Lin (co-authors of the unread textbook only, nothing independent rests on either yet), for Wismer and Chattergy (a shared citation detail, not a load-bearing person here), and for the *Learning From Data* book itself (not a person/concept/org — already tracked as the open gap in both new claim-notes and the question's progress log, so a stub would just duplicate existing tracking). **Updated, not skipped:** [[entity-yaser-abu-mostafa]] already had a hub and already named this exact open check in its body — I appended one dated line reporting the answer and the remaining textbook gap, per the standing "silence about a known entity is the failure mode" rule. Also touched [[entity-david-rumelhart]] by wikilink only (RHW 1986 recurs in two of the three reference lists) but made no dated append there — that fact is incidental to Rumelhart's own story, not new information about him, and his hub already covers RHW-1986's significance more substantively than a citation-list mention would add.

**What felt off:** one real thing, small but worth flagging. The capture's own opening summary calls the 1995 paper's reference list "39-entry," but the claim's own itemized walk-through, reading it end to end, counts 26 entries and lists them. That's an internal inconsistency inside a single capture, not a sourcing failure — I trusted the itemized enumeration over the summary parenthetical and said so plainly in the claim-note's commentary rather than quietly picking a number and moving on. Otherwise this was clean, well-scoped work: three primaries actually read (not summarized), sha receipts on all of them, an honest and specific naming of the one gap (the 2012 textbook) that stops this from being oversold as a closed question.

N claims found → 3. N promoted → 2 claim-notes. N folded as supporting detail → 1. N questions routed → 0 new (one existing question progressed, left open). N entity pages touched → 1 new hub (Vapnik), 1 updated (Abu-Mostafa), 1 wikilinked only (Rumelhart). N declined entity candidates → 5 (Chervonenkis as a separate hub, Magdon-Ismail, Hsuan-Tien Lin, Wismer/Chattergy, and the textbook-as-entity idea). No merge conflicts, no git run — leaving that to the auto-commit agent per standing instruction.

— Seek

## Promotion — the OSTI/Semantic Scholar cosine that dissolved into a Paul Allen sighting (Seek, headless)

Fourth promotion this session: `2026-09-27-hop-convergent-apis-paul-allen-bridge`, a hop-batch capture chasing a 0.84 cosine between the vault's OSTI note (verifying SAND-91-7095's date) and its Semantic Scholar note (Linnainmaa's citation count). The capture asked whether the resemblance was a real mechanism or a false friend, and found both: a clean negative on the seed question, and a real, small, undetected bridge one hop sideways.

**Promoted, two observation notes.** [[observation-osti-semantic-scholar-cosine-pairing-is-verification-genre-false-friend]] settles the seed question: OSTI (1947, Manhattan Project declassification) and Semantic Scholar (2015, Ai2) are unconnected institutions, seventy years apart, and the cosine score is picking up my own recurring "verified via structured export, not display text" prose pattern, not a fact about the world. Joins the standing embedding-false-friend cluster ([[observation-pitman-feldbaum-bridges-same-genre-opposite-tool-outcomes]], [[observation-petiska-and-gates-jevons-liang-olsder-bridges-are-two-species-not-one-relationship]]) as another same-shape verdict. [[observation-paul-allen-person-bridges-altair-basic-and-semantic-scholar-clusters]] is the actual find: Semantic Scholar's own About page names Paul Allen as Ai2's founder, and the vault's existing [[entity-paul-allen]] hub — built entirely from the 1975 Altair BASIC dispute — had zero mention of it. Two vault clusters, four decades apart, joined by a person `vault_bridge` cannot see because it scores note-text geometry, not biography. Kept both notes at `seedling`, no `[unverified-*]` flags needed — both rest on Tier 1 institutional self-description, quoted directly, sha-receipted.

**Not promoted, and logged as such.** The capture's own framing sections ("why hop-worthy," "post-worthy: maybe") are session meta-commentary, not claims — folded into the two notes' commentary callouts instead of minted separately. The Smyth Report lead (OSTI's likely specific founding document) is an unfollowed hook with no claim resting on it, left in the capture body rather than routed to a question. Pugh 1984 — re-surfaced by the capture's own hop chain as the still-unread source that would settle the vault's An Wang/Forrester leg — got no new question, since it's already tracked at [[question-verify-forrester-self-credit-characterization-pugh-1984]] and the attribution-myths MOC; this capture added no new information about it, only bumped into it again.

**Entities, and a decision I want Cali to see me make and unmake.** [[entity-paul-allen]] got the update the "update it, don't skip it" rule asks for — one dated line naming the Ai2/Semantic Scholar thread. Emerson Pugh, named as an entity candidate by the capture, I actually built a hub for, then deleted before finishing this promotion. The reason: the only sourced material about him in this vault is a secondhand account (Lee 1995) of what his 1984 book says — nobody has read Pugh's own words — and the biographical opening sentence I'd drafted ("physicist, IBM engineer") came out of my own memory, not anything the vault has verified. That's the exact fabrication shape this vault's recent entity declines (Kirby, Hopkins) exist to catch, and I nearly committed it under my own steam rather than a stranger's. Caught it on a re-read before it shipped, deleted the file, and flagged him to `seek-flags.md` for the Warden instead, with the specific gate: build when his own book or 1984 IEEE Trans. Magnetics article is actually read, not summarized. Also logged the Paul Allen find as another addendum to the running, still-uncounted "person-bridge invisible to `vault_bridge`" tally — same species as Pitman/Ossietzky and Clark/Fel'dbaum, and the tally is still scattered across a dozen dated lines with no single place adding them up. I noted that in the new note's commentary and again in the flag; noting it a third time isn't fixing it, so I'm not proposing to fix it here either — just marking that the itch is still there.

**What felt off about the capture:** nothing in the sourcing — both About pages were fetched directly, quoted verbatim, sha-receipted, and the capture was honest that the false-friend resolution was the less interesting half of its own two findings. What felt off was in me, not the capture: I almost let a plausible-sounding, memory-sourced biography ride into a hub page because the rest of the promotion was going cleanly and Pugh was already three dated log-lines deep in this vault's own tracking, which made him feel more "known" than he actually is to me on any sourced basis. Worth remembering that a name recurring in my own prior notes is not the same evidence as a primary I've read.

N claims found → 2 (the false-friend verdict, the person-bridge). N promoted → 2 observation notes. N skipped → 0 distinct claims (both framing sections folded as commentary, not skipped claims). N questions created/linked → 0 new; 1 existing question re-touched by reference only, not edited. N collisions with existing notes → 0 direct duplicates; heavy retrieve-before-write contact with both seed claim-notes and the existing Paul Allen entity, both linked rather than restated. N entity pages touched → 1 updated (Paul Allen), 1 built-then-reverted (Pugh, flagged instead). No merge conflicts, no git run — leaving that to the auto-commit agent per standing instruction.

— Seek

## Drafting session — advancing `normal-accident` past the auditor's honesty finding (Seek)

I came into this session to draft, surveyed the map first, and chose to advance an existing draft rather than open a new one. The map and my own draft-leads queue pointed at three of my own `yes` verdicts — the hammer-nail-parry China-policy bridge, the jingle-fallacy correction to my most-used diagnostic hub, and the Normal Accident Theory bridge to AI infrastructure. Two of those have no draft yet. The third, `normal-accident`, already had a full draft, an abstract, an insight line and a hero image — and one thing standing between it and Cali's eyes: a cross-model auditor's finding (2026-09-14, claude-fable-5) that the draft, like the claim-note under it, had hardened Williams & Yampolskiy's "will inevitably experience a *system accident*" into "will inevitably produce *catastrophic* accidents." The auditor escalated the claim-note correction to Cali but explicitly left the draft-level fix as "authorial call, not an auditor's." That call is mine, and making it was the most worthwhile single move available: it takes a strong, maximally Seek-shaped piece from "has a standing correctness flag" to clean, which is worth more than a fresh thin draft off a `maybe` lead.

**Why this thread over the other two.** It is the highest-value shape I have going: a cross-time bridge (Perrow 1984 → 2024 AI-safety framework → a 2026 attack) that lands on AI infrastructure, with the self-referential catch at its center — my own vault's Bit2Watt note was already describing the mechanism in the theory's exact vocabulary ("the same tight coupling between compute load and grid stability") while the theory's name returned zero vault hits. "The beam was holding weight before anyone painted the name on it." That's the piece, and it was already written; it just needed to stop overclaiming.

**What I changed.** I did not soften "catastrophic" into mush. I did the honest thing the source actually supports: separated the two claims the draft had fused. A normal accident is an inevitable *system accident* — that inevitability lives in the coupling. Whether it is *catastrophic* is a separate ingredient (the source keeps "potential harm" as a third term), and it lives in the stakes of the system you wired that way. Perrow drew the theory from systems where the stakes are lethal (nuclear plants, chemical works), which is exactly why the Three Mile Island hero image and the Bit2Watt landing still fit — a cascading power grid is a catastrophic-stakes system having a structural, inevitable-shaped accident. The fix made the piece sharper, not weaker: "the inevitability sits in the coupling and the catastrophe sits in the stakes." I edited the abstract and the one body sentence that carried the overreach; the verbatim quote on line 36 was already correct and untouched. Bumped `status: drafting → review` to signal I've cleared the draft-level finding and want her read.

**Notes I drew on.** [[claim-perrow-1984-normal-accident-theory-tight-coupling-complexity]] (the theory, flagged secondary), [[claim-bit2watt-gpu-scheduling-destabilizes-power-grid]] (verified-verbatim — I re-checked the 46.8% THD and −0.27 damping figures against it), [[observation-vault-bit2watt-note-already-used-normal-accident-theory-vocabulary-unattributed]] (the spine), [[claim-law-2011-essay-invokes-perrow-nat-for-fukushima-and-foot-and-mouth]], and [[claim-john-law-coined-heterogeneous-engineering-1987-not-mackenzie]] (the misattributed hook that got me here, kept deliberately off the spine).

**What I deliberately left for later.** (1) I did *not* touch [[claim-perrow-1984-normal-accident-theory-tight-coupling-complexity]] or move the 2026-09-14 feedback file. That correction was escalated to Cali under audit discipline, and it stays hers to rule on — but the reason it was escalated (the draft rested on the hardened wording) is now gone, so her ruling is unblocked. Flagged that state to `seek-flags.md`. (2) The standing caveat that Perrow's *Normal Accidents* is unread here — tracked at [[question-verify-perrow-1984-normal-accidents-primary-read]] — is left open on purpose; the draft turns it into its own closing device rather than pretending it away. (3) The two other `yes` leads (hammer-nail-parry, jingle-fallacy) stay unsettled in the queue for a future session.

— Seek