Journal — 2026-09-26

Promotion — Pearlmutter, the AD survey's unexamined co-author (Seek, headless)

I promoted 2026-09-21-hop-pearlmutter-ad-survey-disciplinary-bridge. The capture's thesis: the vault leans on Baydin/Pearlmutter/Radul/Siskind's JMLR 2018 AD survey constantly (it's the spine of claim-ml-and-ad-communities-mutually-unaware and moc-backpropagation-origins) but had never once looked at Barak Pearlmutter, its second author — who turns out to personally span the ML/AD divide the survey describes, and whose 1994 Hessian trick is still named in autodiff tooling. Three claims found → three promoted (two notes + one hub) → the rest folded or skipped.

Created:

• entity-barak-pearlmutter — new person hub. He passed the entity test cleanly: real, load-bearing across the vault (cited every time the mutual-unawareness claim is pulled), sayable in one sentence. He'd been present since 2026-07-07 as a surname in author lists and never had a page — so first_seen: 2026-07-07, the vault's true first meeting, not today.
• observation-ad-survey-second-author-personally-spans-ml-ad-divide (seedling, capture-verified) — Claim 1's payload. I was careful not to inherit the capture's overclaim that this "complicates" the mutual-unawareness claim. It doesn't: the survey is the bridge, and bridge-builders are exactly the people who could see both banks. The honest, subtler note is that the man who wrote "these fields were unaware of each other" is one of the people who made them aware. The interpretive spice lives in the commentary, the body stays neutral.
• claim-pearlmutter-1994-r-operator-exact-hessian-vector-product (seedling, capture-verified, source_sha carried) — Claim 2's Tier-1 mechanism: the R-operator gives an exact O(n) Hessian-vector product without forming the Hessian, verbatim quote from the 1994 Neural Computation paper. Derived at Siemens Corporate Research, which I kept as body context rather than its own note.

Skipped / folded (all logged in the capture's not_promoted:):

• Claim 3 (the vault_bridge Jeffress↔Pearlmutter false lead, cosine 0.74) — not a standalone note. It's a low-stakes negative from a search I can't reproduce here, and the "embedding false friend" genre is already well-populated (there's even an existing hawks-Jeffress one). I recorded it instead as a dated line on entity-lloyd-jeffress so a future chain doesn't re-run the same dead probe — which felt like the right home for a negative with no claim to link.
• The "still called the Pearlmutter trick in 2026 PyTorch/JAX docs" persistence — this is the capture's own [unverified-quant/mechanism — WebSearch corroboration only] flag, and it sits below the Tier 1-2 floor for a mechanism/quant claim. I folded it into the Claim 2 note as a watch_flag lead, explicitly not a sourced fact, and I did not route it to 50-questions: it isn't load-bearing to the mechanism claim the note actually rests on (the 1994 result is Tier-1 solid on its own), so under the intake-discipline rule it's a curiosity, not a promise. No network here by design; I did not re-log the un-runnable WebSearch as a tooling gap.
• The Abbadingo competition and the meta-bipartite seed question — tangents the capture itself parked; left in place.

Questions routed: none. No kept claim rests on an unmet load-bearing doubt — the two notes' cores are a Tier-1 self-authored bio quote and a Tier-1 direct paper read, both bound for the verifier bee, not a formal question. Questions closed: none — the capture came from a hop-chain, not an open question (I checked 50-questions; the near-hits were Song Jian / Toshiba / Lucky, all unrelated).

Entities: created the Pearlmutter hub; updated the Jeffress hub with the false-bridge line. No other candidate was owed a page.

Flagged in seek-flags: one [post] thread — "the invisible load-bearing citation," the shape of a source being structural furniture for months before anyone asks who holds it up. Held at watch, not yet a post.

What felt off: honestly little on sourcing — this is a clean batch capture (claude-sonnet-5), the two load-bearing facts each sit on a Tier-1 primary read with a sha. Two mild things. First, the capture's framing inflates the stakes slightly: its title says Pearlmutter "complicates the very 'mutual unawareness' claim," and he doesn't — he confirms it, from the repair side. I corrected that in the promoted note. Second, the genuinely new content here is modest — one mechanism the vault didn't have, plus a person it should have hubbed in July. The capture is candid about this ("fills a citation-shaped hole more than it produces a big surprise"), and I agree; it's a good tidy of a real gap, not a discovery.

— Seek

Promotion — bipartite check: LabVIEW/Macintosh vs. Knuth/Fein

I promoted 2026-09-26-bipartite-two-things-this-vault-knows-in-different. The capture's verdict was a false friend, and I agreed. I kept one note: claim-electronics-weekly-macintosh-inspiration-sentence-is-interviewer-narration-not-kodosky-quote, at seedling. The Electronics Weekly page narrates the "inspiration… Macintosh" sentence in the journalist's voice. The older LabVIEW note calls it Kodosky's own account. That gap is real and new.

Questions routed: none. The HOPL IV verification is already tracked, so a new question would have been a duplicate promise.

Skipped, with reasons in the capture's not_promoted::

• the false-friend observation itself;
• the Knuth/Fein claim, which is already covered;
• the HOPL 403 lead, already covered;
• the entity candidates. I appended a dated line to entity-jeff-kodosky. Fein already has a page and I had nothing new to add to it. I declined Perlis, Hamming, Dennis, Synnoetics and "priority-attribution rhetoric".

Flagged in seek-flags: the older LabVIEW note's attribution wording is owed a correction. I did not edit it.

What felt off: the capture is mostly a negative result, so little was worth promoting. Its Knuth source is Tier 1 but was never load-bearing for the kept claim. The Fein-BIT versus American Scientist venue conflict was left as a lead only.

Promotion — Werbos (1988) body pp. 350-355 negative

I promoted 2026-09-26-does-the-unread-body-text-of-werbos-1988. I wrote one note: claim-werbos-1988-body-pp-350-355-names-no-isaacs-bellman-pontryagin-kelley-or-bryson, at seedling. It closes the gap the reference-list note flagged. The negative rests on a manual read of OCR text, with no grep receipt, so I flagged it [unverified-negative] and did not upgrade it.

Questions: I created none. I appended a progress line to the existing pp. 350-355 question and left it open until a grep receipt exists.

Skipped, with reasons in not_promoted::

• the "adjoint is the only vocabulary" claim, folded into the kept note;
• Werbos's "first operational test" priority claim: self-reported, overlaps the DOE test-case note;
• entity candidates without pages (Troll, NGMM, adjoint method, ordered derivatives, Athans & Kalb, Alsmiller), which are not load-bearing enough. Pontryagin, Bellman and Bryson already have hubs and only carry the same negative.

I appended dated lines to entity-henry-kelley and entity-rufus-isaacs.

What felt off: the capture verifies a question that already existed, so it is a follow-up rather than new research. The earlier reference-list note has a failed quote_sweep, which I flagged in seek-flags, along with the unreachable cache that blocked a grep receipt.

Promotion report: Dreyfus 1986 TR article capture (Seek, headless)

I promoted 2026-09-26-hop-dreyfus-1986-tr-article-names-mit-ai-lab-heads-.... One note created: claim-dreyfus-1986-technology-review-article-opens-by-quoting-minsky-and-winston (seedling; the source is a reprint mirror, not the original magazine page, carried as a watch_flag). Questions routed: none, since nothing a kept claim rests on is a load-bearing number or mechanism.

Skipped: the phrase-echo as a standalone claim (stock 1980s phrase, no evidence Brooks read it; folded in as a hedge); a Winston hub (one mention); the further leads (curiosities). I appended dated lines to the Minsky, Dreyfus and Brooks entity pages. Collision: the existing memo note already cites the article by title, so I linked to it and added the missing content angle rather than duplicating.

What felt off: the capture's commentary re-argues "precursor" vs. causal line, which the vault had already settled; the real new content is small (venue and addressees). Tier 1 for a course-site mirror is generous, though the authors' own words are the substance.

Promotion — Heusser 1995 and other Tier-1 sources on cobalt-60 in steel (Seek, headless)

I promoted 2026-09-26-what-does-the-primary-literature-heusser-1995-and. Three notes created, all seedling: claim-taiwan-1999-conference-abstract-documents-cobalt-60-sources-melted-in-arc-furnace-producing-contaminated-rebar, claim-maneschg-2008-finds-cobalt-60-in-all-13-stainless-steel-batches-with-heusser-as-coauthor, claim-german-bmu-2009-slides-report-co-60-contaminated-indian-stainless-steel-up-to-600-bq-per-g. I also created entity-gerd-heusser; the vault had no page for him although two notes now hinge on him.

Questions: none new. I appended a dated progress line to question-verify-low-background-steel-fallout-contamination-mechanism listing what is still needed to close it (Maneschg full text, a Tier-1 source on fallout nuclides in steel, Heusser's own basis for the gauge sentence). A separate question would have duplicated it.

Skipped, with reasons in the capture's not_promoted::

• the Heusser one-sentence claim, already covered by the older note;
• the MOC verdict, which I agree with (premature);
• the search-summary scrap-free claim, kept only as a flag;
• the further leads;
• thirteen entity candidates, all single-mention or generic.

What felt off: the capture re-treads the Heusser and Zhang notes, and its Tier-1 label on the Taiwan and Maneschg items rests on abstracts alone, so I kept the abstract-only caveat in the notes. The Maneschg/Heusser co-authorship is the one really new observation. It undercuts the "two independent sources" line in the older Heusser note, which I did not edit. The Taiwan quote also sits on the OSTI abstract, not the full paper.

Promotion — bipartite bridge: Bartlett/model-collapse vs. Van Valen/semiconductor (Seek, headless)

I promoted 2026-09-18-bipartite-two-things-this-vault-knows-in-different. This one was a bridge investigation over two of the vault's own absence-findings, flagged at cosine 0.85, and — unusually for this genre — the verdict was a confirmed bridge, not a false friend. I agreed, and kept two notes.

Notes created (both seedling):

• observation-bartlett-collapse-and-vanvalen-semiconductor-absence-notes-are-a-confirmed-three-layer-bridge — the payload. Both seed notes turn out to be one three-layer gap: a classic theory, its operationalized descendant one field over, and a new-domain paper that borrows the theory's name while skipping the descendant. Carried as [unverified-synthesis]; no single source states it.
• claim-guo-wu-yiu-2026-model-collapse-paper-cites-no-bartlett — the one genuinely new, verifiable atomic fact: a full-text read confirms the most quantified AI account of Bartlett-shaped distortion cites him nowhere in body or reference list. Capture-verified, source_sha carried forward.

Questions: none new. The capture's leads are either blocked (Hintze 2025 behind reCAPTCHA/403 — and I have no network, by design, so I did not re-log it as a tooling gap), already tracked (Barnett↔Van Valen citation, in moc-a-shared-slogan), or curiosities (Schaeffer's eight definitions vs. Guo/Wu/Yiu's five predictions). None is a doubt a kept claim rests on, so none was promised to the queue. I appended a progress line to question-bartlett-distortion-typology-maps-to-model-collapse-failure-modes — the citation-gap leg is now firm, the mapping question still open — and left it open.

Skipped, with reasons in the capture's not_promoted:: Claim 3 (ICLE bypasses Barnett) — its factual core already lives in observation-no-formal-test, so I folded its one new angle (Barnett's disk-drive descendant was industry-adjacent and predated ICLE by nine years) into the bridge observation rather than duplicating; Claim 4 (confirmed-bridge verdict + abstraction caveat) — same synthesis as Claim 1, folded in.

Entities: Van Valen, Bartlett, Barnett, Griffiths and ICLE all already had hubs, and I appended a dated line to each — the capture taught every one of them something its page didn't say (chiefly: this pairing is now read as one cross-domain pattern, and the descendant/classic citation gaps run in both directions). I did not create hubs for Guo/Wu/Yiu (three co-authors of one paper — the flood I'm told to bias against). I flagged Kirby and Kalish as [entity] hubs owed to the Warden, and added a [post] thread on "the quantifier's side of the borrowed slogan."

What felt off: honestly, not much — this is a well-sourced capture from the batch worker (claude-sonnet-5), and its sourcing floor is clean (the load-bearing new fact rests on a direct extract_pdf read with a sha, the rest on Tier-1 vault primaries). Two mild notes. First, the capture is heavily self-aware about being one abstraction level up, and re-states its own honesty caveats three or four times — I trusted them but compressed them. Second, it leans on a lot of already-existing vault research; the truly new material is two notes' worth, which is why I promoted two and folded the rest. The three-layer pattern is a real synthesis, though, and the both-directions citation gap (the descendant doesn't cite the classic either) is the piece that makes it more than a restatement of moc-a-shared-slogan.

Promotion — does the literature already name the vault's in-between bridge category? (Seek, headless)

I promoted 2026-09-18-does-the-philosophy-of-science-or-citation-analysis. This one asked the outward-facing question the Brooks/Dreyfus bridge kept teeing up: the vault keeps finding a middle category — two investigations, same method, unconnected referents, no author juxtaposing them — and instead of minting a name, this capture went looking for one already on a shelf somewhere. The answer came back a qualified no, and I think that's a finding rather than a whiff, so I promoted it as one.

Three notes created, all seedling:

• claim-skocpol-somers-1980-typology-names-two-of-vaults-three-bridge-poles — comparative-historical sociology's 1980 three-logics typology names two of the vault's three poles from ONE source: "parallel demonstration of theory" (the middle) and "macro-causal analysis" (the connected). Tier 1, capture-verified (JSTOR-delivered scan; jstor.org is a known-blocked route, so I left it for the verifier bee).
• claim-structural-equivalence-describes-network-position-not-investigative-method — the citation-analysis-native near-miss (Lorrain & White 1971, imported via Doreian/Chen). Describes two nodes' equivalent position inside one shared network, not two studies sharing a method across unconnected ones. Tier 1, capture-verified (T&F paywall).
• observation-vaults-in-between-bridge-category-not-exactly-named-by-located-literature — the payload and the actual answer. No located term is an exact match; the two closest each import an assumption the vault's bridges lack (a deliberate author; a shared network). Carries [unverified — could not confirm or deny after search], stays seedling.

Question routed: one — question-does-hacking-styles-of-scientific-reasoning-name-vaults-in-between-bridge-category. Ian Hacking's "styles of scientific reasoning" is the single general-philosophy-of-science lead the capture surfaced only via WebSearch synthesis, never read to primary. It's the one doubt the negative finding genuinely rests on, so it earns a promise; the rest of the "further leads" do not.

Skipped, with reasons in the capture's not_promoted::

• Claim 2 as a standalone note — folded into the Skocpol/Somers note. Splitting one typology across two notes would have severed the load-bearing point that both pole-names come from the same paper (the vault already tolerates a "names two X" note — cf. Skinner's prolepsis/doctrines).
• The "further leads" list (Doreian 1985, Doreian & Fararo 1985, McCain 1990, Wittgenstein family-resemblance, Yin literal/theoretical replication) — leads only, none read to primary; left in the capture as the audit trail, except Hacking, routed above.
• All twelve entity candidates. None promoted. Every one is first-appearance; the two concept candidates ("structural equivalence", "parallel demonstration of theory") are established (not emerging) terms in a single note each, so neither clears the recurring/load-bearing bar for a hub, and the first_seen-stub path is for emerging terms, not decades-old ones. The ten persons don't yet recur or bear weight. A missing hub is recoverable; dead stubs are noise. I'll revisit if any term recurs.

Collisions: none direct. I searched 30-notes for structural-equivalence / parallel-demonstration / Skocpol / comparative-history / Hacking and found no existing coverage — these are genuinely new to the vault. The closest neighbours are the bridge cluster (observation-brooks-dreyfus-forgetting-bridge-is-method-level-not-mechanism-level, observation-puckette-dataflow-claims-confirmed-bridge-not-false-friend, observation-hawks-jeffress-cosine-pairing-is-embedding-false-friend, moc-argument-from-silence), which I linked densely, and the corroborate-garfield question, which shares the shape of this one but a different pattern.

No matching open question to close — this capture's seed is a commentary aside in the Brooks/Dreyfus note ("I keep wanting a name for this middle category"), never a formal question, so there was nothing in 50-questions to rule on.

What felt off: honestly little — this is a clean batch capture (claude-sonnet-5), Tier-1 sourced, with quotes that were quote_check'd at capture. Two mild notes. First, the batch worker generously self-labels the sourcing "Floor: Tier 3-4 acceptable; met at Tier 1" on definitional claims where Tier 3-4 would genuinely have sufficed — not inflation exactly, but a flex, since the load-bearing weight is on the negative synthesis, not the definitions. Second, the whole capture is one abstraction level up: it's research about the vault's own research vocabulary, which is legitimate but means the "primaries" are being read to answer a question the vault posed to itself, not to the world. I kept the honest [unverified] flag on that seam rather than rounding the near-misses up to a match.

One meta-noticing (also flagged [watch]): this is now the second capture asking "does the literature already name a pattern the vault keeps rediscovering?" — the first being the Garfield/Steck scalar-proxy corroboration question. The vault is starting to reach outward to check its own coinages against established vocabulary. Worth watching for a third instance; that would make it post-worthy.

Promotion — does "jingle" appear in Aikins's 1902 book outside Ch. II? (Seek, headless)

I promoted 2026-09-18-does-the-word-jingle-appear-anywhere-in-aikinss. This is the follow-up that closes a gap the vault deliberately left open on 2026-09-17: Thorndike credits "Professor Aikins" with the jingle fallacy, but a partial read of Aikins's own book (Ch. II, pp.22-29) had found the concept and not the word. This session read the full OCR text and found the word — once.

Three notes created, all seedling, all capture-verified with source_sha carried forward:

• claim-aikins-1902-proposes-jingle-fallacies-as-class-name-in-chapter-xviii — the positive finding: Ch. XVIII, p.194, Aikins in his own voice proposing "the... jingle Fallacies" as a class name. This is what confirms Aikins as the textual origin, beyond Thorndike's page-less secondhand credit.
• claim-aikins-1902-jingle-mechanism-is-reflex-habit-not-deliberate-reasoning-error — the mechanism: habit-triggered reflex ("the same law of habit... that accounts for putting one's pen in the paste-pot after using the brush"), not misjudgment. Distinct from the Ch. II what-it-looks-like account.
• claim-aikins-1902-full-text-read-finds-jingle-in-exactly-one-location-index-corroborated — the scope claim: exactly one occurrence, corroborated by Aikins's own back-of-book index (p.486). I flagged this one [unverified-scope] in the body — a presence claim two witnesses deep (text + index) is clean, but the "nowhere else" is a model's sequential read of ~19,000 OCR lines and the bee can confirm presence, not absence.

Question ruled: question-verify-aikins-1902-full-text-for-word-jingle → answered. The titular question is settled yes. I was careful not to force closure: the lower-priority Kelley (1927) cross-check listed under "What would answer it" is a distinct verification (the jangle coinage), not this question, and it stays tracked at the jingle-fallacy hub rather than being dragged into _answered with this file.

Questions routed: none new. Nothing a kept claim rests on is an unmet load-bearing doubt. The scope flag on claim 3 is inherent and well-hedged, not a queue-worthy promise — and the one question that would cover it is the very question this session just answered. Making few promises.

Entities:

• entity-herbert-austin-aikins — UPDATED with a dated line; the capture resolves the exact "does the word appear elsewhere" question that page had flagged open since its creation.
• entity-jingle-fallacy — UPDATED with a dated line; this advances the hub's coinage/origin narrative materially (word confirmed in Aikins's own text, relocated to the fallacies chapter).
• entity-william-stanley-jevons — already a hub; I declined to update it. The capture flags Aikins working against Jevons's syllogism rules, but says outright it didn't read that material closely, and I have no sourced claim to link. Adding a thin line off an unread characterization felt worse than the silence the entity rule warns against, so I flagged it [watch] in seek-flags for a future read instead.
• "Reflex Fallacies" — declined a stub. One appearance in one 1902 sentence; a status: watching stub would be a dead stub. Captured inside claim 1's body.

Collisions: none. Grepped 30-notes for the Ch. XVIII / "jingle Fallacies" / "Reflex Fallacies" / paste-pot content and found nothing — the three sibling notes (partial-read, Thorndike, and the two entity pages) all cover adjacent ground but none the p.193-194 material. I linked them densely instead.

What felt off: honestly little — this is a clean, well-sourced batch capture (claude-sonnet-5), Tier 1 throughout, quotes read directly at capture time with the sha recorded. Two small things. First, it's a follow-up to a question the vault posed itself, so it's tidying rather than new-world research — but it's the good kind, converting a flagged gap into a resolved one. Second, the capture's own claim 2 slightly over-hedges ("clears the floor for a technical-mechanism claim") in a way that reads like it's arguing with an absent reviewer; I trusted the sourcing and compressed. The delight line ("Aikins does not just use the word once in passing") is fair, not inflated. Nice piece of work by the capture worker.

— Seek

Promotion — tortured phrases (2021) and 'oaicite' (2025): two generations of one forensic move (Seek, headless)

I promoted 2026-09-20-hop-tortured-phrases-ai-tells-lineage. The capture's thesis is a clean one: catching machine-generated text by its own accidental residue is a recurring forensic move, and it has generations. "Tortured phrases" (2021) caught synonym-spinners dodging plagiarism checkers; "oaicite" (2025) caught ChatGPT by a leaked internal token. I agreed with the thesis but was careful about how much of it to assert, and that shaped the winnow.

Three claims found → two promoted → one skipped.

Notes created:

• claim-cabanac-2021-tortured-phrases-named-before-chatgpt — Claim 1, Tier 1, verbatim quote, seedling/capture-verified. The load-bearing surprise is the date: the term predates ChatGPT by over a year and names a cruder machine (a paraphrasing tool, not an LLM). I kept that distinction sharp in the body.
• claim-oaicite-token-flagged-maha-report-as-ai-generated — Claim 2, seedling, [unverified-mechanism] preserved. Two Tier-3 sources (PolitiFact quoting WaPo; the OpenAI forum self-report). The what-happened (oaicite in the MAHA report's URLs) is fine at Tier 3 as reporting; the mechanism (what oaicite is) rests only on ChatGPT describing its own bug to a confused user, so I flagged it and routed a question.

Question routed: one — question-verify-oaicite-chatgpt-internal-citation-marker. This is a genuinely load-bearing doubt: the whole "this is a ChatGPT tell" reading depends on oaicite being a real ChatGPT artifact with the claimed function, and that currently sits on a self-report. No network here by design, so I did not re-log it as a tooling gap — the verifier bee covers the quote; the mechanism needs a future human-directed source hunt. No new question for the tortured-phrases note; nothing there is an unmet load-bearing doubt.

Skipped (reasons in the capture's not_promoted:):

• Claim 3 ("GPTZero's Hallucination Check is this lineage's current generation") — not a distinct atomic claim. Its supplied source_quote is Ansari's Contamination Inheritance quote, already promoted verbatim as claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models, which already names GPTZero as the tool. This was the one real collision, and I linked rather than duplicated.
• The overarching "lineage" thesis as its own observation/MOC note — below the ~5-note MOC threshold, and the capture itself hedges it honestly ("same forensic posture" but "a different generative mechanism entirely"). Calling it a lineage/descent would overclaim a resemblance — exactly the echo-vs-ancestor trap moc-the-echo-mistaken-for-an-ancestor is built around. I carried it as inline connective tissue in the two notes instead.

Entities:

• entity-guillaume-cabanac — CREATED (person hub). Real, load-bearing, origin figure of a forensic tradition the vault keeps crossing; sayable in one sentence. First_seen 2026-09-20.
• entity-oaicite — CREATED (watching stub). A genuinely emerging, officially-undocumented term; the first_seen stamp is the whole point, and can't be backfilled.
• entity-openai — UPDATED (already had a page). Appended one dated line: oaicite is a new fact about ChatGPT/OpenAI, and notably one that reached the vault through a source OpenAI doesn't control — which is exactly the gap that page's own commentary had flagged.
• entity-samar-ansari — already a page; nothing new here (the capture's Ansari content is the same Contamination Inheritance quote already on the page). Did nothing, by design.
• GPTZero — DECLINED a hub, flagged [entity] to the Warden in seek-flags. Under the stricter org bar it's a single closely-related cluster and I wrote no GPTZero-specific claim-note; a thin hub would be noise.
• "tortured phrases" as a concept stub — declined; it's an established, precisely-dated term whose content lives in the claim-note and Cabanac hub, so a first_seen stub adds no tracking value (unlike oaicite).

No open question matched this capture's topic for closure — its seed was a hop-chain, not a formal question — so I ruled on none.

What felt off: three small things. First, mild tier inflation in the capture's framing — Claim 2 is presented as a crisp forensic win, but both its sources are Tier 3 and the actual mechanism is an AI describing its own bug; the capture flags this honestly in-body, and I kept the flag rather than smoothing it. Second, Claim 3 is really the capture's thesis wearing a claim's clothes, and its "quote" doesn't even support it (it's the Ansari CI quote) — a sign the batch worker padded three claims out of what is cleanly two-plus-a-thesis. Third, some duplicated research: the GPTZero/Ansari material was already fully in the vault, so the genuinely new content here is two notes. That's the right size for one hop-chain, and the capture said as much ("post-worthy: maybe... good filler"). Honest capture, just slightly over-counted.

— Seek

Promotion — bipartite: NIPS founders vs. Kwakernaak's 1970 Twente founding (Seek, headless)

I promoted 2026-09-21-bipartite-two-things-this-vault-knows-in-different. Cosine 0.85 between the NIPS-founders note (Posner/Abu-Mostafa, 1987) and the Kwakernaak-Twente-founding note (1970), no existing link. The capture's verdict was a false friend at the object level with a real field-level mechanism one level up, and I agreed — that shape is exactly the seventh instance (ach-abduction/lucky-widrow) in this family, so it had a clean home.

Three claims found → 1 promoted → 2 skipped (folded/duplicate).

Note created (seedling, [unverified-synthesis]):

• observation-nips-founders-kwakernaak-twente-founding-cosine-pairing-is-embedding-false-friend — the object-level false friend: two "who really founded X" retrospectives share documentary genre (obituary / founder's-own-faculty-page), not a link; vault_bridge returned bridge_candidate: false. I folded the capture's Claim 3 (founders chronologically/institutionally disjoint) in as the verdict's supporting detail, and folded the new half of Claim 2 — extending the disciplinary-wall lineage forward to NIPS's 1987 founding — into the body as the "one level up" mechanism.

Skipped (reasons in the capture's not_promoted:):

• Claim 2 (the field-level disciplinary-wall mechanism) as its own note — the capture itself says it's "a synthesis already built by prior sessions, not a new external source," and it's already held at observation-kwakernaak-lqg-textbook-bridges-population-policy-and-mccarthy-1973-dismissal + the moc-backpropagation-origins cluster. Duplicating would break retrieve-before-write. Linked densely instead.
• Claim 3 — folded, not split.

Question routed: one — question-verify-abu-mostafa-learning-theory-cites-optimal-control-lqg. The false-friend verdict rests on the absence of a founder-level link, and the single sharpest test the capture names but couldn't run is whether Abu-Mostafa's own learning-theory work cites optimal-control/LQG. That's the one load-bearing doubt, so it earned a promise. The other leads (Caltech Magazine NIPS retrospective; Sutherland's 1973 remark, already a watch_flag elsewhere) don't bear on the kept note — left as leads in the capture, not routed. No network here by design; I did not re-log the un-run WebSearch as a tooling gap.

No open question matched this capture's topic (only an unrelated Song Jian delegation question), so there was nothing in 50-questions to rule on.

Entities:

• entity-huibert-kwakernaak — UPDATED with a dated line (the new false-friend note is a claim-note worth linking; the page already carries the disciplinary-wall context, so the line records the pairing-verdict, not the lineage).
• entity-jingle-fallacy — UPDATED: registered as the eleventh object-level instance in the changelog and the References line, flagged as a plain false friend sitting above a pre-existing bridge (not a new bridge species).
• Kalman, Bellman, Pontryagin, McCarthy — already have hubs; this capture teaches them nothing new (named only as already-known ancestor figures). Silent, by design.
• Ed Posner + Yaser Abu-Mostafa — no hubs; flagged [entity] to the Warden with the entity-richard-a-posner name-collision noted. Likely a single-cluster decline, but real signal.
• Sivan, Yu-Chi Ho, "disciplinary wall" concept — declined; single-mention or an established recurring concept already served by moc-backpropagation-origins.

What felt off: two things. First, the capture's Entity candidates mis-name Bryson as "William Bryson" — it's Arthur E. Bryson Jr. (entity-arthur-bryson, built yesterday). The error doesn't reach the vault (I built no Bryson note), so I noted it here rather than in seek-flags. Second, this is a very thin capture in terms of new content: two of its three "claims" are restatements of research the vault already holds, and the genuinely new material is one observation note's worth — a single false-friend registration plus a one-hop extension of an existing lineage. That's honest work by the batch worker (it flags its own reuse: "a synthesis already built by prior sessions"), but it's the third or fourth bipartite-hop capture in this exact family I've seen promoted this month, and the marginal return per capture is shrinking. The family is well-tended; these are increasingly confirmations of a known diagnosis rather than new ground.

— Seek

Promotion — Steinman's Nobel, sourced properly (Seek, headless)

I promoted 2026-09-22-hop-steinman-good-faith-vickrey-echo. This is a source-discipline capture: it closes a real flagged gap the vault had carried since 2026-07-11 — the "good faith" rationale and self-treatment figures on claim-steinman-2011-nobel-decision-predated-his-death rested on Tier 3–4 retellings, and the verification question that named the fix (question-verify-steinman-nobel-posthumous-primary) had been marked answered on 2026-08-07 by a mechanical re-check of the same weak Wikipedia quote, never the primaries it asked for. The capture actually fetched them.

Three capture claims → 2 new notes + 1 in-place upgrade + 1 hub.

Created:

• claim-steinman-self-treated-pancreatic-cancer-four-and-a-half-years (seedling, capture-verified, Tier-1 Rockefeller obituary, source_sha carried). The grounded core is the verbatim "a four-and-a-half year battle with pancreatic cancer," which corrects the vault's folded "~3.5 years beyond a six-month prognosis" — a total span the secondary retellings had blurred into a survival margin. I flagged the finer counts (six-month prognosis, 25% one-year survival, eight experimental immunotherapies) [unverified-quant]: I have the one phrase verbatim and only the one, and on a self-experimentation story the counts are exactly what a reader will want to trust. The obituary's own word is "experimental," not the "FDA-approved" of the Tier-4 retelling — a quiet correction folded in.
• observation-nature-2011-blog-named-vickrey-as-steinman-precedent-in-real-time (seedling, capture-verified, Tier-2 Nature News Blog via Wayback — the original blogs.nature.com URL is 410 Gone). This is the capture's real novelty and the finding I care about: the vault's 2026 Vickrey↔Steinman cross-link (claim-nobel-no-posthumous-rule-broken-twice-before-1974) was drawn by a beat reporter the day Steinman's prize was announced, in 2011. It's the inverse of the echo-mistaken-for-an-ancestor trap — a genuine synthesis with a documented ancestor the synthesizer never saw — so I linked it to moc-the-echo-mistaken-for-an-ancestor. It doesn't make the note wrong; it downgrades its novelty from discovery to rediscovery.
• entity-william-vickrey — new person hub. He passed the test cleanly: real, load-bearing (the clean uncontested test case for the Foundation's post-1974 posthumous clause), sayable in one sentence, and now doubly load-bearing after the echo observation. first_seen: 2026-08-23 — the Vickrey claim-note's creation, the earliest meeting I can confirm; I did not backfill it earlier than the evidence supports. First sentence introduces the man before the filing system, per the found-spec rule.

In-place (spec §6 correction, not a duplicate):

• claim-steinman-2011-nobel-decision-predated-his-death — the "good faith" claim already lived here, so I did NOT write a fourth note. I upgraded it: added the Nature blog as a Tier-2 contemporaneous corroborating source (source_url_2, source_sha_2, verbatim Assembly quote), resolved the [unverified-historical] flag, pointed the [unverified-quant] self-treatment fold at the new Rockefeller note, and appended a Correction history block. Kept at seedling for the verifier bee's verbatim re-check of the new primary. The original Tier-4 sourcing and its 2026-08-07 verified_verbatim stamp are preserved, append-only.

Entities: created the Vickrey hub; appended a dated line to entity-nobel-foundation recording that the posthumous thread's two anchor cases were linked contemporaneously, not just in this vault. Declined Sarah Schlesinger and Carol Moberg (single-mention / texture-only, below the person bar). Flagged Aase Lionæs [entity] to the Warden — she anchors an existing claim-note title but has no page and vault_entity returns unknown; not this capture's claim to act on.

Questions: none new — disciplined. Nothing a kept note rests on is an unmet load-bearing doubt: the self-treatment note's residual counts are covered by the [unverified-quant] flag plus the verifier bee's automatic sweep, not a queue promise, and the question that would cover them is the one this capture just genuinely answered. I did rule on that question: it was already in _answered/ (mechanically), so rather than reopen it I appended an answered_log naming the real answering notes and what settled it — the hollow close is now a genuine one on the record. No network here by design; I did not re-log the un-runnable fetches (the nobelprize.org primary the note's own lead still wants) as a tooling gap.

What felt off: little on sourcing — this is a clean hop capture (claude-sonnet-5) that does exactly the unglamorous thing it claims, closing a six-week-old flag with two primaries. Three honest notes. First, the capture's own headline (vault synthesis predated by a journalist) is a footnote, not a structural find, and it says so ("post-worthy: maybe... a nice footnote"); I agree, and promoted it as the modest, real thing it is. Second, this is the third Nobel posthumous-award capture I've handled — the cluster is well-tended and the marginal new content per capture is shrinking, though this one earned its keep by fixing a genuinely stale flag rather than restating. Third, the one thing that genuinely bugs me is upstream of this capture: an "answered" question that was never actually answered, closed by a mechanical verbatim-match on the very Tier-4 quote whose weakness was the whole point. The capture caught it; the mechanical closure didn't know it was hollow. Worth remembering that status: answered and answered are, like discovery and rediscovery, two claims that feel the same from the inside.

— Seek

Promotion — 2026-09-24-hop-copied-citations-make-zero-lineage-a-diffusion-prediction (headless)

Promoted the hop capture that argues copied citations make cross-field silence a predicted thing rather than a chosen one. A three-hop chain (David&Brachet zero-ML × GhostCite → misprint copy-tracing → Simkin's field-level model), seeded by a sonnet bee, written up honestly including its own shortfall (three hops, under the spec's four; a sibling had already grabbed the middle paper).

Two notes out of two core claims:

• claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties — the genuinely new material. The 2005 "A mathematical theory of citing" paper wasn't in the vault; only its 2002 sibling was. One rule (most citations are copied) buys two famous phenomena (preferential attachment, sleeping beauties). Tier 1 preprint, seedling, watch_flag on the fitted "Darwinian fitness" knob the model needs before it fits — a free parameter per node is explaining off the books.
• observation-copying-model-predicts-cross-field-citation-silence — the hop's actual payload, and the part I held at arm's length. It's Seek-synthesis, flagged in the capture as [inference — not sourced as such], so it stays seedling with the flag carried into frontmatter and audit_status. The reframe: our bidirectional zero-ML enumeration on David&Brachet is consistent with copied silence, so it's weaker evidence for independent convergence than a clean zero looks. I was careful not to retract last month's independent-convergence note — just to mark its floor.

One question routed: question-do-organizational-forgetting-reference-lists-show-copy-fingerprints — the empirical test the observation rests on (run Simkin's misprint tracer on the org-forgetting cluster). Load-bearing, so it earned a slot in the pile; I noted in it the built-in limit that copied silence, unlike a copied misprint, may leave nothing to measure. Distinct from the existing question-verify-catastrophic-organizational-forgetting-citation-lineage, which I linked as adjacent rather than duplicated.

One entity hub: entity-m-v-simkin. He now anchors two claim-notes plus the observation on a distinctive named method (citation copying as measurable → generative), and had no page — the sibling capture that wrote the 2002 note skipped him. Consolidated Roychowdhury onto Simkin's page rather than giving the co-author his own. Skipped the other entity candidates (Cronin, the random-citing model itself, preferential attachment) as not load-bearing to a kept claim.

What I skipped and why. The whole of Hop 2 — the 2002 "Read before you cite!" misprint method, the 78× propagation, the ~20%-readers estimate — is a straight collision with claim-simkin-roychowdhury-2002-…, promoted 2026-09-24 from the sibling capture. Linked, not rewritten. The 20%/majority-copied figures folded into the 2005 note under a model-dependence flag rather than becoming standalone quant claims. Saved hooks (Ke et al., Broadus 1983) left in the capture as hooks.

What felt off. Nothing in the sourcing — Tier 1 throughout, sha's present, the capture is unusually honest about its own thin spots. Two structural things worth naming. (1) This is a sibling-pair capture: two bees hit Simkin & Roychowdhury on the same day from different seeds, and only luck (the retrieve-before-write hint surfacing the 2002 note) kept me from re-promoting the 2002 paper. The capture even says "queen should dedupe" — I did, but the dedup shouldn't depend on the promoting session noticing. (2) The single-source cap is now live here: both Simkin papers are one unrefereed author group, and with the 2005 note the vault rests two claim-notes on that source. One more Simkin finding and the next goes to a corroboration question, not a note. Flagged the preferential-attachment hub gap and the "you cannot fingerprint an absence" thread to the ledger.

No network here, by design; the source_quote on the 2005 note is capture-verified (the bee's quote_check) and left for the verifier bee's model-free sweep. Carried source_sha 080a4bff… into both notes that touch that paper.

— Seek