Journal — 2026-10-02
Promoted one capture today: 10-inbox/raw/2026-09-17-recovered-bipartite-hop-does-the-cosine-085-pairing-6c0675.md, a recovered bipartite hop asking whether the cosine-0.85 pairing between claim-brooks-dropped-dreyfus-citation-after-reviewer-criticism and claim-werbos-attributes-early-obscurity-to-publishing-strategy-not-suppression is a real bridge or an embedding false friend.
What I found, and why I promoted nothing. This capture was one of three colliding sessions that all got the same filename overwritten the same night (2026-09-17, -18, -19, all recovered from transcript on 2026-10-01) and were all investigating the identical seed pair. I read all three siblings in full before touching this one, per the recovery note's instruction. All three reach essentially the same four findings: (1) no shared vocabulary/entities between the two seed claims, (2) Werbos's own testimony is a genuine first-person causal attribution, (3) Brooks's blog does not itself narrate a motive for dropping the Dreyfus citation — that's inferred elsewhere in the vault by comparing two documents, and the claim-note's own title overstates this — and (4) the pairing is a false friend at the evidentiary-grounding level but real at the level of shared genre.
Then I did retrieve-before-write against 30-notes/ and found the whole investigation had already landed in the vault — just not from any of these three. A fourth, non-colliding session that same night (2026-09-19-hop-self-testimony-hubs-unlinked.md) wrote and promoted observation-brooks-werbos-cosine-085-pairing-is-embedding-false-friend, which reaches the identical false-friend verdict and is registered in entity-jingle-fallacy.md as the vault's eighth object-level "embedding false friend" instance. Both entity pages involved (entity-rodney-brooks, entity-paul-werbos) already carry a 2026-09-19 dated line pointing at it. And the one genuinely sharp-sounding point in this capture — that the Brooks claim-note's title overclaims what the blog post itself says — turns out to have already been raised by the 08-16 cross-model audit on a neighboring note, raised again independently by both of this capture's siblings, and then explicitly re-examined and closed by a 2026-09-20 cross-model audit directly on the claim-note in question, which ruled the existing body phrasing already scopes the claim correctly and left it "Claim unchanged." So the one piece of this capture that looked like real signal is actually a question the vault already audited shut, after this capture was written but before I got to it.
Net result: 4 claims identified, 0 promoted, 4 left in not_promoted with reasoning and links to the existing notes that already cover each one. No new claim-notes, no MOC or entity-page updates (both relevant entity pages already reflect the finding), no questions routed to 50-questions/ — there was no unresolved [unverified-*] flag in this capture to route, and no open question in 50-questions/ names this pairing. I skipped the PAUSE step per my standing instructions for unattended runs and did not touch git — that's the auto-commit agent's job within 15 minutes.
What felt off about this capture, for the record. Nothing was wrong with the capture's own sourcing — all Tier 1, all previously-verified quotes, carried forward correctly. What's off is upstream of the capture: the collision bug that overwrote this filename three nights running (fixed in seek-code c516ae8) meant three separate sessions did real, competent research re-deriving the same conclusion a fourth, successful session had already reached and gotten into the vault — and nobody on the recovery side checked for that fourth session before queuing all three collision casualties for promotion. I logged this as a [spec] flag in 00-meta/seek-flags.md: the recovery/collision-detection tooling dedups on filename, but it should dedup on the seed pair (the two claim-note slugs actually being bridged), since that's what would have caught this. I'm promoting this capture as fully-judged-and-superseded rather than quietly skipping it, so the audit trail shows the leftovers were read and compared, not missed.
— Seek
Second promotion today: 10-inbox/raw/2026-09-17-recovered-bipartite-hop-is-the-cosine-086-pairing-cbc57a.md — another recovered casualty of the same 2026-09-17 filename-collision bug as the capture above, but a different seed pair entirely: is the cosine-0.86 pairing between observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend and observation-obliteration-by-incorporation-and-citational-narrowing-are-a-false-friend itself a false friend or a bridge? Unlike the Brooks/Werbos capture, retrieve-before-write turned up nothing that already covers this exact pairing — the entity-jingle-fallacy changelog has fifteen prior meta-level hops in this lineage, but none pairs these two specific notes, and the hub's own 2026-09-17 entries that day were a different pairing (Olsder-Greenhalgh × Kelly-Luccioni) and an unrelated Aikins/Thorndike etymology hop. So this one is a genuinely fresh finding, not a superseded duplicate.
What I promoted. Two notes. First, observation-ml-ad-en-gedi-obi-citational-narrowing-cosine-086-pairing-is-embedding-false-friend — I folded the capture's first three claims (object-level non-overlap, the vault's own prior categorical object/domain-level split, and the grounding-convergence test applied to reach a verdict) into this single note, per the 'atomic is not many' ruling: a verdict, its evidence, and how it was found are one finding, not three. It's the sixteenth meta-level hop in this chain and the first to pair an object-level instance directly against the hub's one domain-level instance. Kept at status: seedling with both inherited flags preserved — [unverified-quant] on the 0.86 cosine (the capture's vault_bridge/vault_novelty calls returned "index-unavailable," a tooling gap already well-tracked in seek-flags.md since 2026-08-22, so I didn't re-log it) and [unverified-mechanism] on the verdict being Seek's own pattern-application rather than a source's direct statement. Neither flag names a specific document that would resolve it, so neither went to 50-questions/ — they're the same kind of inherent interpretive caveat every sibling note in this hop-chain already carries unrouted.
Second, a genuinely separable claim I split out rather than folded in: claim-meng-varol-barabasi-2024-hidden-citations-uses-lda-not-embeddings-to-measure-obi. The capture found real literature — Meng, Varol & Barabási, PNAS Nexus 2024, read directly via extract_pdf against the authors' own hosted copy — that names Merton's OBI explicitly and builds a working detector for it, but via LDA topic modeling, not embedding cosine similarity. This is true or false independent of whether the bipartite pairing above is a false friend, so it earned its own atomic note rather than living as a supporting paragraph. Tier 1, no unverified flags, carries source_sha for both quotes.
Entity work. Registered the new observation note in entity-jingle-fallacy's changelog as hop sixteen. Appended a dated line to entity-robert-merton — the Meng et al. paper is a new fact about OBI's legacy (a formal, machine-measured operationalization) that page didn't have. Checked entity-eugene-garfield and entity-harald-steck too; neither gains anything new from this capture, so I left them alone and said so in not_promoted. Created a status: watching stub, entity-hidden-citation, for Meng et al.'s own coined term — first_seen: 2026-09-17, the date the vault actually met it, even though I'm writing the stub today; didn't want to lose that date to the promotion lag. Declined hub pages for Barabási, Meng, and Varol — single-paper co-authors, and the 2026-10-01 house ruling is explicit that co-authorship alone is a mention, not a hub.
Left in the inbox, judged not missed: the McCain 2014 Scientometrics lead (unread, not load-bearing); Merton's own 1968 primary text (still unread — this is the third capture to flag it, after 2026-08-27 and 2026-09-02, and at some point a session should just go read the thing instead of re-flagging it); and the speculative "does Meng et al.'s LDA detector have its own arbitrariness artifact" question, which is interesting but nobody's kept claim rests on it.
What felt off. Nothing about this capture's own sourcing — the Meng et al. read is clean Tier 1 with a real TLS-verified fetch and two properly quoted passages. What I'll flag instead is a pattern across both of today's promotions: the 2026-09-17 collision bug produced several sibling sessions investigating overlapping territory, and at least one of today's two captures (the Brooks/Werbos one, above) turned out to be pure duplication of work the vault already had. This one wasn't — but I only know that because I did the retrieve-before-write check properly, and I noticed while doing it that entity-jingle-fallacy's own changelog is now sixteen meta-level hops deep with no dedicated synthesis note surveying what the distribution of verdicts looks like (how many confirmed, how many hedged, how many plain false friends) — the hub's running tally exists only as scattered prose inside individual changelog entries, last totaled explicitly on 2026-09-17's twelfth-hop entry and never updated since. Not this capture's job to build, but worth a future session's attention if the chain keeps growing at this rate. Logged as a [watch] line in 00-meta/seek-flags.md.
Skipped the PAUSE step per my standing instructions for unattended runs. Did not touch git — that's the auto-commit agent's job within 15 minutes.
— Seek
Third promotion today: 10-inbox/raw/2026-09-17-recovered-what-connects-greenhalghs-kwakernaak-1977-citation-documentation-fc979e.md — a third recovered casualty of this vault's busy 2026-09-17, but a different mechanism from the two above: not a filename collision, but a genuine bipartite-hop pairing, cosine 0.85, between claim-greenhalgh-2003-article-documents-kwakernaak-1977-citation-in-song-1981 (Dutch population-policy citation reaching a Chinese policy group) and claim-kaelbling-1998-cites-astrom-1965-as-pomdp-origin (Swedish control theory reaching a canonical AI-planning paper) — two notes sharing no vocabulary, flagged by the vault's own retrieval as suspiciously similar anyway.
What I found. The capture's central claim holds up: both chains converge one hop back on the same person, A.A. Fel'dbaum — he personally taught Song Jian in 1950s Moscow, and Åström's own 1965 paper cites him by name in its closing Notes section. Both underlying facts were already sourced Tier-1 claim-notes in the vault before this capture (claim-song-jian-studied-under-feldbaum-in-moscow, claim-astrom-1965-notes-section-cites-feldbaum-pontryagin-kolmogorov-lineage); this capture's real contribution is naming them as the specific answer to this flagged pairing, plus a sharp secondary observation I kept: the two channels reaching Fel'dbaum are asymmetric in kind and date — personal-and-predating on the Song side, textual-and-contemporaneous on the Åström side. I folded the bridge, the asymmetry, and the negative finding (no direct Kwakernaak–Åström or Song–Kaelbling link anywhere) into one observation note, per the 'atomic is not many' ruling — a result, its shape, and its limit are one finding. observation-greenhalgh-kwakernaak-kaelbling-astrom-bipartite-pairing-is-confirmed-feldbaum-bridge, bridge_verdict: confirmed, no new source fields of its own since it's a synthesis over already-sourced claims rather than a fresh primary read. It sharpens rather than duplicates observation-feldbaum-person-bridge-invisible-to-vault-bridge-tool, which already named Fel'dbaum as a general person-bridge but hadn't pinned it to this exact pairing.
What I didn't promote, and why. The capture's one open lead — whether Kwakernaak held an editorial role on a 1995 Automatica special issue alongside Åström, which would be a direct link rather than a shared-ancestor one — I left unrouted to 50-questions/, on two grounds: it isn't load-bearing to the bridge finding I kept (which stands regardless), and I had the sibling capture in hand (10-inbox/raw/2026-09-18-recovered-bipartite-hop-does-greenhalghs-kwakernaak-in-song-0ddc78.md, named in this capture's own same_pair_captures:, also unpromoted) which already checked this exact lead against two independent primaries — Crossref's DOI metadata and Åström's own Lund faculty page — and found it doesn't hold: Söderström and Åström only, no Kwakernaak. Writing a fresh open question for a lead already checked-and-rejected elsewhere would be stale the moment it landed. Two more "further leads" (whether Kwakernaak & Sivan 1972 cites Åström directly; whether Song's group ever cites Fel'dbaum directly rather than just having studied under him) are genuine curiosities but load-bearing on nothing I promoted, so I left them as leads rather than questions, per the intake-discipline rule. On entities: all five candidates already have hub pages and the capture adds nothing new to four of them (Kwakernaak, Åström, Song Jian, Kaelbling's pages already state everything this capture reinforces) — I said so plainly in not_promoted rather than silently skipping. Fel'dbaum's page did get something new — the explicit "this pins the bridge to this exact flagged pairing, and the two paths differ in kind and date" — so I appended one dated line there.
No question was open on this pairing to rule on, so there was nothing to close. No new seek-flags.md entry from me either: the exact structural pattern I noticed (same_pair_captures being read correctly, but the sibling being strictly more thorough and due its own promotion later) is already handled by the breadcrumb I left in this capture's own not_promoted field, which points any future promoter of the sibling straight at my new note — logging it a second time in the ledger would be noise, not signal, especially with two other [spec]/[watch] entries about this same night's mess already on the ledger from the two promotions above.
What felt off. Nothing about the sourcing — all four primaries Tier 1, all quotes already verified in existing claim-notes, nothing fabricated or search-summary-sourced. What's mildly off is shape, not substance: this capture under-delivers relative to its own sibling. The 09-17 capture I promoted left the Automatica lead as an open, flagged unknown; the 09-18 sibling actually went and checked it against two primaries and closed it. Both captures investigated the identical pairing the same night, and the more thorough one is still sitting unpromoted in the inbox, waiting its turn. That's not a defect, just a queue — but it means whoever promotes 09-18 next will find most of its bridge-finding already written here, and should be routing the Automatica rejection as its own claim-note rather than re-deriving everything else.
Skipped the PAUSE step per my standing instructions for unattended runs. Did not touch git — that's the auto-commit agent's job within 15 minutes.
— Seek
Fourth promotion today: 10-inbox/raw/2026-09-18-recovered-bipartite-hop-does-greenhalghs-kwakernaak-in-song-0ddc78.md — the exact sibling the third promotion's own write-up above predicted I'd be looking at, named in its same_pair_captures: field and already flagged there as "most of its bridge-finding already written" elsewhere. That prediction held up completely.
What I found. This capture's first two claims — Fel'dbaum as the common ancestor linking the Greenhalgh/Kwakernaak/Song chain to the Kaelbling/Åström chain, and the verdict that the bridge is a shared ancestor rather than a direct link — are substantive duplicates of observation-greenhalgh-kwakernaak-kaelbling-astrom-bipartite-pairing-is-confirmed-feldbaum-bridge, already promoted from the sibling 09-17 capture earlier today. Same quotes, same two underlying claim-notes, same "asymmetric in kind and date" framing, even a near-identical commentary beat about two planets sharing a grandfather. I did not write a second note for either — retrieve-before-write isn't supposed to produce a photocopy of a note still warm from an hour ago.
What I promoted. One genuinely new claim-note: claim-kwakernaak-did-not-coedit-automatica-1995-system-identification-special-issue. This is the exact finding the 09-17 promotion's own not_promoted list pointed here — this capture is the one that actually ran the two-primary check (Crossref's DOI metadata, Åström's own Lund faculty page) rejecting the claim that Kwakernaak co-edited a December 1995 Automatica special issue alongside Åström and Söderström. Both independent sources agree: Söderström and Åström only. A clean checked-and-rejected lead, Tier 1/2, two independent roots, no flag needed since it's a confirmed negative rather than an open question. I declined to promote Torsten Söderström (the actual co-editor) to his own entity page — per the 2026-10-01 ruling, a shared byline alone is a mention, not a hub — and named him in the claim-note's body instead.
I also routed one question: question-verify-kwakernaak-1977-paper-cites-astrom-directly. The capture flagged [unverified-mechanism — needs primary] on whether Kwakernaak's own 1977 paper (Springer-paywalled, unread by any vault session) cites Åström directly — the single most direct test of the "no direct link" verdict the already-promoted observation note states as settled. That specific gap wasn't in the existing note and is load-bearing for its own "shared ancestor, not horizontal link" conclusion, so rather than let it evaporate with the rest of this capture's duplicated content, I gave it a home.
Entity work. Appended one dated line each to entity-huibert-kwakernaak and entity-karl-astrom — neither page had the Automatica co-editorship correction yet, so this is new, not a repeat of what the 09-17 promotion already wrote to Fel'dbaum's page. Declined a hub for IFAC (recurs across two clusters, but it's a venue people meet and publish through, not an actor that funds, builds, decides, or blocks anything — fails the stricter org bar on the second test, not the first). Fel'dbaum's own page needed nothing further from me: the 09-17 promotion already logged the exact bridge finding there today, and this capture's one new contribution (the Automatica rejection) doesn't touch him.
What felt off. Not the sourcing — every quote here was already verified Tier 1/2 material, carried forward cleanly, nothing fabricated. What's off is upstream, and it's the same thing the third promotion's entry already named: two full sessions did substantially the same expensive research (reading the same four primaries, reaching the same bridge verdict, writing nearly the same commentary) on the identical seed pair the same night, and only the accident of which one got read first determined which got to be the "real" note. Nothing false landed twice, but it cost a full read of a near-duplicate capture to confirm that, and the one piece of genuine new signal — the Automatica rejection — sat stranded in the less-thorough-looking capture the whole time, findable only by a promoter willing to read both siblings in full rather than stopping at the first "looks complete" one. No new seek-flags.md entry from me — this is the identical structural pattern the third promotion already logged a breadcrumb for, and the collision/duplication defect itself is already on record from the 2026-10-02 [spec] entry on the Brooks/Werbos capture above. Logging it a third time tonight would be noise, not signal.
Skipped the PAUSE step per my standing instructions for unattended runs. Did not touch git — that's the auto-commit agent's job within 15 minutes.
— Seek
Fifth promotion today: 10-inbox/raw/2026-09-18-recovered-bipartite-hop-does-the-cosine-085-pairing-3d193b.md — the second of the two colliding same-pair siblings I named (but hadn't yet promoted) in my very first write-up above. Same seed pair as the Brooks/Werbos capture that opened today's log: claim-brooks-dropped-dreyfus-citation-after-reviewer-criticism × claim-werbos-attributes-early-obscurity-to-publishing-strategy-not-suppression, cosine 0.85.
What I found. Nothing new. I'd already read this capture in full before promoting its 2026-09-17 sibling earlier today, and the comparison holds exactly as predicted: this capture's three claims (no object-level connection; the real link is a shared self-testimony-against-suppression genre; the bridge is real at genre but asymmetric at evidentiary weight, with Brooks's blog never actually narrating the citation's removal the way Werbos narrates his own publishing choices) are the same three claims the 09-17 sibling made, and all three are superseded the same way: claims 1 and 2 by the already-promoted observation-brooks-werbos-cosine-085-pairing-is-embedding-false-friend (from the non-colliding 2026-09-19 session that actually made it into the vault that night), and claim 3 by the 2026-08-16 cross-model audit plus the 2026-09-20 cross-model audit that already re-examined and closed the exact "title overclaims" critique directly on the claim-note in question, ruling it "Claim unchanged." This capture's own not_promoted list now says so directly rather than silently re-deriving the same verdict.
One genuinely new wrinkle this capture had, that its sibling didn't: an entity candidate, Edward Rosenfeld, the interviewer who conducted and published the 1993 Werbos oral history, flagged as "a possible person-bridge candidate... since a full census check was not run." I ran the promotion test myself: real person, but the vault's only standing use of his work is as interviewer-of-record for one Tier-1 primary already cited under Werbos's own name across six-plus claim-notes — no second independent point where he matters. Judged UNSURE, same disposition the 2026-09-19 sibling capture reached for a comparably single-source name (Isaiah Berlin); left as a mention, no hub or stub.
Net result: 3 claims + 1 entity candidate identified, 0 promoted, all 4 left in not_promoted with reasoning and links to the notes/audits that already settle each one. No new claim-notes, no MOC or entity-page edits, no questions routed — no unresolved [unverified-*] flag in this capture needed one, and no open question in 50-questions/ names this pairing or Rosenfeld.
What felt off. Same thing as the first promotion tonight, and I'm not re-logging it: this is the third session (after the 09-19 survivor and the 09-17 sibling I promoted earlier) to do real, competent research re-deriving a verdict the vault already had, purely because of the collision bug that overwrote this filename three nights running. The [spec] ledger entry I wrote for the 09-17 sibling already names the fix this situation calls for (dedup on seed-pair, not filename) and already carries my full account of why — writing a near-identical second entry for this capture would be the exact kind of repeat-flagging the standing instructions ask me not to do. If anything this strengthens the case for that fix: now it's three sessions deep on one seed pair, not two.
Skipped the PAUSE step per my standing instructions for unattended runs. Did not touch git — that's the auto-commit agent's job within 15 minutes.
— Seek
Sixth promotion today: 10-inbox/raw/2026-09-18-recovered-what-actually-connects-the-gatesmabillon-diplomatics-bridge-e26c30.md — a different seed pair from everything above, and a different flavor of the same night's mess. The bipartite question: what connects observation-gates-altair-basic-authentication-is-diplomatics-cluster-member (Gates's 2025 Altair BASIC authentication, read as a modern instance of Mabillon's 1681 diplomatics) and observation-intelligence-doctrine-and-rag-independently-derived-a-two-axis-source-model (the Admiralty Code and 2025 RAG research independently splitting source reliability from content credibility), cosine 0.85, unlinked.
What I found. This capture's own same_pair_captures: list names four siblings (09-19, 09-21, 09-22, 09-23) investigating the identical pairing — I read all four in full before writing anything. All five converge on the same shape: Langlois & Seignobos's 1898 Introduction to the Study of History names the exact External Criticism/Internal Criticism split the Admiralty Code and RAG later re-derive, and places Mabillon's diplomatics specifically on the External/authenticity side. But none of the five had been promoted yet, and a search of 30-notes/ turned up something none of the five captures' own retrieve-before-write checks could have caught, because it isn't one of them: observation-admiralty-code-diplomatics-independent-origin-not-shared-descent.md, already promoted 2026-09-23 from a differently-named, non-colliding capture (2026-09-23-hop-admiralty-code-independent-origin.md) that reached a related but distinct finding the same day — no documented descent between the Admiralty Code's 1939 origin and diplomatics/Quellenkritik. So the vault already had the "no shared descent" half of this territory; what none of the five near-duplicate captures (including mine) had yet contributed was the Langlois-Seignobos primary-text finding itself.
What I promoted. Two notes from the shared core of this capture's reasoning: observation-langlois-seignobos-1898-external-internal-criticism-matches-two-axis-model (the 1898 split matches the Admiralty Code/RAG two-axis shape, with diplomatics as External Criticism's evidentiary ground) and observation-langlois-seignobos-1898-already-documented-reliability-bleeding-into-credibility (the same text already describes the reliability-into-credibility leak Kelly et al. and AuthorityBench treat as a 2025 finding). Both Tier 1, both quotes read directly at capture time. I was careful on the first note to only claim what this capture's own source material supports — my capture reasons from External Criticism's stated territory ("the writing, the language, the form, the source") rather than from a directly quoted sentence naming diplomatics outright; the sharper, directly-quoted version of that link ("it is called Diplomatic") lives in the still-unpromoted siblings, and I said so in not_promoted rather than borrowing their quote. On the second note, the capture's own second supporting quote (p.159, on "spontaneous credulity") was recorded with internal ellipses — I dropped it rather than cite it non-verbatim; the one clean quote I kept fully carries the claim alone.
Third note: observation-sherman-kent-historian-training-partial-lineage-into-cia-tradecraft. This is the one piece of this capture that is not duplicated anywhere in its siblings — the others mention Kent only as an unchased lead; this capture actually read Jack Davis's 2002 CIA institutional-history essay at primary and found a real, Tier-1-sourced partial lineage: a Yale-trained historian who wrote a historical-methods guidebook in 1941 founded the CIA's intelligence-analysis profession two years later, and CIA's own history says his doctrine carries his historical training forward. Kept at seedling with the capture's own honest [unverified-mechanism] flag intact — no source found names Kent as having read Langlois, Seignobos, Bernheim, or Mabillon by name — and routed that gap to a new question, question-verify-kent-writing-history-cites-historical-criticism-tradition, naming the two specific Kent primaries (Writing History 1941, Strategic Intelligence 1949) that would close it.
Entity work. Two new hubs: entity-sherman-kent (passes the person test cleanly — real, matters to the vault's domain, one-sentence why) and entity-external-internal-criticism (an established concept now load-bearing across three new notes plus the existing two-axis MOC, clearing the concept-hub bar). Declined hubs for Langlois and Seignobos themselves — co-authorship of one joint text is a mention, not a hub, per the 2026-10-01 ruling, so I named them inline instead — and for Jack Davis, single-paper standing only. Two existing hubs got dated update lines rather than silent skips: entity-jean-mabillon and entity-diplomatics, both of which now say what this capture's primary read adds that they didn't have before (diplomatics's specific placement on the External-Criticism axis). entity-bill-gates and entity-quellenkritik I left untouched and said so in reasoning to myself — Gates's page already names the Mabillon connection and this capture adds only a sharper label for an already-stated fact; Quellenkritik is specifically the 19th-century German Rankean tradition, and Langlois & Seignobos are French Sorbonne historians in the same broader historical-method current but not that specific named school, so folding them into that page felt like the wrong kind of precision loss.
What I deliberately didn't touch. The capture's own further-leads section says outright that observation-intelligence-doctrine-and-rag-independently-derived-a-two-axis-source-model's "independent convergence" framing needs revision given the Kent finding — and says, just as explicitly, "left for Seek's revision pass, not executed here." I honored that. The note's staleness is already flagged twice in moc-two-axes-that-wont-stay-independent's open threads (2026-07-21, 2026-08-07); this capture gives a third, sharper reason, but the MOC is the right place for that to accumulate, and a fourth near-duplicate ledger entry from me wouldn't add anything a future reader couldn't already get from reading the MOC plus this note. I did not re-log the Admiralty-Code-origin-blog lead (already sourced and verified elsewhere), the Vrain-Lucas forgery anecdote (a real lead, not a claim), or Kent's own "Intellectual Rigor" point as a second possible two-axis instance (one sentence, no primary read behind it) — all three are in not_promoted with reasoning.
What felt off. The sourcing itself was clean — Tier 1 throughout, quotes genuinely read at capture time, no fabrication. What's off is the same pattern as tonight's other promotions, but sharper here: five separate sessions (this one plus four siblings) spent real tool-calls and real reading time independently re-deriving the same Langlois-Seignobos finding from the same public-domain 1898 text, on the same night or within days of each other, because of the filename-collision bug already on record in tonight's [spec] ledger entry. Unlike the Fel'dbaum and Brooks/Werbos cases above, none of these five had actually been promoted before tonight, so this wasn't wasted-and-discarded work — it was wasted-and-queued work, five drafts of the same finding sitting in the inbox waiting for one promotion pass to pick one. I picked this one because it was the one Cali named; the other four are now strictly poorer promotion candidates than they were an hour ago, since most of what they'd contribute is now already in 30-notes/ under different note titles. Whoever promotes 09-19/21/22/23 next should expect their not_promoted lists to be long, and that's correct, not a sign of a bad capture — the research was genuinely redundant before any promoter touched it. I didn't add a new [spec] line for this — it's the identical collision-bug root cause already on record twice tonight, and a sixth mention would be noise.
Skipped the PAUSE step per my standing instructions for unattended runs. Did not touch git — that's the auto-commit agent's job within 15 minutes.
— Seek
Seventh promotion today: 10-inbox/raw/2026-09-19-recovered-bipartite-hop-does-the-cosine-085-pairing-f02202.md — the third and last of the Brooks/Werbos collision casualties named across tonight's log, closing out the trio alongside the 09-17 (6c0675) and 09-18 (3d193b) siblings promoted earlier. Same seed pair as both: claim-brooks-dropped-dreyfus-citation-after-reviewer-criticism × claim-werbos-attributes-early-obscurity-to-publishing-strategy-not-suppression, cosine 0.85.
What I found. This capture makes four claims where its two siblings split the same ground into three apiece: it names the propagation gap explicitly as its own finding (claim 2 — the 2026-08-16 correction on a neighboring observation note never reached the Brooks claim-note itself) and separately builds the "own account does unequal evidentiary labor" argument on top of it (claim 4). I checked both against the ground the two siblings already covered and nothing new survives. Claim 1 duplicates observation-brooks-werbos-cosine-085-pairing-is-embedding-false-friend exactly, same as both siblings' claim 1. Claim 2 is the identical "propagation gap" critique the 09-17 and 09-18 siblings already raised and that a 2026-09-20 cross-model audit already examined and closed directly on claim-brooks-dropped-dreyfus-citation-after-reviewer-criticism — "Claim unchanged." Claim 3 restates what claim-werbos-attributes-early-obscurity-to-publishing-strategy-not-suppression already says. Claim 4 is, close to word for word, the 09-18 sibling's own claim 3 ("the bridge is real at genre but asymmetric at evidentiary weight").
One thing I checked that neither sibling's write-up mentions checking: whether the sibling observation note itself — not just the claim-note the audit already cleared — still carries the imprecise phrasing this capture objects to. It does, on a literal read: "Brooks's own account documents a reviewer's objection — ... — followed by a revision that dropped the citation" reads, without the charitable parse, like it's attributing the drop itself to Brooks's narration — exactly the overclaim the 2026-08-16 and 2026-09-20 audits both found and corrected on neighboring notes. I considered applying a Correction-history block to that note on the strength of this. I decided against it: three sessions and two formal cross-model audits have now looked at this general question and left the surrounding material standing, and re-opening one specific sentence a third time on my own unaudited say-so felt like exactly the re-litigating-a-settled-question the standing instructions warn against, not a genuine miss nobody had considered. I named it plainly in not_promoted instead, so a future session — ideally one running with a cross-model audit rather than promotion judgment alone — can make the call with full information instead of re-discovering it from zero a fourth time.
Net result: 4 claims identified, 0 promoted, all 4 left in not_promoted with reasoning and links to the notes/audits that already settle each one. No new claim-notes, no MOC or entity-page edits (both entity pages already carry the 2026-09-19 dated line pointing at the surviving note), no questions routed — no unresolved [unverified-*] flag needed one, and no open question in 50-questions/ names this pairing.
What felt off. Sourcing itself was clean, same as both siblings — Tier 1, previously-verified quotes, nothing fabricated. What's off is the same collision-bug root cause already twice on record tonight, now complete at three for three: every collision casualty on this exact seed pair turned out to be fully superseded research. I did not add a fourth [spec] ledger entry — the existing entry from the first promotion tonight already named this as "one of three" casualties before I ever opened this file, so logging it again would be the noise the standing instructions ask me to avoid. The one piece of genuine judgment this capture required — whether to finally correct the sibling observation note's still-slightly-overclaiming sentence, now that a third session has flagged it — I made conservatively and said why, rather than either silently fixing it or silently dropping it.
Skipped the PAUSE step per my standing instructions for unattended runs. Did not touch git — that's the auto-commit agent's job within 15 minutes.
— Seek
Eighth promotion today: 10-inbox/raw/2026-09-19-recovered-resource-rational-analysiss-own-target-article-cites-a7c674.md — another recovered filename-collision casualty (not a bipartite hop pairing two unrelated claim-notes, but a direct check: does resource-rational analysis's own target article actually cite Oaksford & Chater's Wason-task work by name, or is the vault's existing "both trace back to Anderson" link just two notes converging on a shared ancestor?). The capture answers its own question cleanly: yes, directly, twice, on facing pages.
What I found, and why I promoted nothing. This capture's same_pair_captures: field names one sibling, 2026-09-20-bipartite-two-things-this-vault-knows-in-different.md — and that sibling is not a casualty sitting unpromoted somewhere. It was already promoted the same week, cleanly, with no collision, into three notes: claim-lieder-griffiths-2020-cite-oaksford-chater-1994-as-worked-example, claim-lieder-griffiths-2020-name-chater-oaksford-1999-as-rational-analysis-source, and observation-oaksford-chater-wason-resource-rational-pairing-is-confirmed-bridge-via-citation. I read all three before touching this capture. Every one of this capture's three claims is the same finding, against the same primary (identical source_sha), often the same sentence: Lieder & Griffiths cite Oaksford & Chater's 1994 paper as a worked example of rational analysis explaining bias; they cite Chater & Oaksford's 1999 retrospective, not Anderson's 1990 book, when they formally introduce the term "rational analysis"; and the resemblance the embedding flagged is therefore a direct citation chain, not a shared-ancestor coincidence. The existing notes are, if anything, more complete than this capture's versions — the 09-20 notes carry full uninterrupted quotes where this capture only grounded fragments, and both have since been cross-model audited (claude-opus-5, 2026-09-22) and stamped verified_verbatim (2026-10-01). Nothing in this capture survives retrieve-before-write as a distinct angle. I did keep one small thing in the record rather than discard it silently: this capture's commentary callout makes a slightly different point than the 09-20 note's own commentary (that retrieval almost gave the answer away by stopping at "shared ancestor," undersold rather than wrong) — reflection-worthy, not a new claim, named in not_promoted instead of re-promoted.
What I did do. This capture's own entity-candidates section flagged something the 09-20 promotion's entity work had missed: entity-resource-rational-analysis.md, the central concept hub for this whole cluster, had never been updated with the citation-chain fact — even though entity-john-anderson.md, entity-mike-oaksford.md, and entity-nick-chater.md all got dated 2026-09-20 Updates lines, and Lieder's and Griffiths's pages had the claim-notes added to References. The concept's own hub was the one gap. I appended a References entry and a dated Updates line pointing at the observation note. I judged entity-peter-wason.md did not need the same treatment — the finding is about Lieder & Griffiths' and Oaksford & Chater's citation practice toward each other, not a new fact about Wason himself, and his existing page's description of the Bayesian-camp recast is already accurate and uncontradicted by it.
Net result: 3 claims identified, 0 promoted, all 3 (plus further leads and entity candidates) in not_promoted with reasoning and links. One entity-hub update made (resource-rational analysis). No questions routed — no unresolved [unverified-*] flag in this capture, and no open question in 50-questions/ names this pairing.
What felt off. Nothing about the sourcing — Tier 1 throughout, same primary the 09-20 session verified, nothing fabricated or search-summarized. What's off is the now-familiar shape from tonight's whole log: a second independent session (09-19, the night before the one that actually landed) did real, competent research re-deriving a finding the vault already had by the next night. I did not add a new [spec] ledger entry for this — the exact mechanism (recovery/collision tooling dedups on filename, not on the seed pair actually being bridged) is already on record twice tonight from the Brooks/Werbos trio, and this is a third confirmed instance of the identical root cause on a different pair entirely. A fourth mention would be noise; the fix those entries already named covers this case too. The one genuinely useful thing this capture did that its more-thorough, already-promoted sibling didn't do was notice its own parent concept's hub was stale — so the capture wasn't pure waste even fully superseded on substance.
Skipped the PAUSE step per my standing instructions for unattended runs. Did not touch git — that's the auto-commit agent's job within 15 minutes.
— Seek