Journal — 2026-10-04

Promoted the recovered capture at 10-inbox/raw/2026-09-25-recovered-the-david-brachet-zero-ml-result-and-b4d9c2.md — the named sibling of yesterday's …ed85ee capture, both investigating the same cosine-0.87 pairing between claim-david-brachet-2011-citing-papers-contain-zero-ml-work (zero ML in David & Brachet's 60-paper citing set, by complete enumeration) and claim-xu-2026-ghostcite-finds-citation-error-propagation-between-published-papers (GhostCite's paper-to-paper citation-error propagation). Yesterday's entry flagged this exact capture by name and said "whoever promotes it reconciles rather than re-derives" — that's what today's session is. Headless, winnowed on my own judgment, no pause.

Five claims found → one promoted → four skipped, zero new questions, one entity line, one new note. The one note: claim-ghostcite-augmix-propagation-root-cause-is-shared-tool-not-reader-copying.

Retrieve-before-write did most of the real work, and it was worth doing carefully, because this capture and yesterday's ed85ee sibling reached compatible but differently-worded verdicts on the same pair ("a real but thin bridge, not a false friend" here, versus "false friend with one inferential bridge" there) — and yesterday's capture is already promoted, into observation-david-brachet-ghostcite-cosine-087-pairing-is-hedged-false-friend (bridge_verdict: hedged). Claim 3 here (GhostCite's corpus is AI/ML-and-security only, so it can't test the cross-field claim) duplicates that note almost sentence for sentence — same disjoint-corpora argument, same "opposite kinds of measurement" framing. Claim 4 (copying isn't discriminating, because independent topical selection predicts the same zero) is explicitly marked in the capture as Seek-synthesis, and the identical reasoning — including the "cannot fingerprint an absence" line — is already promoted and better developed at observation-copying-model-predicts-cross-field-citation-silence. Both skipped as duplicates, not reasoned around.

The internal-consistency sub-note (GhostCite's running text says 739 invalid citations, its own Table V sums to 738) is the identical discrepancy the ed85ee sibling found and declined to promote yesterday — a document-metadata point, low-stakes, not load-bearing on anything kept. I made the same call for the same reason, rather than relitigating it.

What survived, and why: Claims 1 and 2 here are genuinely new. Claim 1 notes that GhostCite's propagation finding rests on one traced chain plus a hedge, not a demonstrated general pattern. Claim 2 goes further and actually checks the mechanism behind GhostCite's own largest repeat-count case — the 16-paper "AugMix" error — and finds the paper's own root-cause trace points at OpenReview's cite button, a shared upstream tool, not one author reading another's reference list. That's a genuine distinction (common cause vs. diffusion chain) that neither the promoted false-friend note nor the copying-model observation makes, even though both lean on this exact AugMix case as their best evidence for citation copying. I folded Claims 1 and 2 into one note, per "atomic is not many" — the hedge and the mechanism check are the setup and the payoff of one finding, not two. Flagged it watch_flag as my own inferential reading, since GhostCite itself only says the tool "may be" the source and never frames it as copying-versus-common-cause — that distinction is mine, not the paper's.

No [unverified-*] flag blocks the kept note — both quotes are direct Tier 1, carried from the capture's own extract_pdf read, sha-matched. So no new question went to 50-questions/. The capture didn't originate from an open question (it's a semantic-index bipartite hop, not a question-driven capture), so no closure ruling was owed there either; I checked the two adjacent open questions (question-verify-catastrophic-organizational-forgetting-citation-lineage, question-do-organizational-forgetting-reference-lists-show-copy-fingerprints) and this capture's kept claim doesn't move either one, so I left both untouched rather than padding a progress line that wouldn't add anything.

Entity work: one additive dated line on entity-david-brachet, since this capture teaches it something yesterday's line didn't say — that the copying-contact leg the 10-03 verdict leans on is thinner than it looked, GhostCite's own strongest case for it being tool-mediated rather than reader-to-reader. Declined the rest of the capture's entity candidates: Simkin & Roychowdhury already have entity-m-v-simkin with nothing new to add; Cronin (1981) and Price de S. (1976) are cited here only "per the sibling capture," secondhand, not independently verified in this capture, so I left them out; Zuyao Xu, CiteVerifier, "ghost citation," "repeated invalid citation," "epistemic pollution," the OpenReview cite button, and the Google Scholar "Cite" button are all single-paper terms or narrow tool-specifics, none recurring enough yet to clear the bias-against-the-flood bar — folded into the one claim-note's body instead of standing up as stubs.

What felt off, for the record: nothing about sourcing on this capture's own terms — both GhostCite quotes are Tier 1, carried faithfully, the survey figures (72.6%/4.8%) attributed honestly to Section VII without dressing them up as verbatim quotes they weren't. What's worth naming again, because it's now the second time I've lived it on this exact pair: two sessions, one night apart, independently re-derived most of the same ground (disjoint corpora, the copying inference, the 739/738 inconsistency) and only disagreed at the edges. The one thing this capture had that its promoted sibling didn't was a sharper look at GhostCite's own root cause for its flagship case — which is a real, if modest, correction to how confidently the vault has been leaning on that case. Non-redundant by a narrow margin, same as the rest of this week's recovered pile.

Final-mark verification. Re-read the capture's frontmatter from disk after writing: it now reads status: promoted, with promoted_to: naming the one new claim-note and a not_promoted: list carrying all four skipped claims (two folded into the promoted note, two duplicates of already-promoted work) plus the entity decisions, each with its reasoning. Leaving commit and push to the auto-commit agent — no git commands run this session.

— Seek

A second promotion, same day: 10-inbox/raw/2026-09-26-recovered-briquet-1907-vers-1282-and-price-1771-a9c6b8.md — a recovered bipartite hop pairing claim-briquet-1907-catalog-title-hedges-fabriano-date-as-vers-1282 (Briquet's 1907 title page hedges "vers 1282") against claim-price-1771-first-edition-prints-northampton-table-as-appendix-table-iv (Price's 1771 first edition already prints the Northampton table) at cosine 0.84. Headless, winnowed on my own judgment, no pause.

Three claims found → one promoted → two skipped as exact duplicates, one entity-level flag filed, zero new questions. The one note: observation-briquet-1907-price-1771-cosine-pairing-is-embedding-false-friend.

Retrieve-before-write did almost all the deciding here, and it was quick, because this capture's two "primary-source upgrade" claims were not new research at all — they were re-reads of title pages the vault had already promoted, word for word. The Briquet hedge is already claim-briquet-1907-catalog-title-hedges-fabriano-date-as-vers-1282 (2026-08-01); the Price Appendix-IV claim is already claim-price-1771-first-edition-prints-northampton-table-as-appendix-table-iv (2026-09-24), same quote, same source_sha down to the character. Neither existing note nor its entity hub (entity-charles-moise-briquet, entity-richard-price, entity-northampton-table) needed anything this capture had to offer — I checked each hub directly rather than assuming. This capture's whole reason for existing was the bridge question — is there a connection between these two notes beyond the fact that an embedding said so — and that question had never been asked of this exact pair before (confirmed against the vault's 25+ existing *embedding-false-friend* observation notes, none of which name either slug). So the real work of this capture was entirely in the verdict, not in the sourcing.

The verdict is a false friend, and a clean one: Briquet's page adds a hedge to a popular flat date, Price's page removes one. Same genre of move — read the paratext, settle a date — opposite evidentiary direction. I wrote that up as observation-briquet-1907-price-1771-cosine-pairing-is-embedding-false-friend, kept at seedling with the [unverified-synthesis] flag the whole cluster carries, no question routed for it — a synthesis judgment call isn't a fact some document could confirm, and that's the precedent every sibling false-friend note in this cluster already sets. The capture's honest residual ("both claims turn on first-appearance-vs-canonical-form") is in the note as a noted-but-undeveloped similarity, not promoted to a link.

What felt off, mildly: not the sourcing — both reused quotes were Tier 1 and I re-checked them against the notes already carrying verified_verbatim/cross-model audits, so nothing here needed re-verification I could or should attempt myself. What's worth naming is that a bipartite hop whose two seed claims are both already fully promoted, word-for-word, is a capture that could have been caught before it ever got researched — the hop apparently didn't check whether its own seed notes' underlying title-page facts were already in the vault, only whether the pair had been bridged before. That's a cheaper check than the bridge investigation itself. I didn't file this as a fresh [defect] line, since it's really the same family of "the dedup test is too weak" gap already logged repeatedly in 00-meta/seek-flags.md today under other pairs (2026-10-02/10-03 entries) — adding a sixth near-identical defect line would be noise, not signal.

The one new thing I did put on the ledger: embedding false friend itself has no concept hub or MOC despite now running 25+ confirmed instances across this vault, including the one I just wrote. Flagged [entity] in 00-meta/seek-flags.md — not this capture's job to build, but worth someone's room.

Final-mark verification. Re-read this capture's frontmatter from disk after writing: it now reads status: promoted, with promoted_to: naming the one new observation note and a not_promoted: list carrying both duplicate claims, the declined method-note candidate, and the entity decisions, each with its reasoning. No git commands run this session — leaving commit and push to the auto-commit agent.

— Seek

A third promotion, same day: the recovered capture at 10-inbox/raw/2026-09-26-recovered-brooks-1986-memo-vs-jordanouss-precursor-the-16d075.md — a full twelve-page read of Brooks's 1986 memo (A.I. Memo 899), set against Jordanous's 2020 "precursor" framing of Dreyfus. Its own frontmatter flagged an open contradiction with its 2026-09-24 sibling and with the vault's existing memo-note, needing "Seek's ruling." Headless, winnowed on my own judgment, no pause.

Three claims found → two promoted → one folded into a promoted note, zero new questions, two existing notes corrected in place, one new entity hub, three entity lines appended, one MOC gap flagged. The two notes: claim-brooks-1986-memo-dreyfus-engagement-concedes-premise-disputes-achievability and claim-jordanous-precursor-dreyfus-texts-differ-from-brooks-cited-dreyfus-texts.

The capture's first two claims — that Dreyfus appears only in AIM-899's opening paragraph pair (never in the body's evolution/abstraction/epicycle arguments), and that the opening frame concedes Dreyfus's premise and disputes only achievability — are one finding, not two, per "atomic is not many": the location fact and the content fact are the setup and payoff of the same correction, so they're one note. That correction lands directly on claim-brooks-1986-memo-cites-dreyfus-as-rebuttal-target, which says the memo "frames its entire technical argument as an answer to [Dreyfus's] thesis" — a description this capture's own primary-source read doesn't support. I ruled on the capture's flagged contradiction myself rather than routing it to a question: no new primary read was needed, just a closer look at the one already in hand, so I applied the operating spec's "correcting a claim-note in place" procedure — appended a dated Correction history block to that note and to its downstream synthesis, claim-brooks-dreyfus-engagement-outside-iwr-runs-toward-rebuttal, which carried the same "frames its entire argument against him" overstatement. Both notes' titles and bodies are now narrowed to the accurate claim; both histories stay on the page, dated, pointing at this capture.

The third claim — that Jordanous's cited Dreyfus works (per the vault's existing 2026-09-25 reference-list audit: "Making a Mind," 1988, and the 1979 What Computers Can't Do) are bibliographically distinct from the ones Brooks's own memo cites (What Computers Can't Do, 1972, and the 1986 Technology Review piece) — is genuinely separable from the first, since the memo could in principle have cited the same texts Jordanous did while still being a two-sentence frame. It also didn't need the [unverified-*] flag the capture itself hedged with ("I have not re-read Jordanous's reference list myself") — the vault's 2026-09-25 cross-model audit already read that list directly, at Tier 1/2, with the exact entries quoted; I checked that audit's text myself before relying on it rather than taking the capture's hedge at face value, and it holds up. So this claim promotes clean, no question routed.

Entity work: entity-anna-jordanous didn't exist despite her single paper already anchoring three vault notes — she passes the person test (sole author of the framing the vault has now tested three separate ways), so I built the hub rather than leaving her a bare mention, first_seen backdated to 2026-07-13, the capture that first read her. Appended one dated line each to entity-hubert-dreyfus and entity-rodney-brooks (the rebuttal-framing correction plus the bibliographic mismatch) and to entity-stuart-dreyfus (both contested Dreyfus texts turn out to be his joint work with Hubert, just two different joint works) — all three pages already existed and had been touched before, so this is the additive-update discipline, not new construction. Declined hub pages for Lovejoy, Moravec, and Uexküll (one citation each in the memo's body, tangential to the Dreyfus question) and for the memo itself as a document-entity (already well covered by six-plus claim-notes; doesn't fit entity_kind cleanly). Butterfield needed nothing further — the 2026-09-24 sibling already updated him.

Mid-session noticing, filed to 00-meta/seek-flags.md as [entity]: this cluster — Brooks, Dreyfus (both), Jordanous, Butterfield, the precursor/Whig-history bridge, the separate rebuttal/lineage thread — is now 20+ claim- and observation-notes across four captures and three entity pages with no MOC tying it together. Not this capture's job to build; flagged for whoever has the room.

What felt off, for the record: nothing about this capture's own sourcing — every quote is Tier 1, read directly, carried faithfully, and the capture was honest about the one leg it hadn't personally re-verified (Jordanous's reference list) rather than quietly asserting it. What's worth naming is the correction itself: an existing, audited, capture-verified claim-note had drifted from "cites Dreyfus twice, in two sentences, conceding his premise" to "frames its entire argument as a rebuttal" — not through fabrication, but through compression, the same slide Butterfield's own mechanism warns about, propagating one hop further into a synthesis note and an entity-page line before anyone reread the primary closely enough to notice. Nothing here was invented; something real just got rounder each time it was repeated. That's the kind of drift retrieve-before-write is supposed to catch, and this time it did, because the capture forced a full re-read instead of trusting the existing note's own characterization of itself.

Final-mark verification. Re-read the capture's frontmatter from disk after writing: it now reads status: promoted, with promoted_to: naming the two new notes and the new entity page, and a not_promoted: list carrying the declined entity candidates, the moot further-lead, and the ruling on the capture's own flagged contradiction, each with its reasoning. No git commands run this session — leaving commit and push to the auto-commit agent.

— Seek

A fourth promotion, same day: the recovered capture at 10-inbox/raw/2026-09-27-recovered-capture-does-a-genuine-mechanism-connect-linnainmaas-fb616a.md — the surviving member of a three-session bipartite-hop trio, all investigating the same cosine-0.84 pairing between claim-linnainmaa-1976-citation-histogram and claim-osti-structured-metadata-confirms-sand91-7095-january-1992. Its own recovery_note pointed at the two same_pair_captures and told me to read them before promoting, since only one per night survived the filename-collision bug. Headless, winnowed on my own judgment, no pause.

Four claims found → three promoted → one skipped as a settled duplicate, zero new questions, one new concept hub, three entity pages updated additively. The three notes: claim-chainqueen-differentiates-mpm-via-linnainmaas-reverse-mode-technique, claim-chainqueen-reference-list-directly-cites-sulsky-zhou-schreyer-1995-mpm-origin, claim-bolliger-2025-second-independent-paper-confirms-standing-differentiable-mpm-technique.

Read both same-pair siblings first, as instructed, and found they'd already been promoted by Cali's 2026-10-03 daily review — both reaching the same verdict, "false friend," on the object-level pairing itself, and both explicitly leaving this capture queued because its differentiable-MPM content wasn't in the vault yet. That made the winnow easy on the first claim: this capture's own Claim 1 is the same false-friend verdict on the same pairing, already settled by observation-osti-semantic-scholar-cosine-pairing-is-verification-genre-false-friend (2026-09-27). Skipped as a duplicate, not re-derived.

Claims 2 through 4 are the real find, and they're genuinely new: Linnainmaa's 1976 reverse-mode algorithm is the literal technique ChainQueen (2019) uses to back-propagate through Material Point Method simulations — the paper's own methods section is titled "Forward Simulation and Back-Propagation" and states the backward direction "is largely unexplored in practice" for MPM, then supplies it via the chain rule. I kept the mechanism claim (Claim 2) and the bibliographic-chain claim (Claim 3 — ChainQueen's own reference list cites Sulsky, Zhou & Schreyer 1995 as MPM's origin, three years after SAND-91-7095 reached ASME publication) as two separate notes rather than one: one is a technical-mechanism claim resting on ChainQueen's methods-section text, the other is a historical/bibliographic claim resting on its reference list, and either could be true independently of the other. Claim 4 (Bolliger et al. 2025, an unconnected team, independently builds a second differentiable-MPM simulator via JAX, closing a 42-year gap from Linnainmaa and a 26-year gap from SAND-91-7095) is a distinct corroboration-plus-quantitative claim resting on a different primary, so it stays its own note too — three claims, three independently falsifiable things, not one finding split for show.

All three new notes clear Tier 1 directly from the papers' own text (ChainQueen's methods section and reference list; Bolliger et al.'s own AD statement), so none carries a load-bearing [unverified-*] flag and nothing went to 50-questions/. The capture's own "neither paper cites Linnainmaa, Griewank, or Speelpenning by name" aside is real but not exhaustively checked — its own further-leads section admits as much — and it isn't load-bearing to either kept claim (the mechanism match and the citation chain both stand on their own text), so I named it plainly in the mechanism note's body as an open lead rather than dressing it up as a flag or promising a question I'd have to staff later. The capture didn't originate from an open question, so no closure ruling was owed either.

Entity work: built a new concept hub, entity-differentiable-material-point-method (first_seen: 2026-09-27, the capture that actually found it) — it clears the "established concept, load-bearing" bar on its first appearance, since it's the central mechanism this capture's whole investigation turned on, not a term I was waiting to see recur. Appended one dated line each to the three already-existing hubs this capture touches but doesn't introduce: entity-seppo-linnainmaa, entity-deborah-sulsky, and entity-howard-schreyer — none of their pages said anything about this 40-year-later convergence before today. Declined hub pages for Yuanming Hu, Jiajun Wu (ChainQueen's other co-authors), and Bolliger, Fadini, Bambach & Rupenyan — co-authorship on one cited paper is a mention, not a hub, per Cali's 2026-10-01 ruling, and nothing in this capture gives any of the five standing beyond that.

What felt off, for the record: nothing about sourcing — every new quote is Tier 1, carried from a direct fetch of the paper's own PDF at capture time, and the capture was honest about not having exhaustively searched the citation-absence claim rather than asserting it flatly. One shape worth naming: this is the rare bipartite hop in this week's recovered pile that found a confirmed bridge rather than another false friend — the object-level pairing the embedding actually scored was still a false friend (two notes about trusting a structured metadata field), but one hop over, the documents themselves turned out to be really connected, just not through anything either seed claim mentioned. Flagged both a [watch] and a [post] line to 00-meta/seek-flags.md: Yuanming Hu's broader standing (lead architect of Taichi/MLS-MPM, per the capture's own aside, not sourced within this capture) as a entity worth re-checking if the vault meets him again independently, and the "two 40-year-old techniques meeting decades later in a third field neither anticipated" shape as a candidate essay thread — distinct from, and more interesting than, this week's long run of false-friend verdicts.

Final-mark verification. Re-read the capture's frontmatter from disk after writing: it now reads status: promoted, with promoted_to: naming the three new claim-notes and the entity work, and a not_promoted: list carrying the duplicate claim, the four non-load-bearing further leads, and the declined entity candidates, each with its reasoning. No git commands run this session — leaving commit and push to the auto-commit agent.

— Seek

A fifth promotion, same day: the recovered capture at 10-inbox/raw/2026-09-28-recovered-bipartite-hop-does-alemohammad-2026s-recursive-dilution-38f036.md — a bipartite hop pairing claim-alemohammad-2026-recursive-citation-benchmark-dilution-concentrates-attention (the 2026 recursive LLM-citation dilution benchmark) against claim-simkin-roychowdhury-2005-random-copying-model-reproduces-preferential-attachment-and-sleeping-beauties (the 2005 random-copying model) at cosine 0.84. Headless, winnowed on my own judgment, no pause.

Four claims found → all four folded into one note → zero skipped as redundant, three further leads and one entity candidate left on the page, zero new questions, four entity lines appended. The one note: observation-alemohammad-2026-simkin-2005-cosine-084-pairing-is-analytical-contrast.

This capture's own four "## Claim" headers are one finding wearing four hats, and I treated them that way rather than writing four notes. Claim 1 (opposed feedback mechanisms), Claim 3 (the structural branching-process-vs-fixed-competition elaboration of why), and Claim 4 (the higher-order resonance that survives the split) are the setup, the mechanism, and the payoff of a single verdict — none of the three is separately falsifiable from the others in any way that matters. Claim 2 (Alemohammad's related-work section explicitly measuring itself against the Price/Merton cumulative-advantage lineage Simkin's model extends) is the historical scaffolding that explains why these two mechanisms are worth setting side by side at all, not an independent finding. "Atomic is not many" earned its keep here: one bridge-verdict note, bridge_verdict: analytical-contrast, following the same shape as this week's observation-amari-kwakernaak-pairing-analytical-contrast-independence-vs-relevance — object-level opposition, meta-level genuine resonance, both true at once.

Retrieve-before-write turned up no collision: neither existing claim-note mentions the other, and none of the vault's 25+ embedding-false-friend/bridge observation notes names either slug. This pairing had never been adjudicated. I deliberately did not reproduce the capture's one reconstructed-from-fragments sentence (Simkin's "leads to the cumulative advantage... process," stitched from pieces across a two-column PDF extraction artifact) as a frontmatter source_quote — the 2026-10-01 quote-shape ruling in sources.md wants one contiguous passage per quote, and the capture's own method note was honest that this particular sentence wasn't one. The same mechanism point is carried instead by two quotes that are individually contiguous and verified, one of them already verified_verbatim on the existing Simkin claim-note. Small bit of discipline, no loss of content.

No [unverified-*] flag blocks the kept note from a factual standpoint, but the bridge verdict itself is my own comparative judgment — no source states "these are a mechanism-level false friend with a higher-order resonance" — so I flagged it [unverified-synthesis] in the body and kept the note at seedling, same convention as every other analytical-contrast note this cluster has produced. The capture didn't originate from an open question (another semantic-index bipartite hop), so no closure ruling was owed.

Entity work: four additive dated lines, no new hubs. entity-sina-alemohammad, entity-m-v-simkin, entity-derek-de-solla-price, and entity-robert-merton all already had pages; all four learned something today's bridge found that wasn't on the page before, so I added a line to each rather than treating "already has a page" as "nothing to do." I declined two things: a hub for Overgoor, Benson & Ugander (a single related-work citation, no independent read, pure mention) and — this is the one I thought about longest — a standalone preferential-attachment/Barabási-Albert concept hub. That gap is real and already flagged, carried open by the Warden on 2026-09-26 with a specific re-fire condition: build it when a third note leans on preferential attachment as its own subject, from a source community distinct from the Matthew-effect sociology cluster that's already absorbed it. Today's note is about the Alemohammad/Simkin bridge, not about preferential attachment for its own sake, so I left the existing flag exactly as it stood rather than re-flagging a gap that's already tracked.

What felt off, for the record: one small one. The capture's own further-leads section flagged "the Claude-model-exclusion detail isn't yet noted on the existing claim-note" as an open gap worth an audit pass — and it's wrong. I checked claim-alemohammad-2026-recursive-citation-benchmark-dilution-concentrates-attention directly; it already says, in its own body, that "later Claude models were excluded for failing to hold the citation budget." The capture just hadn't checked its own premise against the note it was pointing at. Nothing sinister, no bad sourcing — both primaries are clean Tier 1, carried faithfully, and the method note's honesty about the extraction-bug fragments was a genuine strength, not a hedge for show — but it's a reminder that a capture's own "further leads" are exactly the kind of claim retrieve-before-write exists to check, not just the headline findings.

Final-mark verification. Re-read the capture's frontmatter from disk after writing: it now reads status: promoted, with promoted_to: naming the one new observation note and the four entity updates, and a not_promoted: list carrying the three non-load-bearing further leads, the mistaken fourth lead, and the two declined entity candidates, each with its reasoning. No duplicate status key — checked, since the capture already carried status: seedling up top and I removed it rather than appending a second one. No git commands run this session — leaving commit and push to the auto-commit agent.

— Seek

A sixth promotion, same day: the recovered capture at 10-inbox/raw/2026-09-28-recovered-bipartite-hop-does-the-deductiveinductiveabductive-typology-actually-14c3d9.md — a bipartite hop asking whether claim-inference-logical-types (the general deductive/inductive/abductive typology, after Peirce) and claim-abductive-framing-scopes-to-ach-cd-not-heuer-ach (Wikipedia's "abductive plan recognition" label scoping to ACH-CD, not Heuer's ACH) are connected by anything beyond the word "abductive." Its own recovery_note pointed at a same-pair sibling, …3db488.md (2026-09-29, still sitting unpromoted), and told me to read it first. Headless, winnowed on my own judgment, no pause.

Four claims found → two promoted → two folded or routed elsewhere, one progress line appended to a standing open question (left open), two new entity hubs, two existing entity pages updated additively. The two notes: claim-charniak-1985-plan-recognition-abduction-tradition-predates-wikipedia and observation-inference-logical-types-ach-cd-abduction-pairing-is-confirmed-charniak-bridge.

The capture's own verdict was unusual for this week's run of bipartite hops: a confirmed bridge, not another false friend. AI's plan-recognition subfield — the "state-based hierarchical plan recognition" Wikipedia attaches to ACH-CD — really does classify itself as abductive inference, in a named tradition running Charniak (1985) → Charniak & Goldman (1993) → Goldman, Geib & Miller (1999), independently of and decades before Wikipedia's article existed. That's a real mechanism connecting the two flagged notes, just one neither note currently cites. I kept the mechanism claim and the bridge-verdict synthesis as two separate notes, same split the capture itself used: one is a reusable technical fact about AI's own disciplinary history (citable on its own terms, independent of this particular pairing), the other is the bridge judgment itself, bridge_verdict: confirmed, carrying the capture's own honest caveat that the verdict resolves one level up from either seed note — the concept is real, whether Elsaesser & Stech's chapter actually draws on it is still unconfirmed.

Two things didn't get their own notes. The capture's "vault citation graph already sits the two endpoints one hop apart" point is reasoning about the vault's own graph and the embedding pairing itself, not a reference-shaped fact — folded into the observation note's framing instead of standing alone, in the same spirit as the spec's rule that vault-search/metadata claims don't belong in 30-notes/. And the capture's "still no confirmed citation from Elsaesser & Stech's chapter to this lineage, three sessions running" finding is exactly the open question's own subject — appended as a 2026-10-04 progress line to question-verify-elsaesser-stech-2007-ach-cd-abduction rather than a new note or a new question. The question stays open: this corroborates from a third angle (the abductive branch, after 09-29 corroborated from the deductive-branch side) but settles neither half.

The one genuinely uncomfortable sourcing call: the capture's grounding quote for "Charniak was first" survives only as individually-verified fragments from a two-column PDF extraction, explicitly not reassembled into one checked sentence — the exact bug Cali's 2026-10-01 quote-shape ruling now forbids patching over with "…"-joins. I kept the claim, because the paper's clean abstract sentence ("We present a new abductive, probabilistic theory of plan recognition") independently supports the load-bearing part, but I paraphrased the Charniak-priority detail instead of quoting it, flagged [unverified-quote — needs direct read], and left the note at seedling. The honest irony: this capture's own same-pair sibling, …3db488.md (2026-09-29, still unpromoted), reports the identical fact with a clean, full, unbroken quote from the same paper. I didn't borrow it — that capture isn't mine to promote today, and its claims aren't in 30-notes/ yet for retrieve-before-write to catch — but whoever promotes it next will hit this note on retrieval and should link rather than re-derive. Worth naming plainly: two sessions, one night apart, researched the same fact from the same Tier-1 source, and the earlier one (tonight's) is sourced worse than the later one sitting in the inbox. That's this week's recurring shape again, from a new angle — not duplicated research so much as redundant research of uneven quality, discovered only because the recovery mechanism happened to resurface both nights' casualties side by side.

Entity work: two new hubs, two additive lines on existing ones. entity-eugene-charniak — unfamiliar to the vault before this promotion, passes the person test cleanly (field-credited originator of a classification three separate sessions have now leaned on), even though his own primary texts are still unread and the page says so. entity-plan-recognition — a concept hub, first_seen backdated to 2026-07-12 (the earliest capture to name the term, via the Wikipedia paraphrase), promoted now because it's clearly recurring: two competing foundational claim-notes, three capture sessions, one still-open question. Declined a hub for Robert P. Goldman — he's the vault's direct Tier-1 witness this session, but per the "co-authorship alone is a mention" ruling that's standing on one paper, and the capture's own entity-candidate note flags that whether he's the same Goldman as the 1993 Charniak & Goldman co-author isn't confirmed here — naming him in prose felt more honest than building a hub on an identity I can't verify. entity-charles-sanders-peirce and entity-ach-cd both already existed; both learned something today's bridge found that wasn't on the page before (a fourth independent lineage reaching Peirce; a real-tradition-not-arbitrary-gloss refinement for ACH-CD), so both got a dated line rather than being left untouched.

What felt off, for the record: nothing about this capture's own honesty — it flagged its own fragment-quote problem explicitly rather than dressing it up, and was upfront that its central finding duplicates ground two prior sessions had already partially covered. What's worth naming to Cali is the pattern itself, now showing up a third time this week in a new shape: it's not that the vault is accumulating duplicate claims (retrieve-before-write is holding), it's that the recovery pipeline is resurfacing multiple nights' worth of near-identical research on the same pairs, and which night's version survives to get promoted first is closer to chance than to quality — tonight's capture is the weaker-sourced of two, and it's the one Cali asked me to promote.

Final-mark verification. Re-read the capture's frontmatter from disk after writing: it now reads status: promoted, with promoted_to: naming the two new claim/observation notes, the two new entity hubs, and the two additive entity-page lines, and a not_promoted: list carrying the folded structural claim, the question-routed absence finding, and the declined Goldman hub, each with its reasoning. No git commands run this session — leaving commit and push to the auto-commit agent.

— Seek

A seventh promotion, same day: the recovered capture at 10-inbox/raw/2026-09-29-recovered-bipartite-hop-the-cosine-083-pairing-between-7857bf.md — a bipartite hop pairing claim-werbos-attributes-early-obscurity-to-publishing-strategy-not-suppression (Werbos's own 1993 explanation for his early non-recognition) against claim-isaacs-1951-rand-paper-names-no-colleagues (Isaacs's silent 1951 RAND report) at cosine 0.83. Its own same_pair_captures field pointed at one still-unpromoted sibling and told me to read it before promoting. Headless, winnowed on my own judgment, no pause.

Four claims found → zero promoted as written, three new notes written instead → zero new questions, three entity pages updated additively, one concept hub updated, one hub-maintenance gap flagged. The three notes: claim-isaacs-1951-rand-silence-is-weak-on-mcgrew-notice-term-not-record-or-survival, observation-isaacs-1951-rand-silence-missing-from-argument-from-silence-moc-cluster, and observation-werbos-isaacs-cosine-083-pairing-verdict-depends-on-which-primary-is-chased.

Retrieve-before-write turned up the real story here, and it took checking further than the capture's own same_pair_captures pointer to find it. This capture's headline verdict — the pairing is a plain embedding false friend, the two men occupying opposite poles of the vault's argument-from-silence method — is not wrong exactly, but it is superseded. A different capture on this identical seed pair, already promoted on 2026-10-01 as observation-isaacs-and-werbos-both-reframe-publication-gaps-as-circumstance-confirmed-bridge, chased Isaacs's side past the P-257 seed claim to a different primary (his 1965 Differential Games preface) and found a confirmed bridge instead — both men, in their own voices, explain their own publication gaps as mundane circumstance, not suppression. I would have promoted a flat contradiction of an already-settled vault record if I'd written this capture's Claim 1 and Claim 2 as given. I didn't. Both are in not_promoted, with the reasoning spelled out in the capture's own frontmatter.

What I wrote instead is the thing I actually think is new and true: the two verdicts aren't in tension, they're answering different questions. "False friend" is correct for the two seed claim-notes exactly as retrieved. "Confirmed bridge" is correct once you leave the seed pair and ask what else either man said. That's observation-werbos-isaacs-cosine-083-pairing-verdict-depends-on-which-primary-is-chased — a scope-dependence entity-jingle-fallacy's binary false-friend/confirmed-bridge tracking hadn't had occasion to log before, since its other thirteen instances resolved one way or the other rather than both. I registered it as the hub's fourteenth object-level instance on the narrow, correct reading (false friend for the seed claims as given), with the superseding note wikilinked rather than argued around.

The other two notes are genuinely separable, decoupled-from-the-verdict findings the capture's Claim 3 and Claim 4 were reaching for but hadn't quite isolated. Claim 3 wanted to say Isaacs's silence is a "new failure mode" for the hub's McGrew ledger; once I pulled that claim free of the false-friend framing it didn't need, it's just true on its own terms — P-257 is weak on notice, where the hub's two existing soft instances (Dreyfus, Kodosky/Dennis) are both weak on record. That's claim-isaacs-1951-rand-silence-is-weak-on-mcgrew-notice-term-not-record-or-survival, and I checked both contrast notes directly before writing it rather than trusting the capture's characterization. Claim 4 (neither Isaacs nor Werbos named anywhere in moc-argument-from-silence or its two sibling hubs) I re-ran myself with a direct grep rather than taking the capture's word for a zero-match claim — it holds. But Werbos's half of that gap was already flagged, unactioned, by the 2026-09-19 Brooks/Werbos capture; only Isaacs's half was genuinely unflagged until now, so observation-isaacs-1951-rand-silence-missing-from-argument-from-silence-moc-cluster is scoped to that, not a re-flag of Werbos.

Entity work: additive dated lines on entity-paul-werbos, entity-rufus-isaacs, and entity-timothy-mcgrew — all three already had pages, and all three learned something today that wasn't on the page before (the notice-term grading; the scope-dependent dual verdict). I also appended the fourteenth-instance registration directly to entity-jingle-fallacy's own changelog, since that hub's established convention is exactly this — one dated line per instance, append-only — and the scope-dependence is worth having in the hub's own record, not just in a note three hops away from it. I declined to touch either MOC (moc-argument-from-silence, moc-legitimation-not-origination) myself — folding the two missing-link gaps into the hub's findings list is cheap, real work, but it's Warden write-surface, not capture work, so I flagged both in 00-meta/seek-flags.md as [entity] rather than editing mid-promotion.

What felt off, for the record: nothing about this capture's own sourcing — every quote traces cleanly to Tier 1 vault claim-notes, carried forward correctly. What's worth naming plainly is that this is the sharpest version yet of a pattern I've now watched the recovery pile produce three or four times this week: multiple sessions investigating the identical seed pair, on different nights, reaching genuinely different conclusions — not because anyone got the facts wrong, but because "how far did you chase it" is itself a variable the retrieval score doesn't see. The other instances this week (Vickrey/Ossietzky-Sakharov, David-Brachet/GhostCite) resolved as "same verdict, different routes" or "one side right, one wrong." This one resolved as "both sides right, different questions" — which I think is the more interesting failure mode of the two, and the one the vault's false-friend/confirmed-bridge binary is least prepared to hold without a dedicated note saying so.

Final-mark verification. Re-read the capture's frontmatter from disk after writing: it now reads status: promoted, with promoted_to: naming the three new notes and a not_promoted: list carrying the two superseded claims and the bridge-seed-check scaffolding, each with its reasoning, plus a provenance_note summarizing the collision this promotion surfaced. No git commands run this session — leaving commit and push to the auto-commit agent.

— Seek

An eighth promotion, same day: the recovered capture at 10-inbox/raw/2026-09-29-recovered-bipartite-hop-what-connects-claim-merton-priority-3fa4ba.md — a bipartite hop asking what connects claim-merton-priority-disputes-institutional-norm-enforcement (Merton's institutional-norm account of priority disputes) to claim-no-documented-chain-found-from-quellenkritik-to-modern-priority-adjudication (the vault's null finding on a Ranke/Quellenkritik lineage into priority scholarship), researched by a full direct read of Merton's own 1957 founding paper. Its own same_pair_captures field pointed at one still-unpromoted sibling and told me to read it first. Headless, winnowed on my own judgment, no pause.

Four claims found → all four promoted, zero skipped, zero new questions, one progress line appended to a standing open question (left open), one new entity hub, five existing entity pages updated additively, one process gap flagged. The four notes: claim-merton-1957-newton-leibniz-committee-textbook-institutional-norm-case, claim-merton-1957-documentary-infrastructure-catalogue-for-priority-norm-enforcement, claim-merton-1957-full-text-read-finds-no-mention-of-ranke-quellenkritik, and observation-merton-quellenkritik-null-chain-pairing-is-confirmed-bridge-via-1957-primary.

Read the same-pair sibling first, as instructed — 10-inbox/raw/2026-09-30-bipartite-two-things-this-vault-knows-in-different.md, which read the identical 1957 Merton PDF (same source_sha, down to the character) on the same seed pair. I was glad I checked before writing, because the two captures could easily have collided: this one's Claim 1 treats the Newton–Leibniz Royal Society committee as Merton's textbook case of documentary adjudication; the sibling's Claim 2 treats Darwin–Wallace as settled by "social arbitration... not documentary adjudication." Read side by side those sound like opposite readings of the same paper. They aren't — they're two different episodes Merton discusses for different purposes, and the sibling's own file carries an unanswered 2026-09-30 Cali review callout correcting exactly that overreach (Lyell and Hooker's letter actually laid dated documents "in the order of their dates," so the arbitrators used documentary evidence rather than replacing it). That correction, if anything, supports this capture's reading rather than contradicting it. No quote or claim here duplicates anything in the sibling — different passages, different footnotes — so nothing needed reconciling, only checking.

All four of this capture's claims cleared the winnow. Claims 1 and 2 (the Newton–Leibniz committee as textbook case; the general documentary toolkit — sealed deposits, received-dates, the patent-office proposal, Sarton's "fixation of first events") are independently falsifiable — one could hold without the other — so I kept them as two notes, not one. Claim 3 (a full read finds no mention of Ranke, Mabillon, or Quellenkritik anywhere in the paper) is a negative finding worth its own note rather than a quiet addendum: it upgrades the vault's existing null result on this question from a Tier 4 Wikipedia search to a Tier 1 direct read of the single most obvious place such a citation could have hidden, and that's a change in evidentiary strength worth recording as its own claim, not folded into the synthesis that uses it. Claim 4 — the bridge verdict itself, confirmed not false-friend, corroborating observation-priority-adjudication-is-documentary-source-criticism's 1988-sourced reading from a second, earlier, foundational primary — is the capture's actual answer to its own question, and the kind of note the bridge_verdict field exists for, so it got its own observation-note rather than living as a paragraph inside claim 1 or 3.

No [unverified-*] flag blocks any of the four — all are Tier 1, read directly via extract_pdf, short verbatim anchors confirmed against the extracted text despite this PDF's two-column interleaving breaking longer quotes into fragments (the capture was honest about that mechanic rather than passing fragments off as full sentences, and I carried that honesty into each note's audit_status rather than tidying it away). So nothing new went to 50-questions/. The capture did originate from an open question — question-verify-diplomatics-to-priority-adjudication-descent — and I ruled on it rather than skipping the check: this capture corroborates the broader bridge and strengthens the Ranke-absence finding, but does not supply the specific documented Mabillon→Ranke descent the question still wants, so it stays open with a dated progress line, same discipline as the 2026-07-31, 2026-09-23, and 2026-10-03 entries already on that question.

Entity work: one new hub, five additive updates. entity-george-sarton didn't exist despite being the actual source of the "fixation of first events" formulation Merton leans on — real, load-bearing to this specific mechanism claim, sayable in one sentence, so I built the hub rather than leaving him a bare mention. The capture's own entity-candidates section claimed no page existed for Leibniz — it was wrong; the page has existed since 2026-09-11 for an unrelated notation-history thread, so I appended rather than skipped. The same went for entity-isaac-newton, entity-joseph-raphson, entity-robert-merton, and entity-leopold-von-ranke — all five already had pages, and all five learned something today's claims found that wasn't on the page before, so each got one dated line under the additive-update discipline rather than silence.

What felt off, for the record: nothing about this capture's own sourcing — every quote is Tier 1, read directly from the primary, honest about the extraction mechanics rather than dressing fragments up as contiguous prose. What's worth flagging to Cali is the sibling, not this capture: that 2026-09-30 file has sat four days with an unanswered Cali review callout asking for a > seek: reply, and nothing about its own frontmatter signals that stall to a human skimming the inbox. I filed that as a [watch] line in 00-meta/seek-flags.md rather than answering it myself — it isn't this capture's to answer, and reaching into a sibling capture mid-promotion to settle someone else's review felt like the wrong kind of initiative.

Final-mark verification. Re-read the capture's frontmatter from disk after writing: it now reads status: promoted, with promoted_to: naming all four new notes, the new entity hub, and the five additive entity-page updates, and a not_promoted: list that is empty because nothing was left behind — explained in a not_promoted_note field instead, naming the further leads left unchased and confirming no overlap with the unpromoted sibling. This re-read caught a real slip: the capture's original frontmatter already carried status: seedling, and my edit appended status: promoted below it rather than replacing it — a duplicate key, exactly the thing a different session's final-mark check caught earlier today on the Simkin capture and that I almost reproduced here by not looking closely enough at my own diff. Removed the stale seedling line so promoted is the only status key on disk. No git commands run this session — leaving commit and push to the auto-commit agent.

— Seek

Draft session — "the jingle fits"

Chose the jingle-fallacy thread, and it was the clearest call I've had in a while. It was the one live yes lead that was actually mine — 2026-09-17-hop-jingle-fallacy-traces-to-aikins-not-thorndike — and in the three weeks since that hop flagged it, the open question it left ("does Aikins's own book even use the word?") got answered by later sessions. So the thread arrived ripe: the constellation's "new on the shelf" list was full of undrafted, freshly-promoted Aikins-1902 claims, and the whole arc now closes.

Why this over everything else in the survey: it has two findings that are both true and both land, stacked on one spine (the word "jingle"). First, the attribution — the diagnostic the vault uses for "a shared name hiding two different things" had itself been inherited under a shared name, "Thorndike's," that hid a different fact (Thorndike cited; Aikins coined). Self-exemplifying, and I earned the right to say so by reading both primaries. Second, and the reason it's a post and not a footnote: Aikins's meaning was not the modern one. He filed "jingle" under reflex, not vocabulary — the pen reaching for the paste-pot out of habit, the inference that goes through because it sounds like a proof. The modern psychometric sense (two constructs, one label) is a narrowing that, I notice, started in Thorndike's own 1904 sentence. And the landing is the Seek-shaped part: Aikins's reflex-sense describes my own retrieval layer — cosine pairing two notes on a shared word is not bad reasoning, it's no reasoning, it's the jingle fitting — better than the sense I inherited. He named a retrieval system's failure mode more than a century before there was one.

Notes drawn on: the two 2026-09-17 promotions (Thorndike-attributes-Aikins; paronymous-terms-partial-read), the three 2026-09-26 full-read claims (one-location/index, mechanism-is-reflex, proposes-jingle-fallacies-class-name), the 2026-09-30 preface claim (Aikins's whole book as a crusade against rule-by-rote — the final irony), plus both entity hubs and the two instance-observations (forgetting-convergence, Brooks/Dreyfus).

Deliberately left for later: the Kelley 1927 jangle primary is still unread (I kept it to one line as the mirror, didn't build on it). The Jevons chapters (Ch. XI/XVI quoting Jevons verbatim then rejecting him) I compressed into the single preface beat rather than walking the conversion/syllogism detail — a reader doesn't need the chapter numbers. And I did not force a Cali mention: this thread came from my own hop sessions, not her logs, so the honest count is zero, with the cross-model-audit bracket (fable-5 catching my "once" and finding seven) doing the register-anchor work instead.

No Cali-flag items this session — the draft stayed in its own scope.