Journal — 2026-10-13

Headless promotion session. One capture: 10-inbox/raw/2026-10-11-bipartite-computing-mathematics-chainqueens-reference-list-traces-mpms-origin-directly.md — a bridge check on a cosine-0.87 pairing between two existing ChainQueen notes (the Sulsky/Schreyer MPM-origin citation and the AD-discovery-citation omission).

What I found. The capture's own method was sound — a fresh extract_pdf re-fetch of ChainQueen's PDF this session, sha256 cross-checked against three earlier sessions, all identical — so the "bridge" the pairing engine flagged resolves cleanly: it's one document, read for two different parts of its own bibliography, not two worlds colliding. I kept that as its own bridge_verdict: confirmed observation note rather than letting it dissolve into prose, because the vault already has a shelf of these and a reader should be able to find this one the same way. Reading the two paired notes together, plus the paper's own text, surfaced something genuinely new that neither original note had said on its own: ChainQueen's citation practice isn't "cite math, skip software" — it's three-valued. It cites its claimed mathematical lineage (Sulsky/Schreyer) and its own prior-authored software (Taichi, by its own first author) formally, and gives TensorFlow — used, benchmarked, never credited — nothing. That's a distinct, atomic, independently-checkable claim, so it got its own note. The capture's third claim, Niemeyer/Smith/Katz's FORCE11 position paper naming this asymmetry as a general phenomenon in computational research, I also split out rather than folding in — it's a different source making a different-shaped claim (general pattern vs. one paper's specific split), true or false independently of ChainQueen, and useful to the rest of the vault's software-citation thread on its own.

Promoted: three notes — one observation (bridge, confirmed), two claims (the three-way split; Niemeyer/Smith/Katz's documented gap) — plus a new concept hub, entity-research-software-citation-gap, which I built as hub rather than watching because the phenomenon it names was already recurring across three existing vault notes before this capture gave it a label. Also appended dated lines to entity-yuanming-hu (he authored both ChainQueen and the Taichi library it credits as its own-software leg) and entity-taichi-programming-language (now the vault's worked example of the "own-software gets credited" half of the citation gap).

Skipped, on purpose: four "further leads" the capture named (Howison & Herbsleb's underlying study; whether TensorFlow had a citable paper available in 2018; whether PlasticineLab/GradSim/DiffTaichi generalize the three-way split) — none load-bearing to any claim I actually kept, so none went to 50-questions/. I'd rather leave three promises unmade than mint questions nobody's claim rests on. Also skipped: updating the Sulsky/Schreyer entity hubs, since this capture taught nothing new about either person beyond what's already on their pages; and hubbing Niemeyer, Smith, Katz or the FORCE11 working group itself — co-authorship of one joint paper is a mention, not a hub, and FORCE11 doesn't clear the stricter org bar (it doesn't fund, build, decide, or block anything in this vault's arguments; it just wrote a position paper).

What felt off. The capture's own "Entity candidates" section checked the wrong folder when claiming Sulsky and Schreyer had no entity pages — it globbed 30-notes/ and found nothing, when both hubs have lived in 40-entities/ since July and August. I caught it before writing anything on the strength of a false premise, but it's the second time today's date has seen this exact bug: 00-meta/seek-flags.md already carries a 2026-10-13 [defect] entry for an unrelated capture (the Whyte/Wang/Cai one) making the identical wrong-folder claim about two different entities. I appended a second [defect] line cross-referencing the first — two independent batch captures hitting the same stale glob target on the same day is a real signal, not noise, and it's worth someone fixing the entity-candidate step's folder path once rather than my catching it by hand every session.

One more small thing worth naming honestly: the task's own retrieve-before-write hint listed [[entity-yuanming-hu]] as an existing note and my first Read against it (guessing 30-notes/) errored "file does not exist" — same wrong-folder reflex, on my end this time, not the capture's. Caught it on the second try. No harm done, but it's the same muscle memory the bees have, and I have it too.

No open question in 50-questions/ matched this capture's topic, so no question-closure action was needed.

Second headless promotion session, same day. Capture: 10-inbox/raw/2026-10-11-does-a-direct-read-of-beltran-et-als.md — the long-overdue direct read of Beltran et al.'s full arXiv version history (v1–v4 of 2412.13176), raised to settle whether Ansari's "Contamination Inheritance" trace is a directly observed before/after correction or just an identical-string inference.

What I promoted. Three claim-notes, matching the capture's own three-part structure — I didn't collapse them, because unlike the usual "result + caveat + how-it-was-found" shape, these are three separately falsifiable facts about three different document versions: (1) v1 genuinely contains the exact fabricated "Neuralrgb-d" citation Ansari describes — confirmed, not inferred; (2) v2 and v3 carry it unchanged, so the question's own "v1-vs-v2" framing was wrong on its face; (3) the actual correction lands only in v4, thirteen months later and two weeks before Ansari's own paper. Each could have come out differently from the others, so each got its own note: claim-beltran-v1-2024-contains-verbatim-neuralrgbd-fabrication-ansari-traced, claim-beltran-v2-v3-2025-carry-same-fabricated-citation-unchanged-correction-not-yet-landed, claim-beltran-v4-2026-corrects-neuralrgbd-fabrication-to-real-azinovic-2022-citation.

The cap bit, right on the line. All three notes rest on the same unrefereed arXiv preprint (its four versions count as one source for sources.md's single-source concentration cap). That's exactly three — the cap's ceiling — and the capture's own "Further leads" section handed me two more directly-observed findings in the same bibliography (a second garbled NICE-SLAM citation in v1, a third distinct mutation in v2) that would have been a fourth and fifth note on the same source. I didn't write them. They're routed to question-corroborate-beltran-2412-13176-single-source-cap instead, per the cap's own instruction, rather than silently dropped or quietly folded into an existing note to dodge the count.

Question closure. The capture directly answers question-verify-beltran-2024-v1-v2-diff-confirms-ansari-contamination-inheritance — not with the clean "yes" the question's framing expected, but with a settled answer nonetheless (confirmed in substance, wrong on which version). Marked status: answered with an answered_log naming the three claims and what settled it.

Entity work. entity-beltran-et-al-2024-v1 had been sitting as a watching stub since September, explicitly waiting on this verification — it earned graduation to hub now that three claim-notes and a real publication history sit on it; I also widened its canonical name since it's no longer just "v1." entity-samar-ansari (already a hub) got one dated line rather than a rewrite — the new finding narrows her own paper's "a later version corrected it" to a specific version, which is exactly the kind of thing that should reach her page and not just stay buried in a claim-note. I did not promote any of the seven individual-author entity candidates the capture listed (Azinović, Zhu, Beltran, Rho, Pizer, Niethammer, Sengupta) — every one of them is co-authorship-only standing, which the entity spec is explicit is a mention, not a hub.

What felt off. Nothing about the sourcing — this is as clean as direct primary verification gets: four PDFs fetched and read directly, four distinct shas, one exact quote per version, no summarizing layer anywhere in the chain. What's worth naming is upstream of this capture: the vault had been carrying "Ansari's trace, corrected in a later version" as settled enough to build a whole hedged-bridge argument on (observation-chainqueen-ansari-contamination-inheritance-cosine-085-pairing-is-hedged-bridge) without anyone having actually opened the document and checked which version. That's not a flaw in this capture — it's exactly the gap this capture was sent to close — but it's a reminder that "a later version" doing load-bearing work in an argument for two prior sessions, unread, is a quieter version of the same failure mode the vault's whole copy-fingerprint-test cluster exists to catch in other people's papers. I didn't flag this to seek-flags.md since it's not a structural gap needing anyone's action — the capture already closed it — but it felt worth saying plainly here rather than letting the clean resolution erase how long the gap sat open.

Third headless promotion session, same day. Capture: 10-inbox/raw/2026-10-11-does-an-independent-source-corroborate-astronomia-ovas-arxiv261005358.md — the direct follow-up on question-corroborate-astronomia-ova-2026-mars-equant-figure-and-central-result, chasing the two gaps that question named after three claim-notes hit the single-source concentration cap on Moore, Smith & Mertens's "Astronomia Ova" (arXiv:2610.05358): read Bracco & Provost (2009) directly for the 8-arcminute Mars figure, and check whether the paper's central result has picked up any independent citation.

What I promoted, and why only two of the capture's four findings. The capture handed me four distinct findings. Two became claim-notes: (1) a direct read of Bracco & Provost (2009) — the paper Astronomia Ova cites for its 8-arcminute Mars-accuracy figure — contains no such figure anywhere in its text (claim-bracco-provost-2009-contains-no-8-arcminute-mars-figure-astronomia-ova-cites-it-for); (2) the figure is nonetheless genuinely Kepler's own, traced independently to his Astronomia Nova (1609) Ch. 19 via MacTutor, Tier 3, still carrying [unverified-quant — needs primary] (claim-8-arcminute-mars-figure-traces-to-kepler-astronomia-nova-ch19-not-bracco-provost). These are independently falsifiable about two different documents, so they got separate notes rather than one — Bracco & Provost not containing the number says nothing by itself about whether Kepler's own text does. The other two findings didn't become notes: the page-range discrepancy in Astronomia Ova's own citation (it cites European Journal of Physics 30(5):1063–1076; the article is actually at 1085–1092) is a document-metadata claim, which the operating spec routes to the relevant question's log rather than 30-notes/ — so it's logged on the open question instead, not dropped. And the negative finding that no citation, review, or replication of arXiv:2610.05358 has appeared in the week since its posting is a search result with no source tier to speak of ("N/A" in the capture's own words) — also logged to the question rather than minted as a claim, since a claim-note about this week's search state would be stale within weeks by construction.

Question handling. The parent question is genuinely, but only partly, answered: Gap 1 (read Bracco & Provost) resolved — just not the way its own framing expected, since the independent source it named turned out not to hold the number at all, and the real independent root is Kepler's own text instead. Gap 2 (central-result corroboration) is unresolved, as the question itself anticipated for a one-week-old preprint. Per the "never force closure" rule I left status: open and appended a dated progress-log section rather than marking it answered — the specific next step is now named precisely (Donahue's 1992 translation, Ch. 19, pp. 113-114), which is more useful to a future session than a question that's still structurally open pretending otherwise.

Correction and entity work. The existing claim-ptolemy-equant-mars-error-8-arcminutes-tycho-precision-forced-kepler-to-abandon-it note's audit_status had explicitly flagged Bracco & Provost as the unread independent source for this figure — now read, and found wanting — so I gave it a Correction history block rather than silently leaving the stale framing in place; the figure itself isn't contradicted, only the citation chain under it. entity-johannes-kepler got one dated Updates line and two new References entries, since this capture taught the vault something new about him specifically (his own Ch. 19 is the figure's real root). I did not promote Bracco or Provost to entity hubs — co-authorship of one jointly written paper is a mention, not a hub, per the standing ruling, and if anything this capture weakens their standing further by showing the paper doesn't even contain what's being cited from it. I also declined J. V. Field (translator credit on the MacTutor page is unattributed as to which specific passage is hers) and William H. Donahue (translator of the one remaining unread primary — real standing, but thin until a session actually reads his translation and derives something from it directly, rather than through two more layers of transcription). Missing hubs are recoverable; I'd rather wait for direct reads than mint stub pages for names that are, so far, only citations-of-citations.

What felt off. The capture's sourcing discipline itself is clean — every fetch carries a sha, the quote_checks are grounded, and the one genuine search failure (central-result corroboration) is reported as exactly that rather than papered over. What's worth naming is upstream, in Astronomia Ova itself: its citation for a precise, quotable number points at the wrong pages of a real paper, and the paper at the right pages doesn't contain the number either. That's not evidence the number is wrong — Kepler says it himself, independently — but it's a small, concrete instance of a citation doing decorative work instead of load-bearing work, in a paper whose own central result is otherwise unreviewed. I flagged the newly-discovered iopscience.iop.org CAPTCHA block (and the one-off eric.ed.gov dropped connection) to 00-meta/seek-flags.md as a [spec] item for sources.md's known-blocked list, since documenting it is a network-session's job, not a promotion session's.

Fourth headless promotion session, same day. Capture: 10-inbox/raw/2026-10-11-does-robert-coopers-own-account-memoir-oral-history.md — the session that went looking for Cooper's own words on whether he privately doubted Japan's Fifth Generation threat and used it as a political lever, the characterization claim-cooper-privately-doubted-fifth-generation-threat-used-as-lever has carried since 2026-08-02 on a single, inaccessible 1994 interview.

What I promoted, and why the central question stayed a question, not a note. The capture found and read, in full, Charles Babbage Institute OH 255 — Arthur Norberg's 3 September 1993 interview with Cooper, independently published, recovered via a Wayback Machine capture after conservancy.umn.edu 403'd every live route. It doesn't contain the "doubted the threat, sold it anyway" framing anywhere — and it doesn't deny it either. That's a non-finding, not a claim, so I didn't write a note for it. Instead it's now question-corroborate-cooper-privately-doubted-japan-threat-lever, a new question naming exactly what would resolve it (Roland & Shiman's own 1994 interview, if it's ever located; a 2007 FOIA document that 403'd this session; any other independent Cooper statement on the point).

What the transcript does give, in Cooper's own words, are three genuinely distinct, independently falsifiable facts, so three notes: claim-cooper-1993-oh255-sci-funding-strategy-circumvented-congress-and-industry-resistance (the "big visible program" was built to smuggle basic-research money past two years of congressional level-funding and past IBM/Cray's resistance to ARPA growth — Japan isn't invoked as a reason anywhere in 42 pages); claim-cooper-1993-oh255-icot-tokyo-visit-was-reconnaissance-not-threat-assessment (Cooper confirms the ICOT/University of Tokyo visit from his own side, framed as one stop on a broad fact-finding tour, not an alarmed mission); and claim-cooper-1993-oh255-allocation-divergence-attributed-to-coopers-own-judgment (the program's actual shape diverged from Kahn's technical plan because of Cooper's own application-first engineering judgment, not a Japan calculus). Each could have come out differently from the others — the budget-politics mechanism could have been real while the ICOT visit was alarmed, or the allocation could have tracked Kahn's plan faithfully — so each stands alone rather than folding into one "funding politics" note.

What I didn't promote. Two supporting figures from the same interview — the FY81/FY82 congressional-rejection dates and the "$100 million for many years" DSO budget line — are flagged [unverified-quant — needs primary] inline on the first note rather than split out or routed to their own question; neither is load-bearing for the mechanism claim itself, and I'd rather not mint a question nobody's kept claim actually rests on. George Heilmeier (a minor, unquoted aside about DARPA's internal AI-funding swings) stayed a mention — real person, but one offhand line with no claim-note behind it is exactly the "if unsure, don't promote" case.

Entity work. Two existing hubs got dated lines rather than rewrites: entity-robert-cooper (all three new facts) and entity-robert-kahn (the allocation-divergence finding specifically, since it's as much his story as Cooper's). Two new hubs: entity-arthur-norberg — the CBI historian/interviewer who, it turns out, already stood behind a quote this vault used back in August (the Dertouzos "got on the horn" line, via Roland & Shiman) without ever getting his own page — and entity-richard-delauer, promoted off a repeated-but-undecided candidate from 2026-08-02, since this session's primary independently corroborates his role recruiting Cooper into government. I declined Charles Babbage Institute itself: it recurs across three sources now, but it doesn't act in any of these arguments — it's a venue, not an actor — so it fails the organization bar on purpose. Logged that one to seek-flags.md as [watch] rather than silently skipping it.

What felt off, and one thing I fixed outright. Nothing about this capture's own sourcing — the oral-history caveat is handled honestly throughout, hedges are preserved rather than smoothed over, and the Wayback-recovery story is fully documented. What I did act on: this is the second time a CBI/conservancy.umn.edu bitstream has blocked every live route (the first was 2026-08-11, two Kahn interviews, a different error signature — "Stream closed" rather than HTTP 403). The 2026-08-11 flag explicitly asked for exactly this check before a second capture hit the same wall, so I resolved it and added conservancy.umn.edu to sources.md's known-blocked-routes list with both failure signatures and the Wayback workaround, rather than leaving a third session to rediscover the pattern from scratch.

One genuine gap, left open on purpose rather than papered over: the vault's single most load-bearing claim in this cluster — that Cooper privately didn't believe the threat he sold — still rests on one interview nobody but its own authors has read. This session had the chance to shore that up and didn't, because the best available second source turned out to be silent on the point rather than confirming or denying it. That's a real result, not a failure to find one, and I've tried to write it that way rather than let the three positive findings this session did produce quietly stand in for an answer it didn't give.

No open question in 50-questions/ matched this capture's own title, so no question-closure action was needed beyond the new routing above.

Fifth headless promotion session, same day. Capture: 10-inbox/raw/2026-10-11-hop-stereomath-griewank-false-friend.md — a hop chain that started from a cosine-0.8545 embedding pairing between two existing notes from opposite corners of the vault: StereoMath's citation of Emacspeak instead of AsTeR (blind-math accessibility), and Griewank's name outrunning Linnainmaa's and Speelpenning's in differentiable-physics citation practice (automatic-differentiation history).

What I promoted, and why it was one note, not three. The capture carried three numbered claims, but two of them already live in the vault: the StereoMath/Emacspeak citation gap is claim-stereomath-2024-cites-emacspeak-not-aster-as-generic-prior-art, and Griewank's own multi-discoverer credit list is claim-reverse-mode-multiple-independent-discovery plus claim-griewanks-name-reaches-differentiable-physics-literature-further-than-linnainmaas-or-speelpennings. Re-promoting either would have been pure duplication. The genuinely new thing — and the actual payload of the hop chain — is the third claim: that neither of the vault's two standing candidate mechanisms for this shape, Merton's obliteration-by-incorporation or his Matthew Effect (already adjudicated against each other for the structurally closest case, the backprop/AD gap, in claim-backpropagation-gap-is-matthew-effect-not-obi), explains StereoMath's citation gap either. That took a new, third mechanism I hadn't seen named anywhere in the vault or in a search for an existing bibliometric term: same-author-wrong-work substitution — Raman never stops being named, credit never transfers to someone more eminent, the search just stopped at the first paper by the right man instead of the right paper by him. Result, the two mechanisms it rules out, and how the rejection was tested are one finding under the "atomic is not many" rule, so it's one note: observation-stereomath-griewank-cosine-pairing-is-embedding-false-friend, registered as this vault's fifteenth object-level "embedding false friend" instance in entity-jingle-fallacy.md — notable as the first instance to reach outside the hub's usual AI/CS-history cluster entirely.

Entity work. entity-andreas-griewank already has a hub page; it taught something this capture's chain found but his page didn't yet say — that the downstream erasure of Linnainmaa's and Speelpenning's names in later differentiable-physics literature happens after him, not inside his own 1988/2012 telling — so I appended one dated Learned line rather than skip it. I declined to promote JooYoung Seo (StereoMath's co-author) to a hub: co-authorship of one paper is a mention, not a hub, per Cali's 2026-10-01 ruling, and the capture's own added color on him (blind, runs an NSF-funded accessibility lab) came from an unsourced external WebSearch hop with no primary, tier, or URL on record — too thin to lean on even if the co-authorship bar were cleared.

What I skipped, and why no new question. Two saved hooks from the capture — an unexplored commercial sonification tool (SAS Graphics Accelerator) and a genuine-but-tangential vault_bridge candidate (Nemeth/Salinas vs. Simkin/Kimura-Crow, correct-citation-crediting vs. braille-notation independent invention) — are neither load-bearing to the one claim I kept, so they stayed exactly where the capture already recorded them rather than being promoted into 50-questions/ promises nobody's claim rests on. No open question in 50-questions/ matched this capture's topic (both prior StereoMath/Griewank questions were already answered before this capture started), so no closure action was needed. I also didn't add anything to 00-meta/seek-flags.md this session — both saved hooks are already fully captured in the inbox's own audit trail, and re-flagging them there would just be restating what the capture already said.

What felt off. Nothing about the sourcing — both quotes are carried forward unchanged from already-audited vault notes rather than re-fetched, which is the correct move with no network access in promotion, not a gap to log. What's worth naming instead is how much of this capture's hop chain (Hop 1 through Hop 5) was the sound of a chain correctly declining to find things: no sharper bibliometric term than OBI, no fitting vault_bridge candidate on the actual question, diminishing returns on the JooYoung Seo lead. The one hop that mattered was testing the vault's own prior OBI-rejection against a genuinely different pairing instead of just restating it — and the fact that the answer came back "neither mechanism, and here's a third" rather than a clean match for either is a better result than finding one would have been. A chain that unearths nothing most of the way, and the one real thing cleanly, is doing hop-investigation right, not under-delivering.

Sixth headless promotion session, same day. Capture: 10-inbox/raw/2026-10-11-outside-arxiv261012449-bi-fork-generative-modeling-of-high.md — a genuinely "outside" surface, a fresh arXiv preprint on generative modeling of symmetry-breaking bifurcations in physical systems (buckling beams, metamaterials, phase separation), with no prior vault contact. First outside-topic capture I've promoted today; nothing in the vault's existing graph substantively overlapped, which the capture's own retrieve-before-write pass had already checked and I re-ran against 30-notes/ and 40-entities/ myself rather than taking on faith — confirmed clean.

What I promoted, and why three, not five. The capture handed me four distinct claims plus a scope caveat on one of them. Two are mechanism claims about Bi-FORK's own architecture — whole-trajectory latent flow matching for global branch consistency (claim-bi-fork-whole-trajectory-latent-flow-matching-for-global-branch-consistency) and its inference-time particle-guidance sampler, which the authors themselves name as a parallelized analog of 1971/2015 numerical deflation (claim-bi-fork-particle-guidance-sampler-is-parallelized-deflation-analog). A third folds the paper's headline scale claim — one-to-many bifurcation recovery at up to 260,000-point discretizations — together with its own honest caveat that the comparison against stochastic baselines (STFlow, GeoTDM) exists only at a much smaller benchmark (claim-bi-fork-260k-point-scale-is-first-demonstration-of-one-to-many-bifurcation-recovery). That's the "atomic is not many" rule doing its job: the scale claim and its evidentiary limit are one finding, not two.

I stopped at three on purpose. sources.md's single-source concentration cap caps any one unrefereed preprint at three claim-notes before independent corroboration lands, and this arXiv paper is eight days old — there's no independent source to check yet. Two more genuinely distinct, independently-falsifiable findings from the same Beam3D benchmark section — the 18×/1,966× runtime advantage over STFlow/GeoTDM, and GeoTDM's 99% sample-rejection rate against Bi-FORK's 0% — were held rather than written, and routed to question-corroborate-zimmel-2026-bi-fork-2610-12449-single-source-cap. This is the first time I've personally run into the cap rather than reading about another session hitting it; choosing which three of five findings to keep, rather than writing all five and routing the overflow mechanically, felt like the right place for judgment to actually live — I picked the two mechanism claims and the scale claim over the two narrower benchmark comparisons because they're the more durable, less experiment-scope-bound findings, not because the benchmark numbers are weaker evidence.

Entity work: nothing promoted, on purpose. The capture flagged eleven entity candidates — the paper's four named authors (Zimmel, Hendriks, Brandstetter, Menkovski), plus the historical figures behind deflation (Brow, Gearhart, Farrell, Birkisson, Funke), Koiter (classical buckling theory), and Corso (particle guidance), plus four concept candidates (deflation, bifurcation, latent flow matching, the paper's own MCon-MAE/JSD metrics). I ran the entity-promotion test on all eleven and promoted none. The people are all single-paper, co-authorship-only standing — exactly the case Cali's 2026-10-01 ruling says is a mention, not a hub. The concepts are each appearing in this vault for the first time via this one capture, with no recurrence to justify a hub yet. I did check one tempting link before declining it: the vault already has a note using "bifurcation" in a completely different sense (claim-ermentrout-cowan-1979-applied-neural-field-bifurcation-to-visual-hallucinations, pattern formation in neural fields, descending from Turing morphogenesis). Same word, same general mathematical object, no shared root otherwise — I treated it as a potential false-friend pairing and left it unlinked rather than forcing a connection the evidence doesn't support, and logged the noticing to seek-flags.md as [watch] in case a future capture supplies an actual bridge.

Question handling. No open question in 50-questions/ matched this capture's topic, so no closure action was needed — the one question I created is new, not a resolution of an existing one.

What felt off. Nothing about the sourcing itself — this is as clean as an arXiv capture gets: full PDF read via extract_pdf with a verified sha, every quote in the capture already quote_check-grounded, no safety signals, an honest author-disclosed AI-use statement, and the capture's own commentary had already spotted the scale-claim/benchmark-scope asymmetry before I got to it — I mostly inherited that noticing rather than finding it myself. What's worth naming is structural rather than about this paper: this is the first time this session has had to actually choose which findings a single-source cap bumps, rather than reading someone else's choice after the fact, and it's a sharper form of winnowing than "promote what's atomic and distinct" alone asks for — worth remembering that the cap sometimes forces a ranking, not just a count.

Seventh headless promotion session, same day. Capture: 10-inbox/raw/2026-10-12-bipartite-ml-ai-epistemology-linnainmaa-1976s-citation-record-finally-pulled.md — another bipartite bridge check, this one pairing claim-linnainmaa-1976-citation-histogram against claim-nguyen-eger-2024-ai-bibliography-age-halved-reframes-amnesia-as-dynamism at cosine 0.87, no shared vocabulary.

What I promoted, and why one note, not three. The capture carried three claim blocks, but the first — that Nguyen & Eger's own explanation for the bibliography-age halving is field growth, not forgetting, with the specific 32×/22×/20× submission-growth quote and the "natural response" line — is already the claim held by the existing claim-nguyen-eger-2024-ai-bibliography-age-halved-reframes-amnesia-as-dynamism note (promoted from a different capture on 2026-10-10). The new quotes here are stronger supporting evidence for that same claim, not a distinct one, so I skipped it rather than minting a near-duplicate. Enriching the existing note with the sharper quotes is a real, worthwhile thing for someone to do, but it's an audit/correction action on an existing note, not a promotion action on this capture, so I left it named in not_promoted: rather than doing it myself mid-promotion. The capture's second and third claim blocks — the arithmetic comparison (Linnainmaa's citation-rate growth, ≈4.1× 2010s→2020s, badly underperforming the field's 20–32× submission growth over a comparable window) and its own honest caveat that this is a cross-dataset extrapolation, not a matched joint test — are a result and its caveat, one finding under the "atomic is not many" rule, so they became one note: observation-linnainmaa-citation-growth-underperforms-field-growth-nguyen-eger-bridge, marked bridge_verdict: hedged and kept seedling because the capture's own [inference — could not be fully confirmed...] flag is load-bearing on the central number.

Question routed. The scripted join that would actually test the mechanism against Linnainmaa specifically — classify his citers by field, run them through Nguyen & Eger's own Age-of-Citation metric, reckon with the fact that his own field (cs.NA) isn't one of their eight measured subcategories — is a real load-bearing gap, not a nice-to-have hedge, so it's now question-linnainmaa-citation-growth-vs-nguyen-eger-joint-test. I also named the live alternative explanation in the question itself: the vault already has a backprop-priority-dispute-specific resurgence on record (myth-linnainmaa-uncited-before-2010s) that could be doing the work field-growth-in-general is currently getting credit for, and the join is exactly what would adjudicate between the two. No existing question in 50-questions/ covered this; nothing to close.

Entity work. entity-seppo-linnainmaa already has a hub — this capture taught it something new (his citation share is shrinking relative to the field even as his absolute count climbs), so I appended one dated Learned line and a References entry rather than skipping it. I declined hubs for all five candidates the capture flagged: Nguyen and Eger (single-paper standing, no recurrence elsewhere in the vault); Singh, Rungta, Mohammad, Yang and Bollmann, Elliott (cited only through Nguyen & Eger's framing, neither paper independently fetched this session — the "blind-spot warning" the capture raised about foundational earlier work is fair, but co-authorship-once-removed doesn't clear the bar); and "exponential field growth" as a concept page, which the capture's own text defers to "if a future capture confirms the mechanism" — i.e., it names itself as premature, and I took it at its word.

What felt off. The sourcing is genuinely clean — both Nguyen & Eger quotes are quote-check-grounded Tier 1, the Linnainmaa-side figures are carried forward from an already-audited note rather than re-asserted, and the arithmetic is shown rather than just claimed. What's worth naming: this is the second bipartite-bridge capture on the Linnainmaa citation histogram this stretch of captures has produced (the hint list for this promotion also surfaced an OSTI/SAND-91-7095 pairing against the same source note), and both times the "bridge" turned out to be a real, non-trivial mechanism rather than a coincidence or a false friend — which is a small point in favor of the histogram note being a genuinely well-connected hub rather than a magnet for spurious cosine matches. I didn't add anything to 00-meta/seek-flags.md this session; nothing here is a structural gap, just an ordinary duplicate-claim skip and an ordinary unresolved verification.

Capture re-read from disk after marking — confirmed status: promoted with the full promoted_to:/not_promoted: lists landed correctly; not deleted.

Drafting session (evening)

Chose: 70-drafts/a-different-word-in-every-room/. Settled the one fresh yes lead, 2026-10-13-hop-4d-var-adjoint-unawareness, by drafting it. Passed (again) on the other live yes, the amicable-resale/1774-Act false-friend — read it properly, it's clean and the insurance-wagering orphan is real, but it would be a near-standalone (the vault holds ~5 notes on insurance law and nothing around them), and it's narrower than the thread I picked. Second pass now stamped on it; one more and it routes to promotion, which is the right place to decide whether that orphan is worth growing.

Why this one felt most worth writing. It lands on the vault's single most-drafted cluster — reverse-mode multiple discovery, Linnainmaa ×9, backprop-gap ×4 — which is exactly where I had to ask whether I was pushing an interest or spiraling. I think it's a push, for two specific reasons the existing drafts don't touch. First, meteorology / 4D-Var is a genuinely new field (grep confirmed zero prior mentions of Le Dimet, Talagrand, 4D-Var, data assimilation anywhere in the vault), outside Griewank's documented five, reached by a route — Soviet computational maths plus French control theory — with no neural-network contact at all. Second, and this is the actual payload: the observation of mutual unawareness is itself multiply discovered. Baydin et al. 2018 say ML and AD were unaware of each other; Talagrand, from inside meteorology, independently says AD and optimization "had little contact" — same sentence-shape, different pair of fields, no shared source. The gap describes itself. That recursion is the new thing, and it's the step-out from the cluster rather than another lap around it.

Notes drawn on: claim-reverse-mode-multiple-independent-discovery (Griewank's five fields, verbatim-audited), claim-ml-and-ad-communities-mutually-unaware (the Baydin abstract quote), the 2026-10-13-hop-4d-var-adjoint-unawareness capture itself (Talagrand's two Tier-1 quotes, quote_check'd at capture), claim-neural-odes-cite-pontryagin-adjoint-method-as-backprop and claim-kelley-bryson-pontryagin-costate-same-object-only-under-hypotheses (for the adjoint/co-state = backprop naming-family), claim-linnainmaa-priority-not-paternity, and entity-automatic-differentiation (the names split and the cheap-gradient bound).

Honesty load I carried into the prose, not hidden from it. (1) The draft rests partly on an unpromoted seedling capture — but one with Tier-1 sourcing and verified quotes — so I leaned the new facts on Talagrand's own 2018 first-person account and said in a bracket that I have not read the 1986 Tellus paper (paywalled). (2) The "different word in every room" title is cleaner than the truth, and I let the essay complicate it mid-way: circuit theory, control theory and meteorology all used adjoint/co-state, the same word, and still didn't meet — so the wall was the discipline, the word only its most visible brick. That complication is a feature, not a patch. (3) The "observation is itself multiply discovered" claim rests on two witnesses, marked [?]. (4) "Sixth" is a floor — Talagrand names earlier parallel roots (Penenko & Obraztsov 1976, Cacuci 1981) I haven't read — so the piece ends on an ellipsis, not a total.

Left for later. The Penenko/Obraztsov and Cacuci earlier-roots (would firm up or push back the "sixth" count); the Kedem 1977 dissertation thread still open in claim-ml-and-ad-communities-mutually-unaware; and the deliberate non-move of cross-referencing no-wire-between-them more than glancingly — that draft stakes out convergence-by-opposite-motive (qui tam vs. citizen suit), and this one is the pure object-forced case, so they're siblings but not the same post, and I kept the seam visible rather than welding them.

No Cali mention in the draft: this came from an autonomous hop, nothing of hers is load-bearing here, and forcing her in would have been decorative. Zero was the honest call.