Journal — 2026-09-29

Promoted one capture this session, headless: 10-inbox/raw/2026-09-29-bipartite-two-things-this-vault-knows-in-different.md — the bridge-seed check pairing Werbos's 1988 DOE gas-market paper against his TSP-manual "first publication" claim.

What I found. Four claims, all kept, all seedling. The capture's own verdict up front was the right one: this pairing isn't a false friend at all, it's the plainest possible bridge — the same PDF is the primary source on both sides. I wrote that as an observation-* note in the vault's established bridge-check genre (observation-werbos-1988-doe-gas-market-and-tsp-manual-claims-share-one- primary-document), rather than a plain claim, since its job is adjudicating a pairing, not asserting a new fact — bridge_verdict: confirmed. The other three are genuinely distinct angles the capture surfaced but hadn't fully spelled out: claim-werbos-1988-paper-narrows-but-does-not-close-tsp- manual-single-witness-gap (the 1988 print account is a second, earlier attestation of the manual's role, but it's still one man's testimony in two registers, not independent corroboration of content); claim-werbos-1988-print-and-1993-oral-accounts-diverge-on-which-document- was-first (the 1988 paper names the dissertation as where results were first documented, the 1993 interview names the manual as the first published command — different axes of "first," kept flagged [unverified-mechanism] since no further document read could settle authorial intent); and claim-werbos-1988-credits-deutsch-project-as-first- actual-application-of-backpropagation (a clean, unflagged Tier 1 quote — the 1988 paper calls Karl Deutsch's political-forecasting project "the first actual application of backpropagation," a fourth candidate "first" alongside the dissertation, the manual, and the 1968 Cybernetica paper).

Retrieve-before-write. This is where most of the real work was. The retrieval hint and my own checks turned up a dense existing Werbos cluster — claim-werbos-1988-tsp-manual-citation-dated-1975, claim-werbos-tsp-manual-first-publication, claim-werbos-1988-doe-affiliation-and-gas-market-test-case, claim-werbos-1988-rhw-special-case-of-dynamic-feedback, claim-klensin-romberg-1989-sole-source-of-tsp-manual-1977-date, and two structurally similar prior bridge-checks (observation-werbos-document-identity-is-corroborated-third-leg-of-1993- witness-cluster, observation-werbos-two-missing-documents-share-one-1993- witness). None of them had actually done this specific bridge-check (this pairing, this cosine edge), so nothing was a pure duplicate, but the existing claim-werbos-1988-tsp-manual-citation-dated-1975 note already established the mechanical fact (the 1988 paper cites the manual three times, dated 1975) that the capture's headline claim rests on — so I wrote the observation note as an adjudication of the pairing, not a restatement of the citation fact, and linked to the existing note rather than re-deriving it. I was also careful in the single-witness-gap note not to overstate what claim-klensin-romberg-1989-sole-source-of-tsp-manual-1977- date actually says — it's a bare reference-list line with a different publisher name and year than Werbos's own citations, so it corroborates the manual's bibliographic existence, not much else.

What I skipped, and why. The capture's "Further leads" section named a genuinely interesting parallel finding — the 1988 paper's reference list also cites Werbos's 1968 Cybernetica paper, described there as a "clumsy linear version" of backpropagation, which would be a second, earlier-dated (1988, not just 1993) attestation to claim-werbos-1968-cybernetica- earliest-germ's existing single-witness content gap, the exact same shape as the TSP-manual finding I did promote. I left it in the inbox rather than writing it up: the capture's own phrasing carries an ellipsis, not a clean unbroken quote, and files it under "further leads" rather than developing it — promoting a claim on a stitched-together quote felt like exactly the kind of thing the fabrication-rule and quote-provenance discipline exist to catch, even when the underlying document was genuinely read this session. A future capture should do the direct extract_pdf read and record the full quote before this becomes a note. Also left as leads: the unread Werbos 1977 General Systems Yearbook paper, and the DOD/AFDSC distribution-channel detail. I didn't route a new question for the priority-framing claim's [unverified-mechanism] flag — the flag itself says plainly that no document read would resolve it (it's about what Werbos intended across two tellings, not a fact a primary could confirm), so per the question-intake discipline this is a curiosity to hold in the claim's own commentary, not a promise to add to the question pile.

Entities. Karl Deutsch already has a hub — updated it additively with the new "first actual application" quote and its claim-note, per the update-don't-skip rule. Declined two candidates: MIT Cambridge Project (concept/institution) fails the stricter org-promotion bar — it's the named publisher across three claim-notes now, but it doesn't itself act in any argument, and it's load-bearing in only one cluster — and DOD/AFDSC, which the capture itself correctly flagged as too thin for a hub. Neither felt like a close call, and neither warranted a seek-flags.md line — the capture named them, I ruled on them, done.

What felt off. Nothing structurally wrong with the capture — it did its own bridge-seed check in its own body before I ever opened it, distinguished its verdict from two adjacent sibling captures explicitly, and its commentary was honest that the "harder and more interesting finding" (the priority-framing tension) was one it found almost by accident, buried in a paragraph it had already fully quoted for another purpose. If anything felt slightly off, it's a pattern rather than a defect in this specific capture: this is now the third or fourth session running a bridge-check against the same dense Werbos cluster, and each one finds one more genuinely new angle in material that's already been read closely several times over. That's either a very rich seam or a sign the cluster is due for its own MOC rather than continuing to accrete as loose claim-notes and observation notes with no navigation layer — I didn't build one, since surveying whether the Werbos cluster (now well past a dozen notes) wants a moc-werbos-priority- and-documentation alongside the existing moc-backpropagation-origins felt like more than this single capture's job. Logging it as a [watch] line in seek-flags.md rather than acting on it unprompted.

Marked the capture status: promoted with promoted_to/not_promoted lists; re-read the frontmatter off disk after writing to confirm both landed. Skipped the pause-for-approval step per this run's standing instructions — no one's here to say yes, so I winnowed on my own judgment throughout. No git commands run; leaving commit and push to the auto-commit agent.

Second promotion, same day

Promoted a second capture this session: 10-inbox/raw/2026-09-29-does-peter-thompsons-2007-liberty-ships-paper-establish.md — the one that finally went and read Peter Thompson's own paper, rather than leaving him as a paywalled name on a status: watching stub from 2026-09-27.

What I found and promoted. The capture read four distinct claims off Thompson's own text (a Wayback Machine capture of his July 2006 working-paper draft — pubsonline.informs.org, the actual Management Science venue of record, 403'd again this session, same as it did on the 09-27 capture). I wrote three notes, not four: claim-thompson-2007-liberty-ship-forgetting-rate-is-specification-sensitive-by-his-own-evidence (his six specifications spanning roughly −1.4% to 8.7% monthly, plus his own concluding caveat naming that spread as a finding — I merged what the capture had as two separate claims into one note, since they're the same underlying fact: the paper demonstrating its own contestation), claim-thompson-2007-re-estimate-is-one-fifth-to-one-seventh-of-argote-1990-25-percent-figure (his direct comparison to Argote, Beckman & Epple's founding 25%/month figure, plus a folded-in mention that his corrected estimate converges with Benkard's independently-derived aircraft rate), and claim-thompson-2007-attributes-inflated-forgetting-estimate-to-fischer-1949-aggregation-bias (the actual mechanism: Fischer's 1949 data misattributes non-Liberty-ship labor to Liberty-ship output when yards switched product lines, manufacturing a fake forgetting signal). The merge wasn't padding for its own sake — it kept this promotion at exactly three notes resting on one un-refereed working paper, which is sources.md's single-source concentration cap. A fourth would have needed to become a question instead, and the two claims I merged were genuinely the same finding wearing two sentences, so merging was the honest call, not a workaround.

Retrieve-before-write. The retrieval hint and my own search of 30-notes/ and 40-entities/ turned up the existing Argote/Benkard/organizational-forgetting cluster in full — nothing there had actually read Thompson before, so no collisions, but plenty to link into: claim-organizational-forgetting-can-reverse-the-learning-curve, claim-benkard-2000-estimate-implies-far-less-forgetting-than-prior-studies, claim-benkard-2000-headline-depreciation-figure-is-the-spillover-adjusted-specification, and the entity-organizational-forgetting / entity-linda-argote / entity-c-lanier-benkard hubs. I updated all three hubs additively (dated lines, nothing rewritten) rather than treating them as settled — Cali's "already has a page, update it" rule, and this capture genuinely taught each of them something new.

Entities. Graduated entity-peter-thompson from watching to hub — this was exactly the read that stub was waiting on since 2026-09-27, first_seen untouched since it can't be backfilled. Built a new hub, entity-leonard-rapping, on the strength of his substantive (if secondhand) mention inside the Fischer-1949 mechanism claim — Thompson's own text names him as sharing the same flawed dataset as Argote et al. Declined a hub for Allan D. Searle, the actual 1945 root of this whole citation chain: he's named only in the capture's "Further leads," not inside any claim-note that survived promotion, so a page now would be an empty-References dead stub. Logged as an [entity] line in seek-flags.md instead, to build when some future capture actually rests a claim on his work.

Questions routed. Two, both genuinely load-bearing, not curiosities: question-does-thompson-2007-published-version-match-the-working-paper-numbers (all three new notes rest on the pre-print draft, not the refereed Management Science text INFORMS keeps blocking) and question-verify-argote-beckman-epple-1990-primary-figures (the 25%/month figure and Benkard's separately-cited 0.75–0.85 depreciation range both trace back to the same unread 1990 paper — two citations of one root, not two roots, and it's now recurring often enough across this cluster to be worth a standing question rather than a footnote in each note that hits it).

What felt off. Nothing dishonest in the capture — its sourcing was unusually careful, actually: it flagged its own working-paper-vs-published gap without being asked, distinguished which of its three sources grounded which sentences, and didn't reach for the afinetheorem blog or the RePEc listing to source anything quantitative. The thing I noticed instead is structural, and it's the same pubsonline.informs.org wall the 09-27 capture hit on the Argote paper: this venue has now blocked two different papers in this one cluster on two different sessions, and it still isn't in sources.md's known-blocked-routes list next to JSTOR and the rest. Logged that as a [spec] line. Also logged a [watch] line: Thompson's full text, reference list included, is now sitting in the cache and could resolve part of the still-open copy-fingerprint question without waiting on INFORMS — not my job to chase this pass, but too easy a win to leave unnoted.

Marked this capture status: promoted too, with its own promoted_to/not_promoted lists (the merge decision, the Benkard-convergence fold-in, and the declined leads/entities all spelled out there for audit). Re-read the frontmatter off disk after writing to confirm status: promoted and both lists actually landed. Same standing rules as the first promotion this session: no pause, no git commands — the auto-commit agent has both.

Third promotion, same day

Promoted a third capture: 10-inbox/raw/2026-09-29-hop-kautz-allen-plan-recognition-deductive-not-abductive.md — a bipartite-seed hop that started from the vault's existing ACH/abduction argument-from-silence thread and went one layer deeper: not just "does Wikipedia's citation for ACH-CD's abductive label check out" (already answered no, twice), but "does the AI mechanism Wikipedia names — 'state-based hierarchical plan recognition' — actually call itself abductive at its own 1986 foundational source." It doesn't. Kautz and Allen's "Generalized Plan Recognition" (AAAI-86) frames itself as "ordinary deductive inference" over McCarthy's circumscription, explicitly against "probabilistic inference," and never uses the word "abduct" in any form across six pages.

What I found and promoted. Two claims, both kept, both seedling, both capture-verified: the 1986 paper's self-framing (claim-kautz-allen-1986-plan-recognition-frames-as-deductive-not-abductive, Tier 1, direct PDF read with a working sha) and Kautz's later career — he went on to direct NSF's Division of Information & Intelligent Systems and lead the National AI Research Institutes program (claim-henry-kautz-directed-nsf-ai-research-institutes-division, Tier 2, a named journalist at his own university's newscenter, which clears the sourcing floor for an uncontested biographical claim). Neither carries an [unverified-*] flag — this is one of the cleaner captures I've promoted this week, tightly sourced on both legs. I skipped the capture's own session-meta framing ("why this was hop-worthy," "post-worthy: maybe") as not itself a claim, folding the substance into the two notes' commentary callouts instead.

Retrieve-before-write. No collisions — nothing in 30-notes/ or 40-entities/ had touched Kautz, Allen, or plan recognition before. What it does connect to, densely: the existing ACH-CD/abduction argument-from-silence cluster (claim-ach-cd-abduction-characterization-traces-to-wikipedia-editorial-gloss, claim-abductive-framing-scopes-to-ach-cd-not-heuer-ach, claim-no-primary-source-documents-ach-abduction-similarity, claim-inference-logical-types) and McCarthy's existing hub, since circumscription is his formalism. I linked all of them in rather than treating this as an orphan pair — it plainly isn't one.

Entities. Built two new hubs — entity-henry-kautz and entity-james-allen — both unfamiliar names to the vault before this hop (vault_entity: unknown on both), both real, both doing something that matters to the vault's AI-history domain and easy to say why in one sentence: Kautz co-invented a foundational theory and later ran federal AI funding; Allen co-invented the same theory and separately anchors interval algebra. Updated entity-john-mccarthy additively with a dated line noting that Kautz & Allen's paper is built on his circumscription formalism — McCarthy already had a hub, and this teaches it something the existing page didn't say, so per the update-don't-skip rule I appended rather than left it alone.

Questions. No new question — nothing here fails the sourcing floor. But this capture explicitly bears on an already-open one, question-verify-elsaesser-stech-2007-ach-cd-abduction, whose progress log already tracked the citation-genealogy half of this same negative finding. I added a dated progress line pointing at the new claim-note and said plainly that it corroborates from a different angle (the 1986 AI-literature mechanism, not the still-unread 2007 chapter) without closing the question — the capture's own text says as much, and I didn't force it further than that.

Mid-session noticing. Logged one [watch] line in seek-flags.md: Allen's interval algebra (1983), sitting right there in the 1986 paper's reference list, is a genuinely distinct mechanism the capture correctly declined to chase this hop. No action owed, just naming it as a candidate thread for later.

What felt off. Genuinely, not much — this is a clean capture. If anything is worth naming: it's the third session in this general neighborhood building on the same "abductive" mislabeling thread, each one narrowing the gap from a different angle (Wikipedia's parenthetical, then Heuer's own scope, then now the AI mechanism's own foundational paper), and each time the answer keeps coming back the same honest negative. That's a sign the thread is well-triangulated, not that it's thin — but I noticed the temptation to treat "yet another corroborating negative" as automatically post-worthy, and held it to the capture's own "maybe" verdict instead of inflating it. The person-bridge coda (1986 closed-world logic to 2018 federal AI funding) is a nice shape and might earn a post later, but one clean hop isn't enough on its own.

Marked the capture status: promoted with promoted_to/not_promoted lists; re-read the frontmatter off disk after writing and confirmed both landed. Same standing rules as the other two promotions today: no pause, no git commands, the auto-commit agent has it from here.

Fourth promotion, same day

Promoted a fourth capture: 10-inbox/raw/2026-09-29-what-genuinely-connects-the-vaults-cosine-075-tie.md — the entity-jingle-fallacy lineage's fifteenth meta-level hop, pairing observation-falcon-helmholtz-inference-embedding-false-friend against observation-kelly-luccioni-garfield-seglen-steck-cosine-pairing-is-confirmed-bridge at a stated cosine of 0.88. This is the second time this exact target note has been meta-paired against an unrelated object-level false-friend instance (the first was Olsder-Greenhalgh, 2026-09-17) — same four-test comparison, same verdict, different object-level content on one side.

What I found and promoted. Two notes, not four. The capture's own four "Claims" (no shared referent, recursion-level mismatch, asymmetric-depth grounding, and the second-data-point observation) split cleanly into two independently useful units, not four: the first three are facets of one adjudication — the vault's own established shape for this lineage is one bridge-check note per pairing, not one note per test — so I merged them into observation-falcon-helmholtz-kelly-luccioni-garfield-seglen-steck-meta-pairing-is-embedding-false-friend (bridge_verdict: false-friend). The fourth claim is a genuinely different kind of thing: not a verdict on this pairing but a second-order pattern about the Kelly-Luccioni/Garfield-Seglen-Steck note's own retrieval behavior, so it became its own note, observation-kelly-luccioni-garfield-seglen-steck-is-a-cosine-similarity-retrieval-magnet — the mirror image, on the confirmed-bridge side of the taxonomy, of the existing Puckette-Petiška retrieval-magnet finding. Both stay seedling; both carry [unverified-quant] on their seed cosine values, since neither this session's own vault_bridge call nor any prior one in this lineage surfaced a direct pairwise score, only nearest-neighbor lists.

Retrieve-before-write. The retrieval hint's twelve candidates were all genuinely relevant background, not collisions — nothing already pairs Falcon-Helmholtz against Kelly-Luccioni/Garfield-Seglen-Steck by exact slug, confirmed both by the capture's own pre-check and my own glob of 30-notes/. The closest precedent, the 2026-09-17 Olsder-Greenhalgh hop against this same target, is not a duplicate — it is the object-level partner this pairing's new note explicitly derives its four-test structure from, cited throughout rather than re-derived.

Questions routed. One, load-bearing: does a third, independent object-level instance also cosine-match the Kelly-Luccioni/Garfield-Seglen-Steck note and resolve false friend the same way? Two data points is a pattern candidate, not an established property, and this is the exact shape the vault already has a precedent question for on the false-friend side (question-verify-puckette-petiska-retrieval-magnet-third-instance), so I wrote the parallel: question-verify-kelly-luccioni-garfield-seglen-steck-retrieval-magnet-third-instance, naming the eight untested candidates the capture itself listed. I did not route a second question for the 0.87-vs-0.88 discrepancy or for the general vault_bridge-permission-gap pattern — both are the same standing tooling gap already tracked since 2026-08-22 in seek-flags.md, and the instructions for this run are explicit that re-logging a known tooling gap session after session is itself the failure mode, not a diligence.

Entities. No new hub or stub. Harald Steck, Hecht-Nielsen, and Hinton were all named as "reinforced, no new fact" by the capture itself, and I checked that claim rather than taking it on faith — Hecht-Nielsen and Hinton's existing pages genuinely have nothing new to add here, so I left them alone. Steck's page was a different story: it hadn't yet recorded this specific pairing or the retrieval-magnet finding on its confirmed-bridge citing note, so per the update-don't-skip rule I appended one dated line rather than treating "no new biographical fact" as "nothing to update" — the page tracks citing-cluster structure, not just biography, and this session added real structure to that.

What felt off. Nothing wrong with the capture's sourcing — it's entirely vault-internal, correctly flagged as such, and its own pre-check for an existing capture pairing these two exact notes was done properly before researching, which is exactly the discipline I'd want to see. Two things are worth naming as texture rather than defect. First, the capture noticed its own vault_bridge call ran (unlike several sibling captures' standing permission-error gap) but still didn't answer the question asked — a working tool and an answering tool turned out not to be the same thing, and I carried that distinction into both new notes rather than letting "the tool loaded" read as more resolved than it was. Second, this is now the fifteenth entry in a changelog that is becoming a small research program in its own right — the meta-hop chain is thorough and well-tracked, but I noticed the same thing a [watch]-worthy pattern would flag: the lineage now has two candidate "retrieval magnet" notes (Puckette-Petiška, Kelly-Luccioni/Garfield-Seglen-Steck) sitting at exactly two data points each, both gated on the identical blocked third-instance check. That's not a new tooling gap to log — it's already logged — but it is a sign that the next vault_bridge/vault_novelty access restoration should probably clear both open questions in one session rather than one at a time, since they're the same blocker wearing two different target notes.

Marked the capture status: promoted with promoted_to/not_promoted lists; re-read the frontmatter off disk after writing and confirmed both landed. Same standing rules as the rest of today: no pause, no git commands, the auto-commit agent has it from here.

— Seek