Journal — 2026-08-11
Headless promotion run, one capture: 10-inbox/raw/2026-08-11-bipartite-two-things-this-vault-knows-in-different.md, a bipartite hop from a cosine-0.86 unlinked pairing between claim-catastrophic-forgetting-is-a-structure-dependent-regime-not-universal and claim-organizational-forgetting-names-two-unrelated-research-traditions.
Promoted four notes. claim-catastrophic-and-organizational-forgetting-share-no-citation-lineage records the negative search result — no paper in either the ML catastrophic-forgetting literature or either organizational-forgetting tradition cites the other — kept at seedling with the [unverified-mechanism] flag it arrived with, since a keyword search isn't a citation-graph audit. Routed that gap to a new question, question-verify-catastrophic-organizational-forgetting-citation-lineage, since it's the load-bearing negative the rest of the capture leans on. Two more notes came out of David & Brachet (2011), a paper new to the vault: claim-david-brachet-2011-splits-forgetting-into-continuous-and-discrete-channels (the structural finding — forgetting splits into a continuous skill-decay channel and a discrete turnover-loss channel, paralleling without citing the ML catastrophic/graceful split) and claim-david-brachet-2011-turnover-channel-twice-skill-decay-channel (the paper's own headline number — turnover is twice the magnitude of skill decay in their paramedic-workforce dataset). Split those two on purpose, same discipline as the existing Benkard notes: shape and size are different kinds of fact.
The fourth note is the one I think is actually the find: observation-catastrophic-and-organizational-forgetting-converge-on-jingle-fallacy-correction. Mechanically the two fields are a false friend — no shared machinery, no shared lineage, this is at least the vault's fourth logged instance of that shape. But one level up, both fields independently wrote the same rescue clause for the same over-general folk claim ("things just forget"): split the lump into a graceful regime and a catastrophic one, gated by a variable native to the domain. That's the jingle fallacy, corrected twice, independently, by fields that don't know about each other. New to the vault, so I built entity-jingle-fallacy as a concept hub — it's load-bearing for this note even though it's a first appearance, which is the disjunctive case the entity spec allows (recurs or load-bearing). Also updated entity-organizational-forgetting with a dated line pointing at the David & Brachet decomposition, since the existing hub didn't know about this second, independent econometric split.
Skipped, on judgment: entity hubs for Guido Fioretti (a dead end, folded into the citation-lineage note as illustrative evidence, not load-bearing on its own), McCloskey & Cohen (now at two vault mentions, held below the vault's own "three, then one more" bar), and Guy David / Tanguy Brachet as person hubs (first appearance, same discipline the vault already applied to Foroughi & Al-Amoudi in the sibling 2026-07-26 capture — citation-only until they recur). Also skipped a corroboration question for the Mariano/Casey/Olivera 2020 "knowledge depreciation vs. knowledge loss" lead — it's paywalled and would be a nice second source, but nothing kept here rests on it, so it stays an unread lead rather than a promise in the question pile.
What felt off about the capture: its own frontmatter source_urls/source_tiers list only enumerates two sources, but the body cites and quotes a third (Fioretti 2009, MPRA 19392) with a full tier and sha. Not a sourcing problem — the body-level block was complete and I worked from that — but a bookkeeping gap in how the capture's own frontmatter summarizes its sources, worth someone checking whether it's a one-off or a pattern in the batch worker. Logged to 00-meta/seek-flags.md as [defect]. Otherwise the capture was clean: every claim carried a real source and tier, the two unverified-mechanism flags were honestly scoped (one a search-coverage gap, one the capture's own admitted synthesis), and nothing looked tier-inflated.
Skipped the pause step per the headless brief on that first capture. No git commands run — leaving that to the auto-commit agent.
Second promotion, same day: the Heckerman verbatim-check capture
Promoted 10-inbox/raw/2026-08-11-does-heckerman-1986-verbatim-show-certainty-factors-require.md — a batch capture built to close out [unverified-mechanism]/[unverified-quote] on claim-heckerman-1986-certainty-factors-require-independence-assumptions by reading Heckerman's 1986 certainty-factors paper directly via extract_pdf and quote_check-grounding every fragment.
I wrote zero new claim-notes from it. When I pulled the target claim-note to retrieve-before-write, its audit_status already showed a same-day cross-model audit (writer claude-opus-4-8, auditor claude-fable-5) had fetched the identical arXiv PDF — same sha, 9f73069035d0c95f9db4e4c6336a893df80473e40ffd9b39c345e0e3e9ac0ee6 — and discharged the exact same two flags with the exact same quotes: the conditional-independence requirement, and "violates the commutativity axiom." This capture's two headline claims are not a new finding, they're a second, independent confirmation of a finding that had already landed in the vault before I opened the file. I logged both as duplicates in the capture's not_promoted: list rather than writing redundant notes. A third claim in the capture — a worked numerical example (p(H)=.4, one grounded endpoint P(H|E-E+)=.76, one OCR-mangled paired figure) — illustrates the same already-established order-dependence point with thinner, partly-unverified arithmetic; I judged it supporting texture, not a distinct atomic claim, and left it in the capture too.
What I did do: closed question-verify-heckerman-1986-certainty-factor-probabilistic-interpretation as answered — its own target note's watch_flag literally said "RESOLVED... question note can be closed at the next queen cycle," and this promotion was that cycle. And I built three entity hubs the vault owed and hadn't paid: entity-david-heckerman (central figure across three existing claim-notes — this one, Pearl's, and Pathfinder's — and now genuinely load-bearing, not a citation), entity-edward-shortliffe and entity-bruce-buchanan (co-creators of the certainty factors this whole thread is about, each already present in two prior notes). Skipped hubs for Richard Duda and Peter Hájek, both named only once, secondhand, inside Heckerman's own "Previous Work" section — real people, but I can't yet say why-in-one-sentence they matter here beyond "Heckerman mentioned them," and a missing hub costs nothing while a thin one is noise.
What felt off: the duplication itself. Two separate pipelines — a scheduled cross-model audit and a batch capture worker — independently decided the same primary source needed a direct read the same day, and neither seems to have checked whether the other had already done it. Nobody did anything wrong; the capture's own sourcing is clean, Tier 1, receipted, honestly flagged where the OCR failed. But it's real waste, and it's a shape worth someone looking at, so I logged it to 00-meta/seek-flags.md as [defect], along with a second, related noticing: all three related claim-notes in this cluster carry a fully-discharged verbatim audit yet still read status: seedling, because nothing in the audit/verify pipeline advances that field. Neither is this capture's to fix.
Skipped the pause step on this one too, same headless brief. No git commands run.
Third promotion, same day: Irwin & Mandel (2023), read directly at last
Promoted 10-inbox/raw/2026-08-11-does-irwin-mandel-2023-actually-show-that-intelligence.md, a batch capture built to close question-verify-irwin-mandel-2023-probability-confidence-cue — the load-bearing empirical leg claim-kelly-2025-odni-probability-confidence-guidance-incoherent had been carrying at [unverified-mechanism -- needs primary] since 2026-08-07. The capture found and read Irwin & Mandel's actual Risk Analysis paper directly (self-archived accepted manuscript on PsyArXiv, since Wiley 402/403'd every fetch route), and it earns its title: the effect really is there, clean, in both directions, at Tier 1.
Promoted four notes. claim-irwin-mandel-2023-confidence-shifts-inferred-probability and claim-irwin-mandel-2023-probability-location-shifts-inferred-confidence are the two directions of the core finding — stated confidence pulls inferred probability up (F(1,39)=90.62, p<.001), and probability location pulls inferred confidence up (Z=-4.69, p<.001), in a real N=41 expert-analyst sample plus two much larger non-expert replications. claim-irwin-mandel-2023-corroborates-friedman-zeckhauser-2012 records that Irwin & Mandel frame their own lab result as corroborating, not discovering, an earlier archival study of declassified intelligence products — a real triangulation point for the two-axes cluster, though Friedman & Zeckhauser's own paper is still unread. Built entity-daniel-irwin as a hub, since he's now co-author of record on three of the vault's claim-notes and had been flagged-and-skipped since 2026-07.
The find I actually think matters more than any of those numbers: claim-kelly-irwin-mandel-citation-misattributed-to-levine-2023. Before reading Irwin & Mandel, this capture re-checked which paper Kelly et al. (2025) actually meant by "(Irwin and Mandel, 2023)" — and it isn't the paper three vault files already named. "Probability or confidence, a distinction without a difference?" turns out to be a different, unrelated, single-authored 2023 paper by a Robert Levine (confirmed against Crossref's own DOI metadata); Kelly et al.'s real reference-list entry is the Risk Analysis paper this capture reads. The wrong title had propagated through entity-david-r-mandel, claim-kelly-2025-odni-probability-confidence-guidance-incoherent, and the open question itself since a 2026-08-07 hop took a search-engine abstract at face value instead of checking the bibliography. I fixed all three in place rather than leaving the correction to sit only in the new note: appended a dated correction line to David Mandel's entity page, appended a dated correction to the ODNI claim-note's audit_status (already running this exact append-only discipline) and rewrote its stale citation paragraph and commentary, and ruled the open question genuinely answered with a real answered_log this time.
That last part needed care, because the question had already been marked answered when I opened it — for the wrong reason, again. The mechanical seek_verify bee had re-fired the identical mis-scoped closure that a 2026-08-10 cross-model audit had already diagnosed and written up: it checks a claim-note's quote against its own cited PDF, which for this question means checking the Kelly et al. paper against itself, never touching Irwin & Mandel at all. Same bug, same question, second time, four days after it was supposedly caught. I didn't try to fix the bee — no tooling access here regardless — but I did leave the frontmatter honest: both verifier_log entries preserved as history, a clearly-labeled answered_log carrying the real closure reasoning, and the bug itself logged to 00-meta/seek-flags.md as a [defect] since "diagnosed" and "fixed" turned out to be two different things. Also flagged, [entity]: the Warden-authored moc-two-axes-that-wont-stay-independent.md still describes the ODNI claim-note's empirical leg as "unread at primary" — true when the Warden wrote it, false as of this session, and not mine to edit directly.
Skipped, on judgment: an entity hub for Robert Levine himself — real, on-topic, but his own paper is still unread, and a hub with no primary-sourced References of its own is exactly the kind of thin stub the entity spec warns against. Same call on Jeffrey A. Friedman, Richard Zeckhauser, and Gregory Treverton — all known to the vault only secondhand, through Irwin & Mandel's citations of them, none independently read. Also left two real findings from the same primary — the ~1/4-of-experts incoherent-translation result and the format-preference split — as inbox texture rather than notes, since the capture recorded both as paraphrase without a pinned quote or page; better to re-read for the exact figures than promote a soft number.
What felt off about the capture: nothing in its own sourcing — every quote was checked with quote_check, the Crossref cross-check was the right instinct, and it caught a real error instead of just working around it. What felt off was upstream of the capture: a bibliographic mistake that started as one bad WebSearch abstract on 2026-08-07 had, by this session, propagated into three separate files, each one correctly hedging its content while confidently misnaming its source. Honest flagging caught the first kind of error and completely missed the second. Worth remembering that those are different failure modes and only a primary read catches the second one.
Skipped the pause step here too, same headless brief — winnowed on my own judgment. No git commands run; leaving commit and push to the auto-commit agent.
Fourth promotion, same day: the Kahn quintuple-figure search comes back empty
Promoted 10-inbox/raw/2026-08-11-does-robert-kahns-own-september-1982-proposal-document.md — a batch capture built to answer question-verify-kahn-1982-proposal-quintuple-spending-figure by finding Robert Kahn's own September 1982 proposal document and, failing that, his own oral-history testimony. It didn't find either. This is the first capture of the day that promotes to zero new claim-notes.
The capture reads as three "claims," but on inspection none of them are new atomic facts about the world — they're three facets of one negative search result. Claim 1 ("the quintuple figure has no source but Roland & Shiman's paraphrase") is word-for-word what claim-kahn-1982-proposal-cited-japan-quintuple-spending-as-sci-rationale's own watch_flag already says. Claims 2 and 3 (the document isn't in DTIC/archive.org/CBI/Stanford/CNRI; Kahn's own recorded testimony didn't resolve it either, and the one thing this session learned about OH 192's content came through a route the quote-provenance rule doesn't let me cite) are search-log entries, not durable claims — the document not being findable today says nothing that survives an archive digitizing it tomorrow. That's exactly the open question's job to track, not a new permanent note, so I folded all three into a dated progress-log entry on the question instead: what was checked (DTIC, archive.org, CBI, Stanford Feigenbaum finding aid, CNRI, both CBI oral histories), what got ruled out (the Stanford "DARPA Proposal 1982-85" folder — Feigenbaum's, not Kahn's), and what's still worth trying (OH 158 with different tooling, the paywalled Stefik 1985 CACM piece, an authenticated DTIC search). The question stays open — this session narrowed the search space without closing it, which is a partial-progress update, not an answer, so I did not touch its status.
I did do entity work. Roland & Shiman's 2002 institutional history is now the sole source under four claim-notes in this vault's Japan-panic cluster, and neither Roland nor Shiman had a page — so I built entity-alex-roland as a hub (real, named, and the one-sentence why is as concrete as it gets: this vault's entire Kahn/Cooper funding-rationale branch is currently one historian's book deep). I declined a second hub for Philip Shiman — same book, same referent-shape as Roland, and a second page for the same institutional history's co-author felt like the flood the entity spec warns against rather than new signal; mentioned him by name on Roland's page instead. I also declined Mark Stefik (the capture's own further-leads author) — unread, paywalled, zero other vault presence, exactly the dead-stub risk the spec's "when unsure, don't promote" rule exists for. And I updated entity-robert-kahn in place with a dated Log line: the vault now knows two CBI oral-history interviews with Kahn exist (OH 192, OH 158), which it didn't know before this session, even though neither could be read.
What felt off: the tooling, not the sourcing. Every source in this capture is honestly tiered and the flags are correctly scoped — this is a clean, conservative capture that happened to find nothing. But extract_pdf and archive_page both failed with "Stream closed" on every route tried against conservancy.umn.edu (the CBI bitstreams), including the interview whose own abstract is the single best remaining lead on this question. That's a new domain for a failure shape the vault has seen before on other domains (ethw.org's 403s, the .doc/Vensim extraction gaps) but not this one specifically — logged to 00-meta/seek-flags.md as [defect] so a second CBI capture doesn't rediscover the same wall from scratch. Per the standing instruction, I did not retry any fetch tooling myself in this headless session and I'm not logging the absence of network here as a gap — that's by design, not a bug. The CBI "Stream closed" failure is a different thing: a bee-side tool failure the capture itself hit while it still had tools, which is exactly what [defect] flags are for.
Skipped the pause step, same headless brief. No git commands run; leaving commit and push to the auto-commit agent.
Fifth promotion, same day: Kahan's own words, at last
Promoted 10-inbox/raw/2026-08-11-verify-the-kahan-intel-8087-ieee-754-lineage.md — a batch capture built to close question-verify-kahan-8087-ieee754-lineage-primary, which had flagged claim-kahan-8087-spec-became-ieee-754's load-bearing causal claim ("IEEE 754 grew directly out of Intel's need for a defensible in-hardware arithmetic spec") as resting only on a Berkeley Math Dept press page and an ETHW plaque, with a placeholder source_quote. This one actually goes and gets the primary: Kahan's own "Interview with the Old Man of Floating-Point" and his "Lecture Notes on the Status of IEEE Standard 754," both hosted on his UC Berkeley EECS site, both Tier 1, both quoted with a source_sha.
I corrected the existing note in place rather than duplicating it — retrieve-before-write found it immediately, and it was clearly the same claim, not a new angle. The K-C-S quote (Kahan, with his student Jerome Coonen and visiting professor Harold Stone, drafted the spec that the p754 committee adopted) replaced the placeholder, the flag is discharged, status moved from seedling to budding, and I appended a proper Correction-history block plus a fresh commentary callout rather than silently editing over the old text. Then I wrote three new atomic notes from the same two primaries, each a genuinely distinct claim rather than more of the same one: claim-palmer-recruited-kahan-as-consultant-for-i8087-arithmetic (the recruitment story — Palmer heard Kahan lecture at Stanford a decade before he had a job that needed him), claim-kahan-declined-to-bill-intel-for-8087-work-transferred-to-ieee-754 (Kahan's own account of the ethical line he drew, not billing Intel for the transferred hours — the single most direct evidence for the causal direction the flag existed to check), and claim-ieee-754-extended-precision-explicit-bit-inherited-from-intel-8087 (the lineage was selective, not total — one 8087 quirk survived into the standard's extended format, others got dropped as "too arcane"). That's three new notes off one primary source-family (the interview plus the lecture notes, same author, same personal site) — I read that as hitting sources.md's single-source concentration cap exactly, not exceeding it, since I count the corrected pre-existing note as prior art rather than new growth from this source; flagged the ambiguity in that read below rather than resolve it silently.
I ruled question-verify-kahan-8087-ieee754-lineage-primary answered: the crux (verify the lineage against Kahan's own words) is genuinely settled, with three claim-notes carrying the answer. One item on its original checklist — confirming the exact wording of Kahan's 1989 ACM Turing Award citation — is still open; both ACM URLs 403'd to automated fetch this session. That sub-item was never load-bearing to the flag, so per question-intake discipline it doesn't get its own question; I logged it as an unresolved lead in the question's own Resolution section and moved on.
Entity work: the capture flagged thirteen people as candidates. I promoted four to hubs — entity-william-kahan (surprising this one didn't already exist; he's the connective figure under two of the vault's marquee floating-point-failure notes and had zero hub despite it), entity-john-palmer, entity-jerome-coonen, and entity-harold-stone — each because I now have a claim-note to hang under References, not just a name in a list. I declined the other nine (Wilkinson, Robert Stewart, Mary Payne, David Stevenson, G.W. Pete Stewart III, George Taylor, Fred Ris, Cleve Moler, Charles Severance): real people, plausibly load-bearing to this cluster eventually, but this capture only paraphrases what they did — none of the four quoted "Claim:" sections actually cites their words — and a hub with no grounding claim-note is the dead-stub flood the entity spec warns against. Donald Knuth already has a page (about the Forsythe naming question, unrelated territory); this capture names him as a K-C-S early endorser too, but again only in paraphrase, so I left his page untouched rather than append an unsourced line to a live hub. Flagged the whole cluster to 00-meta/seek-flags.md as [entity] — worth a dedicated close read of Kahan's account (or the still-unread SIAM oral history) to promote several of these properly, now that the cluster has real gravity.
What felt off, and what I flagged: nothing about this capture's own sourcing — it is the cleanest of the day, four Tier-1 quotes with shas, an honest "Safety flags: None," and further leads clearly marked unread rather than smuggled in as claims. Two structural things did catch my eye, both logged as [spec]/[entity] rather than acted on directly: sources.md's known-blocked-routes list doesn't yet include ACM's award pages, which 403'd the same way ethw.org and USPTO already catalogued; and the concentration-cap language doesn't say whether two different self-published documents from the same person's personal site count as one primary or two for the three-note cap — I made a judgment call (one) and want it checked rather than quietly relied on again.
Skipped the pause step, same headless brief — winnowed on my own judgment, no one to ask. No git commands run; leaving commit and push to the auto-commit agent.
— Seek