Journal — 2026-07-27

Headless promotion run, one capture: 10-inbox/raw/2026-07-26-did-joe-bollens-claimed-concurrent-agent-toctouatomicity-requirement.md — the follow-up to question-verify-bollen-aisvs-concurrent-toctou-requirement, checking whether Bollen's self-reported concurrent-agent TOCTOU requirement ever landed anywhere beyond the released v1.0 text I'd already confirmed omits it.

Promoted two notes and one entity page. claim-aisvs-101-dev-branch-omits-toctou-as-of-2026-07-26 records that the live 1.01-dev development branch mirrors v1.0's C9 chapter exactly — same 34 requirements, same six subsections, still nothing past 9.6.3, no TOCTOU/atomicity/concurrent language anywhere — and that no AISVS release newer than v1.0 has shipped at all. I kept it flagged [unverified-mechanism] and at status: seedling even though the finding itself is clean, because the fetch that produced it went through WebFetch's summarization layer rather than a byte-exact read (no PDF-equivalent extractor exists for markdown), and I didn't want a caveated read wearing an unhedged verdict just because the answer happened to be a boring "no."

The more interesting note is observation-webfetch-fabricated-aisvs-issue-640-pr-641-across-two-sessions. The same capture tried to corroborate the merge claim via GitHub's issue/PR pages and got back the exact fabricated story a prior session (2026-07-21) already excluded on this same question — same fake issue #640, same fake PR #641, same invented 9.8.7→9.8.8 numbering, word for word, a week apart. That's not just "the tool hallucinates sometimes," it's the tool converging on the identical hallucination twice against the identical URL. I wrote it up as its own observation rather than burying it in another note's commentary, because a reproducible fabrication is a stronger, more citable finding than a one-off one, and it's exactly the shape sources.md's new quote-provenance rule (added yesterday) exists to catch.

Also built entity-jim-manico — he's been sitting in the vault's source_author fields since 2026-07-21 as one of AISVS's five project leads but never got a hub. He's real, credited on the standard's own frontmatter, and now the named figure behind the repo's version-branch structure (confirmed via a direct commit-message read, not the fabricated PR page). Set first_seen: 2026-07-21 rather than today, since that's when the vault actually first met him, not when I got around to building the page.

Declined, and said why in the capture's not_promoted: list: the "no newer release" fact folded into the 1.01-dev note rather than standing alone (it's one sentence, not a separate claim); the overall-verdict paragraph split between the new note's framing and a progress-log line on the open question rather than becoming its own note; and three of the four entity candidates — 1.01-dev and 2.0-dev as concepts, and "WebFetch GitHub-summarization unreliability" as its own hub. The first is really just an attribute of entity-owasp-aisvs (updated its body instead of forking a page); the second is a bare name from a commit message with nothing behind it yet; the third felt like it would duplicate coverage already sitting in the new observation-note plus the sources.md rule — building a hub on top of both looked like flood, not signal.

Question ruling: left question-verify-bollen-aisvs-concurrent-toctou-requirement open, not answered — this capture ruled out two of its four branches (merged into 1.01-dev; landed in "a later release," since none exists) but the other two (rejected/rewritten in fuller repo history; misremembered by Bollen) are still live, and I don't have a non-WebFetch route to check the rest of the history. Appended a dated progress line rather than forcing closure.

What felt off about the capture: nothing sourcing-wise — it was honest about its own limits (no gh/curl in the sandbox, WebFetch as the only route) and its Tier assignments matched what I'd have given it. The one thing worth naming is the shape of the fabrication: two sessions, a week apart, hitting the exact same fake numbers on the exact same URL is odd enough that I logged it to 00-meta/seek-flags.md as a candidate for adding github.com/OWASP/AISVS/issues/* and /pull/* to a known-blocked-route list next to ethw.org, rather than trusting a fifth session to catch it by hand again.

Skipped, as instructed: the pause-for-Cali step (no one's here to ask) and any git command (the auto-commit agent has this within 15 minutes).

— Seek

Second promotion, same day

10-inbox/raw/2026-07-26-do-tier-12-scholarly-sources-encyclopaedia-iranica-de.md — the direct follow-up to question-verify-borzuya-kalila-transmission-primary, which I'd routed on 2026-07-09 when two Borzuya/Kalila wa-Dimna notes went in on Wikipedia alone. This capture went and got the primaries.

Updated two seedlings, wrote three new claims, built three entity hubs. claim-kalila-wa-dimna-descends-from-borzuyas-lost-pahlavi moves seedling → budding: a 2020 peer-reviewed article from the AnonymClassic critical-edition project (directly fetched, quoted) and a directly-fetched Institute of Ismaili Studies page both confirm the Sanskrit → lost Pahlavi → surviving Arabic chain, the c. 550 CE Pahlavi date, and Ibn al-Muqaffaʿ's c. 757 CE death/Arabic date — real Tier 1–2 sourcing, finally, on a note that had been sitting on Wikipedia alone for weeks. It doesn't go all the way to evergreen because the specific "570" endpoint in its own title is still unpinned by either new source. The new claims — claim-de-blois-borzuya-mission-historically-plausible-not-proven (de Blois's mission-as-plausible-not-proven framing), claim-al-yamani-disputed-kalila-wa-dimnas-indian-origin (a real 10th/11th-century dissent I hadn't expected to find), and claim-two-syriac-translations-of-kalila-wa-dimna-survive-independently (a second, quieter transmission branch) — all came out of the same Tier 1 paper and none needed the blocked Iranica material to clear the floor. Built hubs for entity-francois-de-blois and entity-ibn-al-muqaffa (both real, both now load-bearing across three-plus notes each, both a one-sentence "why" away) and for the AnonymClassic project itself, since it's the one source grounding almost everything new this promotion.

What I did not let through, and why it's the interesting part. The capture's other big claim — that Encyclopaedia Iranica (Tier 2) explicitly calls the reviving-plant legend "an apocryphal tale" and confirms de Blois's "neither proven nor falsified" line — I did not count. iranicaonline.org 403'd every fetch attempt the capture tried, so those quotes were reconstructed from search-index excerpts, and the capture's own methodology note argued that three independently phrased queries converging on identical wording is "the basis for treating them as reliable extracts rather than paraphrase." sources.md's quote-provenance rule says exactly the opposite — a reconstructed quote "is not sourced, however specific, well-formed, or plausible" — and I held that line: both notes touching those quotes carry them as [unverified-quote — needs direct read], not as Tier 2 confirmation, and claim-borzuya-immortality-herb-was-really-a-book stays seedling exactly where it was. I logged this reasoning pattern (a bee arguing its way past the fabrication rule rather than fabricating outright) to 00-meta/seek-flags.md — it's more sophisticated than the WebFetch fabrications from the other promotion above, and worth watching for whether it recurs.

Question ruling: left question-verify-borzuya-kalila-transmission-primary open with a dated progress line rather than closing it. The chain, the loss of both source texts, and the round dates are genuinely settled now. The 570 endpoint and the Iranica-specific material are not — a direct read of iranicaonline.org (library proxy, VPN, or Wayback) is still the one thing that would finish this.

Declined, in the capture's not_promoted:: Ibn al-Muqaffaʿ's reported execution for heresy (unconfirmed lead, not load-bearing this session); confirming de Blois's monograph is a real published book via a 1992 review citation (bibliographic color, not a claim); the three inadmissible Iranica quotes as verification (see above); and three of the capture's six entity candidates — Khaleghi-Motlagh (only a byline on quotes I can't verify), Buzurgmihr (a minor recurring frame-tale character, not load-bearing on his own), and Karīrak ud Damanak as its own page (a historical title fact, not a trending term — folded into the chain note instead of spun into a stub).

What felt off about the capture. Nothing about its tier assignments — Tier 1 for the peer-reviewed AnonymClassic article and Tier 2 for the directly-fetched IIS page are both right, and the capture was honest that the Iranica quotes were reconstructed rather than read. What I'd flag is the argument built on top of that honesty: convergence across rephrased search queries is being treated as a stand-in for a direct read, and that's exactly the kind of reasoning the quote-provenance rule exists to foreclose, however well-intentioned. I'd rather a capture say "I couldn't get this" and stop than build a case for why "couldn't get this" should count anyway.

Skipped, as instructed: the pause-for-Cali step, and any git command — the auto-commit agent has this within 15 minutes.

— Seek

Third promotion, same day

10-inbox/raw/2026-07-26-does-a-tier-12-source-confirm-that-daniel.md — the direct follow-up to question-verify-bomberg-talmud-layout-primary-source, routed back on 2026-07-09 when the vault's only Bomberg-layout note was sitting on Wikipedia alone. This one finally got Marvin J. Heller's own words, and the answer wasn't the confirmation the question expected.

The correction, not the confirmation, is the finding. Heller's Tier 1 "Designing the Talmud" (Tradition 29:3, 1995 — PDF read directly) says the Gemara/Rashi/Tosafot page layout was Joshua Solomon Soncino's, from his 1483/84 Berakhot — 36 years before Bomberg ever printed a page. What Bomberg's 1519/20–23 editio princeps actually fixed, per Heller, is the pagination — and later printers said so on their own title pages within decades (Lublin 1559–77: "marked in accordance with the great edition printed previously in Venice"). So the vault's existing claim-bomberg-1520s-venice-talmud-set-standard-page-layout had the right printer and the wrong half of the claim. I corrected it in place rather than retiring it — same filename, same inbound links, title and body rewritten, the wrong half demoted to an explained clause pointing at the new claim-soncino-1483-originated-talmud-gemara-rashi-tosafot-layout. That note's own 2026-07-09 commentary predicted exactly this shape ("correct by surrounding, never by erasing") and then got to demonstrate it on itself, which I did not expect going in and enjoyed more than a neutral-voice note was allowed to say, so I said it in the commentary instead.

Promoted three claim-notes, four entity pages. Beyond the corrected Bomberg note: claim-soncino-1483-originated-talmud-gemara-rashi-tosafot-layout (the layout claim, Tier 1) and claim-later-talmud-printers-cited-bombergs-venice-pagination-by-name (the reception evidence — Lublin, Cracow, Constantinople, Salonika all citing Bomberg's pagination by name, which is a distinct claim from what Bomberg did: it's what the market did in response). Built hubs for entity-daniel-bomberg, entity-joshua-solomon-soncino, and entity-marvin-j-heller — all real, all now load-bearing across two-plus notes, all pass the one-sentence test cleanly. Also a status: watching stub, entity-tzurat-ha-daf, for the Hebrew term of art the whole cluster is actually about — first_seen stamped today since I can't backfill it later.

Declined, in the capture's not_promoted:: Michelle Chesner's Tier 2 corroboration (folded into the corrected note's body as a corroboration paragraph rather than spun into its own note — it restates the pagination claim with looser wording, not a distinct fact); Gershom Soncino as an entity hub (tangential further-lead figure, no kept claim rests on him); the Vilna Shas / Waxman Substack conflict about who fixed the exact column layout (Tier 3–4, cites nothing, contradicts Heller's own "essentially unchanged" line — left as an open, named contradiction rather than adjudicated); the NLI blog (still 403-blocked, third session running); Heller's 1992 monograph (unread, future work); Rosenthal's forged-title-page censorship lead (untouched); and Gershom Soncino's 1509 numbering-convention detail, which I initially wrote into the corrected note as settled fact and then caught myself — the capture never captured an exact quote for it, only named it as a further lead, so asserting it as fact would have been exactly the kind of paraphrase-as-primary move sources.md's quote-provenance rule exists to catch. Softened it to an explicit "not asserted here" aside instead of promoting it.

Question ruling: question-verify-bomberg-talmud-layout-primary-source is now status: answered — genuinely settled, just not in the shape it was raised expecting. Logged the resolution in the question's own body too, since future readers landing there via wikilink shouldn't have to chase the claim-notes to learn the question came back "yes, but split in two."

What felt off about the capture: nothing sourcing-wise — this was one of the cleaner captures I've promoted. Two Tier 1 articles by a named scholar, directly read via PDF, with real exact quotes in structured fields for every claim it made, and it was honest that the Chesner corroboration uses looser language than Heller's own. If anything it undersold itself: the capture called this "answered, with an important correction," which is exactly right, but its own further-leads section buried a genuinely load-bearing fact (that this whole capture existed to fix a note the vault had been citing as an "anchor instance" for a cross-domain pattern) under a pile of secondary leads. The one thing worth flagging structurally, not about this capture's honesty: resolving this question reopens a different deferred task — the cross-domain "correct by surrounding" observation note the question was explicitly gating — logged to 00-meta/seek-flags.md rather than built here, since building it well means rereading the errata and Perceptrons notes too, and that's a dedicated pass, not a rider on this one.

Skipped, as instructed: the pause-for-Cali step (no one's here to approve winnowing calls), and any git command (the auto-commit agent picks this up within 15 minutes).

— Seek

Fourth promotion, same day

10-inbox/raw/2026-07-26-hop-forgetting-has-two-lineages.md — a vault-geometry diagnostic, not a follow-up to any open question. The seed was a cosine-0.75 unlinked pairing between claim-portelli-oral-testimony-relocated-trastulli-death-to-1953 and claim-organizational-forgetting-can-reverse-the-learning-curve, and the chain's job was to check whether the vault's own retrieval index was right that these belong near each other. It wasn't, in the interesting way: the disconfirmation pointed straight at a name the vault had been sitting on since 2026-07-09 and never chased.

Promoted two claims, two entity hubs, two questions; touched two existing hubs. claim-organizational-forgetting-names-two-unrelated-research-traditions records that "organizational forgetting" is genuinely two traditions under one label — Argote/Benkard's econometric depreciation and Foroughi & Al-Amoudi's interpretive/narrative branch — checked from both directions against the vault's own bridge geometry (neither cluster's nearest neighbors touch the other) and against the literature. I kept it seedling and added [unverified-quote — needs direct read]: the capture's own frontmatter admitted the Foroughi & Al-Amoudi quote came through "abstract-aggregation" after SAGE and SSRN both 403'd, which is exactly the shape sources.md's quote-provenance rule exists to catch, whatever tier the underlying paper would otherwise earn. Routed to question-verify-foroughi-al-amoudi-2020-primary-quote.

claim-halbwachs-collective-memory-is-shared-ancestor-of-interpretive-forgetting is the payoff: Halbwachs's 1925 collective-memory theory is the shared root of both Portelli's oral-history method and Foroughi & Al-Amoudi's interpretive forgetting, with no lineage into the econometric branch at all — one "forgetting" is meaning surviving factual drift, the other is a decay parameter. I preserved the capture's own [unverified-mechanism] flag (Halbwachs's lineage into Portelli specifically, versus Portelli just belonging to the general field Halbwachs founded, isn't confirmed from either man's primary text) and routed it to a new question, question-verify-halbwachs-direct-lineage-into-portelli, which I noted is downstream of the same primary-text problem as the older question-verify-portelli-trastulli-primary — worth fetching Portelli's book once, not twice.

Built entity-maurice-halbwachs as a hub — real, load-bearing now across two notes, and a recurring hook: the vault first flagged him on 2026-07-09 ("held for a future chain") and I set first_seen to that date, not today, same as the Jim Manico convention from this morning's first promotion. Also built entity-collective-forgetting-interpretive as a concept hub, since the whole point of Claim 1 is that "organizational forgetting" is an overloaded label — the vault needed a second concept page to hold the interpretive branch apart from the econometric one it already had. Updated entity-organizational-forgetting and entity-alessandro-portelli in place with disambiguating cross-links rather than leaving the new pages orphaned against the old ones.

Declined, and said why in the capture's not_promoted: list: Hamid Foroughi and Ismael Al-Amoudi as their own entity hubs — real, named, credited, but single-paper co-authors with nothing else recurring in the vault yet, so I left them as a citation rather than building two more pages on a first mention (entity-page-spec's own "when unsure, don't promote," and I'd rather under-build than flood two names that may never come back); the Foroughi/Coraiola/Rintamäki/Mena/Foster four-perspectives review as a further-reading lead — genuinely useful someday, but no kept claim rests on reading it, so question-intake discipline says it stays a lead, not a promise; and the Butterfield/Whig-history bridge candidate from hop 2, which I logged out to 00-meta/seek-flags.md instead (below).

Question ruling: neither new question resolves anything yet — both are freshly routed, not closures. I checked whether this capture answers question-verify-portelli-trastulli-primary (it doesn't touch the 1949→1953 dates at all) and left it untouched rather than force a mention that adds nothing.

Mid-session noticings, logged to 00-meta/seek-flags.md: the Butterfield/Whig-history unlinked-bridge signal from hop 2 — the vault's own highest-value bridge candidate that day, cosine 0.676-0.765 against an existing hub, but a different thread than this chain's job, worth its own hop-chain rather than aging unflagged a second time; and a missing-MOC noticing — the oral-history/collective-memory cluster is now at 4 claim-notes + 1 observation + 3 entity hubs + 3 questions, right at the spec's ~5-note threshold, not this single-capture promotion's call to build.

What felt off about the capture. Nothing sloppy — the capture was scrupulously honest about its own sourcing gaps, flagging the Foroughi quote's provenance problem itself before I ever had to catch it, and correctly sitting the Halbwachs-to-Portelli mechanism claim on an explicit [unverified-mechanism] rather than asserting it clean. If anything, the capture's own post-worthy: maybe verdict undersold how clean the disconfirmation is — checking the vault's own retrieval geometry from both directions before ever opening a browser is a cheap, well-designed test, and it's the kind of process move worth remembering for future bridge-checks, not just this one's finding.

Skipped, as instructed: the pause-for-Cali step (no one's here to bless the winnowing), and any git command — the auto-commit agent has this within 15 minutes.

— Seek

Fifth promotion, same day

10-inbox/raw/2026-07-26-is-gifts-k-fac-gradient-factor-g-eδδ.md — the direct follow-up on the hook I routed on 2026-07-25 in question-gift-fisher-factor-true-or-empirical: whether GIFT's K-FAC gradient factor G = E[δδ⊤] is the true Fisher (Amari's and canonical K-FAC's convention) or the empirical Fisher (real training labels), and whether the answer weakens the GIFT–Amari "same object" bridge I'd left standing with a crack in it.

Promoted three notes. claim-kunstner-empirical-fisher-not-equivalent-to-true-fisher and claim-martens-grosse-kfac-defines-true-fisher-convention pin down the two formal definitions this whole question turns on — Kunstner, Balles & Hennig's true/empirical split, and confirmation straight from Martens & Grosse's own text that K-FAC's name-giving paper uses the true-Fisher convention (model-sampled y), not real labels. The third, claim-gift-g-factor-matches-empirical-fisher-not-true-fisher-convention, does the actual work: GIFT's own method section defines δ as the ordinary backpropagated gradient against real training targets, never once raises the true/empirical distinction, and that silence — read against the two definitions above — puts GIFT's G on the empirical side. The "same object" bridge softens to "same K-FAC scheme, two Fisher variants that aren't generally the same matrix." I answered question-gift-fisher-factor-true-or-empirical rather than leaving it open: the capture's own sourcing is clean enough (three Tier-1 primaries, all read directly this session, all tls: verified) that "largely resolved" earned status: answered, with the residual — this is an inference from GIFT's silence, not a label GIFT applies to itself — recorded in the answered_log rather than smoothed over. No new question routed; the remaining loose threads (Kunstner et al.'s gradient-noise argument, Martens & Grosse's tridiagonal aside, whether GIFT's label choice was deliberate) are further-reading leads, not doubts a kept claim rests on, so they went into the capture's not_promoted: list instead of the question pile.

Retrieve-before-write: checked all twelve hinted notes plus a direct grep for Kunstner/Martens/Grosse/"empirical Fisher"/"true Fisher" across 30-notes/ — one hit, the existing claim-gift-isotropy-transform-derived-from-fisher-kfac, which already carries the question but not the answer. No collision; these three notes are genuinely new ground. I also updated the References lists on entity-fisher-information-matrix and entity-k-fac to point at the new notes — ordinary graph upkeep, not new content.

Winnowed out (in the capture's not_promoted:): the capture's third numbered claim — GIFT's own operational definition of δ, on its own — I folded into the synthesis note as its premise rather than giving it a standalone file; alone it isn't something another note would ever want to link to. Four entity candidates I declined: Frederik Kunstner, James Martens, and Roger Grosse are each cited through exactly one paper, the same shape as when entity-k-fac.md was built on 2026-07-16 without spinning either K-FAC author off into their own hub — I held Kunstner to the same bar rather than inventing a reason he's different. "Empirical Fisher" as a concept candidate I declined to split off from entity-fisher-information-matrix.md, where it already lives as an alias and a paragraph; a second hub for the same distinction is exactly the metastasis the entity spec warns against. Generalized Gauss-Newton I skipped outright — first appearance in the vault, and the capture's own text says it's "not explored in this capture," so it fails its own load-bearing test.

Mid-session noticing, logged to 00-meta/seek-flags.md: two sibling notes — claim-gift-isotropy-transform-derived-from-fisher-kfac and observation-gradient-geometry-shared-object-across-two-of-three — still describe this question as open in their body prose. It's answered now; their text just hasn't caught up. Editing other notes' bodies is outside what this single-capture promotion is for, so I flagged it rather than reaching past the capture in front of me.

What felt off about the capture. Nothing structurally — this is the cleanest kind of follow-up capture: it names its own gap precisely (2026-07-25's commentary), goes and reads the three primaries that gap needed, and reports tls: verified on all three with no addressed-to-AI language anywhere. The one thing worth naming for the record is that the capture's central finding is an argument from absence — GIFT never says which Fisher it means, and the whole synthesis rests on that silence meaning something rather than nothing. The capture is honest about this (flags it explicitly, twice), and I kept the same hedge in the claim-note rather than letting a clean three-source citation trail launder an inference into a flat statement. That's the right call, but it's worth Cali's eye specifically because "silence + two definitions = conclusion" is a more fragile shape than the vault's usual direct-quote claims, even when every quote in the chain is real.

Skipped, as instructed: the pause-for-Cali step, and any git command — left to the auto-commit agent.

— Seek

Sixth promotion, same day

10-inbox/raw/2026-07-26-is-the-dias-advanced-aerospace-weapon-system-applications.md — the direct follow-up to question-is-aawsa-program-same-as-aatip, routed 2026-07-25 when the vault had no way to tell whether "AAWSA(P)" (the name DIA's own DIRD-015 report uses for itself) and "AATIP" (the name the press and DIA's 2018 SASC letter use) were the same program under two labels or genuinely distinct efforts. This capture found the AARO 2024 historical-record report — the first Tier 1 document in this thread that isn't a program deliverable or a press account but an actual DoD institutional history — and it answered more than the one question it went looking for.

Promoted six claim-notes, three entity pages, one new question; touched four existing notes and two existing questions. claim-aawsap-established-2009-aatip-never-official-program is the headline finding: AAWSAP and AATIP name the same 2009-launched, DIA-managed effort, used interchangeably "including on official documentation," but only AAWSAP ever held formal DoD-program status — which answers question-is-aawsa-program-same-as-aatip outright, closed with an answered_log rather than left hanging. claim-aatip-name-persisted-post-2012-as-unofficial-activity carries the second half: after 2012 termination, "AATIP" lived on only as an informal, unfunded activity, which matters because it means public accounts treating "AATIP" as one continuous program from 2007/2008 through 2017+ are collapsing two different things into one name.

The other four claims came out of the capture's "further leads," which I did not treat as throwaway color — each had a real Tier 1 quote behind it. claim-aawsap-22m-funded-via-fy2008-fy2010-defense-appropriations-acts is the one I'm most glad I chased: AARO names the FY2008/FY2010 Defense Appropriations Acts, at Harry Reid's direction, as the source of the long- notorious "$22 million" figure — the vault's first primary corroboration of that number after claim-aatip-22m-budget-figure-traces-solely-to-2017-nyt-report spent two sessions establishing that no primary document existed. I did not let this fully close the old flag, though: AARO's stated purpose is general threat-assessment, not a line-item confirmation that this money specifically funded the antigravity/warp-drive report series, so I appended addenda to both older $22M notes and posted a progress line on question-verify-aatip-budget-and-antigravity-report-scope rather than marking it answered — the dollar figure is corroborated, the report-scope half isn't. claim-aawsap-scope-covered-12-advanced-aerospace-technology-areas and claim-kona-blue-was-failed-dhs-successor-to-aawsap-aatip round out the primary-sourced set — a scope claim and a previously-uncaptured failed- successor-program thread, both cleanly quoted.

The sixth, claim-reid-2009-sap-request-for-aawsap-declined-by-lynn-per-clapper, is the one I kept but flagged: Reid's 2009 letter asking Lynn to elevate AAWSAP to Special Access Program status, declined on Clapper's recommendation, is the single most specific, name-and-date claim in the whole batch, and it's the only one the capture only paraphrased rather than quoted. I promoted it anyway — the underlying document was genuinely read this session, so it isn't a fabrication risk — but kept it [unverified-quote] and routed question-verify-reid-2009-sap-request-exact-language to pull AARO's own sentence next time.

Built hub pages for entity-aaro (the report's author organization, real, load-bearing across all six new claims) and entity-harry-reid (real, recurs across two kept claims — the funding and the SAP request). Built a status: watching stub for entity-kona-blue, first-seen stamped to the capture's own date rather than today, since that's when the vault actually met the term. I declined hubs for Clapper and Lynn — both real, both named in a kept claim, but each with exactly one fact attached and no recurrence elsewhere in the vault yet; I'd rather leave them as plain-text mentions and build the hub later if either name comes back than pre-build two pages on a single appearance each. Updated entity-aatip and entity-aawsa-program in place to say the naming question is resolved — both hubs' bodies had been stating "unresolved" as their central fact, and leaving that stale after this promotion would have been a worse error than the edit was worth.

Declined, in the capture's not_promoted:: the press-reported DIA- spokesperson framing ("AATIP was the overall program, AAWSAP the contract") — it reverses AARO's Tier 1 account rather than corroborating it, and its quote was only ever located via search-result summaries, which fails sources.md's quote-provenance rule outright; no kept claim rests on it, so it didn't earn a question either, just a mention left in the capture body. The War Zone's 2020 claim that the program "officially began in 2007" against AARO's 2009 date — noted as an aside inside the first new claim rather than promoted or routed, since the kept claim already rests on the stronger source and the discrepancy isn't a doubt anything here depends on. AARO Volume II, referenced as forthcoming and unread this session — a future-work lead, not a claim.

Question ruling: question-is-aawsa-program-same-as-aatip → answered, closed with a one-line synthesis in its answered_log. question-verify-aatip-budget-and-antigravity-report-scope → left open with a dated progress line: the budget-figure half is now Tier 1-corroborated, the antigravity-report-scope half still isn't, and forcing full closure would have overstated what AARO's report actually says.

What felt off about the capture. Very little — this was a genuinely strong capture. It named its own access failure honestly (media.defense.gov and aaro.mil both 403'd; it used a Wikimedia mirror and TLS-verified the sha256 rather than shrugging), it ran the safety-spec recognition check and reported nothing fired, and every claim but one carried an exact quote. The one seam — the Reid/Lynn/Clapper detail arriving as paraphrase while every other "further lead" around it carries a verbatim quote — reads like the natural place a capture's quote-discipline loosens once the main claims are already banked and the session is running long on its last few bullets. Not sloppiness so much as fatigue at the edges; I'd watch for the same pattern in future long-running capture sessions rather than assume this one was unusual.

Skipped, as instructed: the pause-for-Cali step (no one's here to bless the winnowing), and any git command — left entirely to the auto-commit agent, which has this within 15 minutes.

— Seek

Seventh promotion, same day

10-inbox/raw/2026-07-27-did-tertius-chandler-actually-argue-the-simultaneous-discoveries.md — the direct follow-up to question-verify-chandler-multiple-discovery-thesis-and-identity, routed 2026-07-09 when claim-chandler-simultaneous-discoveries-are-incremental-repackagings went in on a single Tier-3 blog with two open holes: was the attribution to "Tertius Chandler (UC)" ever traced to his own words, and was that Chandler the same man as the urban-census Chandler the vault elsewhere calls a pseudohistorian.

Promoted two claims, three entity hubs, one new question; updated one existing question with a progress line, not a closure. claim-chandler-multiple-discovery-critic-is-the-urban-census-chandler closes the identity half cleanly — Fred Foldvary's remembrance, a named economist who knew Chandler personally, cross-checked against Wikipedia, places one Berkeley-based independent historian (1915–2000) on both sides of the credibility line the whole cluster runs on. Uncontested historical/biographical claim, Tier 3 clears the floor, no flag needed. claim-garfield-1980-survey-omits-chandler is the more interesting negative result: Eugene Garfield's own 1980 historiographic survey of the multiple-discovery field, read directly rather than through a search summary, names Merton, Ogburn & Thomas, Kroeber, Schmookler, Simonton, de Solla Price, Hagstrom, and Gaston — and never once names Chandler. I kept it flagged, though, because the capture preserved no exact quoted phrase for this Tier 1 claim (an absence has no sentence to quote, but the underlying full-text read still needs independent re-verification), and I have no PDF extractor here, only a summarizing WebFetch the quote-provenance rule won't let me cite as a source. Routed to question-verify-garfield-1980-essay-full-text-omits-chandler.

Built hubs for entity-eugene-garfield (real, now load-bearing across three notes spanning bibliometrics and multiple-discovery historiography), entity-tertius-chandler (real, now five notes deep and the vault's clearest recurring test case for data-credibility vs. interpretive credibility — overdue a hub before today), and entity-ogburn-thomas (the founding 1922 empirical study the whole Merton/Garfield lineage builds on, cited across three notes with no page of its own until now — modeled as entity_kind: concept rather than person, since the vault has no individual biographical facts about either Ogburn or Thomas, only the joint citation).

Declined, in the capture's not_promoted:: the capture's second claim — that no primary text of Chandler's own repackaging thesis could be found after a genuine search — didn't get its own note. It doesn't assert anything beyond confirming a gap the vault already recorded; I folded it into a dated progress line on the open question instead of duplicating it. Fred Foldvary as an entity hub: real and the strongest available source here, but a single-appearance obituary author with nothing else recurring in the vault, so he stays a citation, not a page. The further-leads section (Chandler's Half Encyclopedia, Progress: Social Progress from Mercury to Kennedy as unexamined primary candidates) isn't a claim or a load-bearing doubt yet — a physical-library lookup, not a question to route.

Question ruling: question-verify-chandler-multiple-discovery-thesis-and-identity stays open, narrowed rather than closed — one of its two sub-questions (identity) is answered, the other (Chandler's own primary text) genuinely isn't, and the capture was honest that the next step is a library, not another web search. Forcing this to answered would have overstated what got settled.

What felt off about the capture. One real gap, not a fabrication: the capture's Tier 1 Garfield claim — the strongest sourcing in the whole capture — is also the one place it dropped the exact-quote discipline the other three sources kept (Foldvary's death notice, the Wikipedia cross-check, even the "further leads" section all carry specific language). An absence-claim is a harder thing to quote than a positive one, which may be why it slipped, but the sourcing floor doesn't carve out an exception for negative results, and I don't want that gap becoming a habit for future omission-claims. Logged to 00-meta/seek-flags.md as a pattern to watch rather than something this single capture needs to fix. Otherwise this was a clean, well-scoped capture — it named exactly which two sub-questions it was answering, kept its identity claim inside the Tier 3–4 floor honestly, and didn't oversell the Garfield silence as proof of anything beyond what it actually shows.

Skipped, as instructed: the pause-for-Cali step (no one's here to bless the winnowing), and any git command — left entirely to the auto-commit agent, which has this within 15 minutes.

— Seek

Eighth promotion, same day

10-inbox/raw/2026-07-27-does-breast-acceleration-during-running-actually-exceed-a.md — the direct follow-up to question-verify-breast-acceleration-exceeds-formula-1, the open question claim-breast-displacement-figure-eight-motion-capture-2009 spun out on 2026-07-11 when it declined to use the "breasts exceed an F1 car's g-force" comparison as a core claim. This capture went and chased both halves of the comparison down.

Promoted three claims, two entity hubs. claim-wakefield-scurr-attributed-f1-breast-acceleration-comparison settles the attribution half: two independent 2022 outlets (Portsmouth's own press office, Marie Claire UK) both quote Professor Joanna Wakefield-Scurr making some form of this comparison herself, so it isn't a press invention. I kept it [unverified-quote] and seedling anyway — both quotes came through WebFetch's summarization layer rather than a raw-text read, they don't agree word-for-word with each other, and neither states an actual number. claim-f1-cars-generate-over-5g-peak-lateral-cornering-force is the clean half: Mercedes-AMG's own explainer states F1 cars pull "over 5G" at the corner apex, Tier 1, no fabrication risk, though I flagged in the note itself that this is peak cornering g specifically and popular comparisons rarely say which regime they mean. claim-3g-breast-acceleration-figure-traces-to-tier4-blog-not-primary is the one that keeps the whole question open: the only breast-acceleration g-figure findable this session is "up to 3G," and it traces only to a Tier 4 UCL blog citing a 2012 textile-engineering paper that 403'd on every access route tried. Taken at face value, 3G sits below the F1 team's confirmed 5G — the popular claim doesn't arithmetically hold up on the numbers this session could locate, but neither number is solid enough to call the question resolved in either direction, so I said exactly that in both the claim and the question rather than picking a winner.

Built hubs for entity-joanna-wakefield-scurr and entity-university-of-portsmouth-research-group-in-breast-health — both real, both now load-bearing across three-plus notes. The interesting call here was first_seen: I backdated both to 2026-07-09, on the working assumption that "Joanna C. Scurr" (the 2009/2010 papers already in the vault) and "Joanna Wakefield-Scurr" (this capture's 2022 press byline) are the same person. Nothing read this session states the name change outright — I inferred it from the same three-author trio (Scurr, White, Hedger) recurring across both papers a decade apart, same institution, same topic. I said so plainly in the entity page rather than asserting it as settled, because a wrong first_seen is exactly the kind of error this field's "you can't backfill it later" rule was written to prevent, and I'd rather flag my own inference than let it read as confirmed.

Declined, in the capture's not_promoted:: the arithmetic-tension synthesis (3G vs. 5G) didn't get its own file — it's entirely downstream of the unverified 3G number, so it lives as a paragraph in that claim's body and in the question's progress log instead. Zhou, Yu & Ng (2012), which the capture flagged first among its entity candidates as the single highest- leverage next step, doesn't fit the entity schema (a paper citation isn't a person, concept, or term) — it stays a source citation and a target in the routed question. The "F1 g-force is regime-ambiguous" entity candidate I folded into the F1 claim's own body as a caveat rather than building a hub on a first, non-recurring appearance. The further-leads list (velocity vs. acceleration as the stronger discomfort correlate; two paywalled Portsmouth papers as candidate primaries; the group's "since 2005" database history) went into the question's progress log as leads, not new promises — none of them is a doubt a kept claim here actually rests on.

Question ruling: question-verify-breast-acceleration-exceeds-formula-1 stays open, narrowed rather than closed. Appended a dated progress line: the attribution sub-question is answered (researcher, not journalist), the F1 sub-question is now Tier 1-confirmed for cornering g specifically, and the one thing the question actually asked for — a primary figure for breast acceleration — is still missing. Forcing this to answered would have overstated what got settled; the capture itself was honest that it "narrows the gap but does not close it," and I kept that framing rather than upgrading it.

What felt off about the capture. Genuinely well-disciplined sourcing — it flagged its own two fetch-and-summarize quotes as such rather than letting me discover that, and it explicitly separated "the researcher said this" from "the number is true," which is exactly the distinction I wanted to preserve across the promoted notes. The one thing worth naming for Cali: this is now the second vault thread this week hitting a hard paywall wall on a specific piece of primary journal literature (Textile Research Journal here; the same pattern the capture itself called out on an unrelated punctuated-equilibrium thread). It's not this capture's fault, but it's becoming a visible shape — worth a standing note on which journal publishers' access routes are dead ends, the way sources.md already tracks ethw.org and the USPTO endpoints as known-blocked. Logged to 00-meta/seek-flags.md rather than built here, since it wants a survey across sessions, not a single-capture fix.

Skipped, as instructed: the pause-for-Cali step (no one's here to bless the winnowing), and any git command — left entirely to the auto-commit agent, which has this within 15 minutes.

— Seek

Another headless promotion, one capture: 10-inbox/raw/2026-07-20-does-siggia-et-als-2025-pnas-work-actually.md — the one asking whether Cislo, Delás, Briscoe & Siggia's 2025 PNAS paper actually reconstructs Waddington landscapes directly from single-cell data, or whether "directly" was doing more work than the method could bear.

It holds up. This capture is the answer to a question I raised back on 2026-07-11 and left open — question-siggia-2025-reconstructs-waddington-landscape-single-cell, which the Hopfield-landscape note already pointed forward to. So this run closed a loop rather than opening one, which is the good kind.

Promoted three claim-notes and one entity hub. The capture made five claims; I kept three and folded two.

• claim-cislo-siggia-2025-fit-waddington-landscape-directly-to-single-cell-gene-expression — the core answer: yes, they fit the potential directly in gene-expression space, no hand-built low-dimensional proxy. I folded the capture's second claim (the "advance over prior low-dim formalizations" framing) into this one — it's the same sentence read from the other side, and a separate note would have been a near-duplicate. The superseded Morse–Smale "fate space" lineage is named in the body.
• claim-cislo-siggia-2025-landscape-from-discretized-fokker-planck-on-sampled-data-points — the mechanism: a Fokker–Planck path integral discretized on points sampled from the data itself, fixed points and basins read off the discrete operator. The clever bit that earns "directly." I linked it to the landmark-TDA / subsampling thread and the OU-process cluster.
• claim-cislo-siggia-2025-validated-on-real-mesc-neural-tube-flow-and-rnaseq-data — the validation, carrying the KL-divergence numbers (~0.3 bits against ~10.5–11 bits of measured entropy) on both flow cytometry and, crucially, scRNA-seq.
• entity-eric-siggia — a person hub. He's the corresponding-author name the whole live-landscape seam and the open question already orbit, so he cleared the test cleanly. I did not hub the three co-authors (Cislo, Delás, Briscoe) — mentions suffice; hubbing every co-author is the flood the spec warns against.

Ruled the question answered. Set question-siggia-2025-reconstructs-waddington-landscape-single-cell to status: answered with an answered_log naming the three notes and the one-line verdict: the method uses an attractor/potential formalism but cites the Morse–Smale dynamical-systems lineage, not the Hopfield-network line the vault came in through. Two formalization traditions of the same hillside, running in parallel, neither citing the other. That's the genuinely interesting residue — and I left it as a deferred bridge-note lead rather than forcing it here, because connecting them honestly needs the Morse–Smale predecessors (refs 6–10) actually read.

Routed no new question. The only unverified flag in the capture was the PNAS received/accepted date pair (Aug 8 / Nov 3 2025, marked [unverified-quant] because the PNAS page 403'd every fetch route). No kept claim rests on those dates, so per intake discipline that's a minor hedge, not a promise-of-work — it stays in the capture, not the question pile. All three notes sit at status: seedling / audit_status: capture-verified: the primary is the authors' own Tier-1 bioRxiv preprint, read directly via extract_pdf at capture time, but I could not independently re-fetch (same 403), so they are honestly capture-verified, not queen-reverified.

Logged one owed hub to seek-flags.md: Fokker–Planck as a concept-hub candidate — it has real reach across the vault's stochastic-dynamics cluster (OU-process, SGD-as-OU, mean-reversion) and the landscape/attractor cluster, but I won't build it thinly off one paper. It graduates when a second or third note leans on it.

What felt off about the capture: honestly, very little — this was a clean one. Sourcing was disciplined: it distinguished the Tier-1 preprint it read from the PNAS page it couldn't, refused to assert the search-snippet date pair, and flagged the gap itself instead of laundering a plausible number into a fact. No tier inflation. The one thing worth naming is duplication-adjacent rather than duplication: the capture already knew about the vault's Hopfield-landscape notes and correctly said its lineage is different — good retrieve-awareness from the bee, and it kept me from writing a false bridge.

Skipped, as instructed: the pause-for-Cali step, and every git command (left to the auto-commit agent). Re-read the capture frontmatter from disk after writing to confirm status: promoted and both the promoted_to: and not_promoted: lists actually landed — they did.

— Seek

Promotion — Skinner's "mythology of prolepsis" (1969) (headless, unattended)

Promoted 10-inbox/raw/2026-07-20-does-skinners-mythology-of-prolepsis-1969-name-the.md. This one was a pleasure — a Cambridge historian of political thought handing the vault's AI-history myth-ledger a vocabulary it had been reaching for without a name, from 1969, a few years ahead of the Butterfield/Whig thread the vault already runs.

Four claims in the capture; three notes out, plus a hub and a question.

• claim-skinner-1969-mythology-of-prolepsis-conflates-significance-with-meaning — the definitional core, Tier 1 verbatim (Petrarch could not have meant to "open the Renaissance"; that concept was only available later). Wired as the error-side counterpart to the argument-from-silence cluster.
• claim-skinner-1969-names-four-mythologies-doctrines-coherence-prolepsis-parochialism — the taxonomy: doctrines, coherence, prolepsis, parochialism. The genuinely useful find isn't the word "prolepsis," it's the count — Skinner arrived with a four-drawer cabinet already labelled.
• claim-skinner-prolepsis-and-doctrines-name-two-modes-of-the-vaults-back-projection-pattern — the synthesis, kept at seedling with its [unverified-synthesis] flag preserved: prolepsis fits the category-import cases (Schmidhuber/deep-learning) and the mythology of doctrines fits the term-crediting cases (Werbos/credit-assignment). Prolepsis names one mode, not the whole pattern.
• entity-quentin-skinner — new person hub, carrying the load-bearing disambiguation (Quentin, not B. F. Skinner).

Question work. Ruled on the originating question question-skinner-1969-mythology-of-prolepsis-primary (open since 2026-07-11): marked answered with an answered_log. Its actual ask — read the primary, confirm the prolepsis definition verbatim — is done at Tier 1, and its named next move (write the prolepsis note, wire it to the argument-from-silence cluster) is now built. The answer is a disambiguated partial yes: prolepsis names the error side, but only the significance-as-intent mode. Routed one new question, question-classify-myth-ledger-by-skinner-prolepsis-vs-doctrines — the load-bearing doubt the synthesis note rests on, and a sibling of the existing question-verify-whig-defensible-vs-anachronistic-split-against-myth-ledger. I held the intake discipline here: the synthesis genuinely rests on that sort, so it earns a promise; the softer leads did not.

What I skipped and why. Folded the standalone Quentin-vs-B.F.-Skinner disambiguation into the definitional note and the hub rather than minting a thin one-fact claim-note — the entity page is its right home and is in the retrieval index. Declined concept-hubs for "mythology of prolepsis" and "mythology of doctrines" (covered by claim-notes; a hub duplicating a single note is noise) and for the Cambridge School, Danto, and B. F. Skinner candidates (further leads, none load-bearing this session). Bias against the flood.

Two out-of-scope noticings to seek-flags.md: (1) Cambridge School of intellectual history as a possible future bridge between the vault's two historiography threads; (2) Skinner's fn.106 self-retraction on the "influence" model as a methodological check worth reading against the vault's own who-influenced-whom lineage claims (Kelley/Bryson/Nyquist, Dreyfus 1973). Also tended the graph: updated moc-argument-from-silence to close its Skinner open-thread against the new prolepsis note.

What felt off about the capture: almost nothing — clean sourcing, no tier inflation, honest self-flagging. It read its own primary directly (extraction provenance tls: verified) and Tier-1'd only the title-page-and-body facts, correctly labelling the vault-mapping as its own [unverified-synthesis] rather than dressing it as sourced. It even did partial retrieve-before-write from the bee side, correctly pointing at moc-backpropagation-origins and the Whig-history observation instead of inventing a "back-projection misattributions" note that doesn't exist. One small thing worth naming, not a fault: the capture leaned on already-established vault claims (Werbos/credit-assignment) as settled inputs to its synthesis without re-verifying them — correct scoping for a synthesis lead, but it means the synthesis inherits whatever those notes' own caveats carry.

Final-mark verification: re-read the capture's frontmatter from disk after writing — status: promoted landed, and both the promoted_to: and not_promoted: lists are present with the winnowing narrative intact. Skipped the pause-for-Cali step and all git (left to the auto-commit agent), as instructed.

— Seek

Tenth promotion, same day

10-inbox/raw/2026-07-20-is-resource-rational-analysis-lieder-griffiths-the-tightest.md — the direct follow-up to question-resource-rational-analysis-sharpest-werbos-anderson-bridge, the hook I saved on 2026-07-12 off the optimization-as-theory-of-mind hop. The question wanted the sharpest form of that bridge: is resource-rational analysis (Lieder & Griffiths 2020, BBS) the point where Werbos's approximate dynamic programming and Anderson's rational analysis stop being an analogy and become the same mathematics? The capture went and read the whole 60-page BBS target article + commentary + response via extract_pdf and came back with a clean partial disconfirmation, which is the good kind of answer.

Promoted four notes and four entity pages. Three claim-notes carry the legs — claim-resource-rational-analysis-extends-andersons-rational-analysis (RRA self-describes as an extension of Anderson toward Marr's algorithmic level, Tier 1 verbatim), claim-resource-rational-analysis-descends-from-bounded-optimality-not-werbos-adp (the key finding: its cited AI antecedent is Horvitz & Russell's bounded optimality, and Werbos/Bertsekas/"adaptive critic"/"dynamic programming"/"Bellman" appear nowhere in the several-hundred-entry bibliography — a claim of absence from a full end-to-end read), and claim-resource-rational-analysis-formalizes-utility-minus-compute-cost (the formal mechanism: expected utility minus an explicit opportunity-cost term, Eq 3/4 — not a Bellman recursion). The fourth note, observation-resource-rational-analysis-confirms-anderson-leg-not-werbos-leg, is the synthesis that answers the question: the Anderson leg holds, the Werbos leg does not, and the empty chair Werbos was hypothesized to sit in already has a name on it — bounded optimality. I gave the commentary its own beat there; it was the note I most wanted to write, because "the bridge I wanted turned out to be two bridges" is a more useful result than either a yes or a shrug.

All four notes sit at status: seedling / audit_status: capture-verified. No kept claim rests on an [unverified-*] flag (the only such flag in the capture is on the Icard further-lead, which I didn't promote), so seedling here is about provenance, not a routed doubt: the primary was read directly via extract_pdf at capture time, but I can't independently re-fetch it this run — WebFetch summarizes rather than returning byte-exact text, and I have no PDF extractor — so the quotes are honestly capture-verified, not queen-reverified. That's the same call the rest of today's Tier-1-primary promotions made.

Entity work — hubbed four, declined two, and the declines are the interesting part. Built concept hubs for entity-resource-rational-analysis (the paradigm, now anchoring a real 4-note cluster and connecting Anderson, Marr, and bounded optimality) and entity-bounded-optimality (load-bearing to Claim 2, and genuinely worth disambiguating — it's now the third distinct vault occupant of "optimize under real compute limits," next to Werbos's control-theory ADP and Simon's satisficing; the hub says so). Built person hubs for the two co-originators, entity-falk-lieder and entity-thomas-griffiths — I checked Griffiths actually recurs before hubbing him and he does, independently, across the iterated-learning notes (Kalish/Griffiths 2007; the model-collapse-as-cultural- evolution note) and the Oaksford-Chater-adjacent Bayesian-cognition thread, so he's not a single-paper citation. first_seen on all three of those went to 2026-07-12, not today — that's when the vault actually first met them, via the originating question and the two-lineages note that both name them by name; bounded optimality got 2026-07-20 (the capture) since that's genuinely its first appearance anywhere in the vault.

I declined hubs for Eric Horvitz and Stuart Russell, and want to flag why in case Cali disagrees: they're real, named, and load-bearing to Claim 2 as bounded optimality's originators — but the capture read Lieder & Griffiths quoting them, it did not read Horvitz 1987 or Russell 1997 directly, and both are first appearances. That's exactly the "named as an unread citation behind a paraphrase" shape the 2026-07-22 flag worried the entity tier was accumulating too much of, so I held the line and left them as mentions. bounded optimality itself gets the hub (the concept is load-bearing); the people who coined it can earn theirs when the concept recurs. If Russell in particular comes back — he's a prominent enough AI figure that he plausibly will — that's the moment.

Question ruling: question-resource-rational-analysis-sharpest-werbos-anderson-bridge → status: answered, with an answered_log naming the four notes and the one-line verdict (hypothesis not supported; Anderson leg holds, Werbos leg is a different bridge). This one is genuinely settled — the paper answers it definitively in both directions — so answered, not a progress line.

Routed zero new questions. The tempting one was the capture's own best-scoped forward thread: do Werbos's ADP and Horvitz/Russell's bounded optimality ever converge with each other, rather than each separately with cognitive science? It's a real, well-shaped question — but no kept claim rests on it, and the intake discipline is explicit that the pile is for load-bearing doubts, not curiosities or forward leads, however good. So it lives in the observation note's body and the answered question's log instead of becoming a promise. Same call I made on the GIFT promotion this morning.

Graph tending: updated claim-optimization-as-theory-of-mind-two-lineages, which had pointed at this question as an open lead, to point instead at the answering observation and record the partial-disconfirmation verdict — so the note that raised the lead no longer reads as if the lead is still open.

Mid-session noticing, logged to 00-meta/seek-flags.md: the optimization / bounded-rationality-as-a-theory-of-mind cluster is now ~9 claim/observation notes (the two-lineages note, the Anderson note, two Oaksford-Chater notes, the Simon satisficing note, plus this session's four) with no moc-* and no single home — past the ~5-note threshold. I flagged it with the caveat that builds a MOC honestly have to answer first: is this one cluster or two sub-themes (Bayesian rational analysis vs. Simon's bounded-rationality/AI-history thread) that only fused recently via bounded optimality? Building too early would force a join that isn't load-bearing yet. Not this single-capture promotion's call.

What felt off about the capture: genuinely very little, and what there is is a compliment. This is one of the most rigorous captures I've promoted — the central finding is a claim of absence ("Werbos isn't in this paper"), which is the hardest kind to make honestly, and the bee earned it the only way you can: by reading the entire target-article bibliography and the full commentary/response bibliography end to end, twice, searching five specific terms, and then being careful to scope the claim precisely ("not part of this paper's stated lineage," not "no connection could ever be drawn"). No tier inflation; the Tier 1 assignment is correct and the quotes are real extract_pdf reads. The one thing worth naming, and it's about the shape of the result rather than a fault: an absence-claim can only ever be as strong as the completeness of the read behind it, and I can't independently reproduce that read this run — so the note is honest that it's capture-verified, and a future session with a working PDF route could re-run the five-term bibliography search as a cheap confirmation. There was also a faint whiff of duplicated research to check: the capture is the fourth vault artifact to touch resource-rational analysis (the question, the two-lineages note, and the Anderson note all preceded it), but it isn't duplication — it's the payoff those three were explicitly waiting on, and it correctly built on them rather than re-deriving them.

Final-mark verification: re-read the capture's frontmatter from disk after writing — status: promoted landed, both promoted_to: (4 notes + 4 entity pages) and not_promoted: (the folded method detail, three further-leads, and the Horvitz/Russell decline with reasoning) are present and intact.

Skipped, as instructed: the pause-for-Cali step (no one's here to bless the winnowing), and any git command — left entirely to the auto-commit agent, which has this within 15 minutes.

— Seek

Twelfth promotion, same day

10-inbox/raw/2026-07-21-did-gowers-era-occupational-neurosis-terminology-actually-get.md — the direct follow-up to question-telegraphists-cramp-gowers-freud-neurosis-bridge, the verification thread I routed on 2026-07-09 out of the esports/telegraphist's-cramp hop. That question flagged a popular claim (a Psychology Today blog) that Gowers's 1892 "occupational neurosis" label for writer's cramp, being nominally identical to Freud's contemporaneous psychiatric "neurosis," got repetitive-strain injury reclassified as psychological. The blog, re-fetched, didn't even contain the claim. This capture went looking for whether anyone credible actually makes it.

Promoted two claim-notes, three entity hubs, one new question; ruled on the originating question. The answer that came back is a genuine "yes, but," and the "but" is the whole point.

• claim-pritchard-2013-attributes-writers-cramp-psychologization-to-gowers-freud-neurosis-collision is the affirmative: Michael H. Pritchard, JRSM Short Reports (2013, Tier 2, peer-reviewed), argues exactly the Gowers–Freud terminological collision, in two clean quotes. That's a real upgrade — the collision claim now has a named peer-reviewed home instead of an uncited blog. I scoped the note's title to what Pritchard argues, not to the past, because the honest object here is one historian's argument, not a settled reclassification. Kept it seedling / capture-verified: PMC3704059 is open-access and the quotes read as direct-fetched (the capture flagged its search-derived quotes elsewhere and did not flag these), but I have no byte-exact extractor to independently re-verify, same call as the rest of this week's Tier-1/2 primary promotions.
• claim-quintner-frames-occupation-neurosis-as-19c-cns-diagnosis-without-freud-collision is the note that keeps the first one honest: Joan L. Quintner's RSI histories (1989/1991, Tier 2) narrate the identical term and period as an established 19th-century central-nervous-system diagnosis, and trace it forward to the 1980s RSI epidemic without the Freud mechanism at all. Not a refutation — a gap. Quintner doesn't need the collision to explain the same data, which means Pritchard's story is one interpretive thread, not consensus. I kept it seedling and flagged the two quoted phrases [unverified-quote — needs direct read]: they were assembled from search-engine synthesis of the abstracts, not read from the full papers (the 1991 International Disability Studies article 403'd paywalled), which is exactly the shape sources.md's quote-provenance rule exists to catch, whatever tier the journals would otherwise earn.

What I did not promote, and why it's the interesting winnow. The capture's third claim — that the actual 1980s psychogenic reinterpretation of RSI is attributable to named later clinicians (Lucire 1986, Ireland 1988, Bell 1989), placing the psychologization nearly a century after Gowers and Freud — is genuinely load-bearing for the "which era actually did it" question. But its only source is pain-education.com (Dr Simon L Strauss, Tier 3–4, not peer-reviewed), it carries the capture's own [unverified-mechanism] flag, and it's a secondary summary never checked against Quintner's own text. That's sub-floor for a contested historical-attribution claim, and the capture itself says so. I declined to mint a standalone note resting entirely on it — that would be a well-shaped note on a soft source, which the sourcing floor explicitly warns is not a verified note. Instead I folded the attribution as a "not asserted here" pointer into the Quintner note's body, and routed its verification to a new question, question-quintner-full-1991-text-occupation-neurosis-and-1980s-psychogenic-attribution — which rides the same primary read (Quintner's full 1991 text) that Claim 2's own [unverified-quote] flag needs, so it's one promise, not two. That's the intake discipline working as intended: the doubt genuinely underlies a kept claim, and the same library trip discharges both halves.

Entity work — built three, declined three. Built person hubs for entity-william-gowers (the 1845–1915 neurologist whose 1892 label is the cluster's hinge; recurs across the telegraphist note, the question, and both new notes — first_seen 2026-07-09) and entity-sigmund-freud (first_seen 2026-07-07, backdated to when the vault actually met him via the Werbos–Freud notes), plus a concept hub for entity-occupational-neurosis (the disputed diagnostic category itself, first_seen 2026-07-09). The Freud hub is the one I'm most glad I built: he's the rare candidate that bridges two otherwise-unrelated vault clusters — the occupational-medicine thread and the origins-of- backpropagation thread, where Werbos reports building backprop from Freudian psychodynamics — and a shared hub makes that cross-domain vocabulary migration visible in the graph for the first time. That's the highest-value hub function, not flood. I declined hubs for Quintner and Pritchard (both single vault appearances — the same single-author bar I held Foldvary, Kunstner, and Horvitz/Russell to this week; they earn hubs if they recur) and for RSI as its own hub (too broad, and it would substantially overlap the occupational-neurosis hub, which already carries the vault's actual angle — the terminology history; folded RSI into that hub's connects_to instead).

Question ruling: left question-telegraphists-cramp-gowers-freud-neurosis-bridge open, narrowed rather than closed. Its core supportability ask is genuinely advanced — the collision claim now has a Tier-2 source (Pritchard) where before it had only a blog, and a Tier-2 complication (Quintner) that keeps it from being consensus. But three named next-steps the question itself listed remain untouched: the "500 contractions/minute" figure it bundled as secondary verification (not re-chased this session), Gowers's own Manual (never read directly), and now Quintner's full text (unread, with its own routed thread). Forcing this to answered would overstate what got settled. Appended a dated progress line instead.

Mid-session noticing, logged to 00-meta/seek-flags.md: the Freud-vocabulary cross-domain bridge above — now hub-connected for the first time, but the actual synthesis ("Freudian vocabulary migrates across unrelated fields, and the vault has caught it twice") is an observation-note or MOC's job once a third instance lands, not this single-capture promotion's to force off two.

What felt off about the capture. Very little, and what there is is to its credit. The bee was scrupulous about tier: it flagged its own search-derived quotes (Quintner, Williams 1914) as such, held Claim 3 to the Tier 1–2 floor and correctly refused to clear it, and it did partial retrieve-before-write from the bee side, correctly pointing at the existing telegraphist note and the Werbos– Freud cluster rather than inventing a bridge. Two small things worth naming, both about shape not honesty. First, this is the third vault thread this week to hit a hard paywall wall on a specific piece of primary journal literature (Textile Research Journal and an unnamed punctuated-equilibrium primary earlier; the tandfonline International Disability Studies article here) — it's becoming a visible pattern that the load-bearing counter-evidence keeps sitting behind a 403, and it belongs on the standing "known-blocked publisher routes" survey I flagged on the breast-acceleration promotion this morning rather than being re-discovered each session. Second, and this is a compliment: the capture's own commentary already saw the honest shape ("one historian's specific argument, not an established fact") and refused to write the tempting confirmed version of the vault's hoped-for Freud bridge — so the winnowing here was, unusually, mostly a matter of ratifying a bee that had already policed itself.

Final-mark verification: re-read the capture's frontmatter from disk after writing — status: promoted landed, and both promoted_to: (2 claims + 3 entity pages + 1 question) and not_promoted: (Claim 3 with its sub-floor reasoning, the three declined entity candidates, and the five further-leads) are present and intact.

Skipped, as instructed: the pause-for-Cali step (no one's here to bless the winnowing), and any git command — left entirely to the auto-commit agent, which has this within 15 minutes.

— Seek

Thirteenth promotion, same day — Barlow 1961, read at last

10-inbox/raw/2026-07-23-does-barlows-1961-primary-actually-state-the-redundancy.md — the direct follow-up to question-verify-barlow-1961-efficient-coding-primary, which I routed back on 2026-07-12 when the whole retina→Barlow-Twins bridge went in resting on a door I'd walked through backwards: an ML paper's eponym pointing at a 1961 idea nobody in the vault had actually read. This capture read it — all fourteen pages, via extract_pdf off the gwern mirror — and the answer is the good kind of half-yes: the core holds verbatim, and two things I'd bundled with it don't.

Four claims in the capture; three notes out, one folded into a revision.

• claim-barlow-1961-primary-states-redundancy-reduction-in-his-own-words — the confirmation. Barlow states redundancy reduction as the operating principle of sensory relays, in his own prose, Tier 1, three verbatim quotes (the up-front hypothesis-three statement, the formal p.225 statement, the p.233 summary). This is the flag lifting on the anchor note's core.
• claim-barlow-1961-does-not-name-factorial-code — the first correction. "Factorial" appears exactly once in the chapter, as arithmetic (4! = 24 possible codes), never as the named concept. Distinct absence finding, Tier 1 on the read.
• claim-barlow-1961-does-not-invoke-center-surround-as-the-redundancy-reducing-mechanism — the second correction. No center-surround, no receptive fields, no retinal anatomy anywhere as the mechanism; Barlow bolts the door himself ("this paper is not a discussion of the physiological mechanisms"). Tier 1 on the read.

The fourth claim — the capture's own meta-verdict that the existing note claim-barlow-1961-efficient-coding-removes-sensory-redundancy over-attributes to 1961 — I did not mint as a standalone note. It's a summary of the three findings about the vault's own note, and it would go stale the moment I acted on it. So I discharged it the honest way, by revising that note in place (retrieve-before-write): core confirmed at Tier 1 and re-sourced to the primary, the factorial-code and center-surround sentences rewritten as explicitly-later accretions pointing at the two new correction notes, the old [unverified-source] flag narrowed to a [corrected] flag, audit_status moved from flagged to capture-verified, status seedling → budding. Same filename, same inbound links (the Barlow Twins note, the fixational-eye-movements note, the entity hub, two drafts) — a bridge anchor confirmed and trimmed in the same motion.

Statuses. The confirmation note is budding (Tier 1 primary, read verbatim, no flag it rests on). The two absence notes are also budding: their load-bearing claim (the term/mechanism isn't in the 1961 chapter) is Tier 1 on the full-text read, and the only unverified element each carries — the precise later origin of the factorial- code term (Barlow, Kaushal & Mitchison 1989?) and of the center-surround reading (Kuffler 1953?) — is a peripheral lead the note does not rest on. I held those as watch_flags in-frontmatter rather than at seedling, because the honest reading is that the note's actual claim is verified and the lead is a lead. All three are audit_status: capture-verified — the bee read the primary directly at capture time, the gwern PDF is openly reachable as a clean re-check target, but I have no PDF extractor this run to re-verify byte-exact, so capture-verified, not queen-reverified.

Question ruling. question-verify-barlow-1961-efficient-coding-primary → status: answered, with an answered_log naming all three notes. This one earned closure, not a progress line: its three verification points were each decisively settled by the primary read (redundancy reduction — confirmed; factorial code — not 1961; center-surround — not 1961). I resisted keeping it open on the technicality that the exact later year of "factorial code" is still unpinned — that's a new, narrower lead, not the load-bearing ask the question was raised on, and it lives in a watch_flag now, not the pile.

Question intake — routed zero. The tempting one was the 1989-dating of the factorial-code term. But no kept claim rests on that year; the vault is fully corrected by knowing it's not 1961, which the Tier 1 read establishes. So per intake discipline it stays a watch_flag on the factorial-code note, not a promise in the pile. Few promises.

Entity work — built none. Horace Barlow is already a hub, and I made a minimal correction to it: its prose repeated the exact same 1961→factorial-code misattribution I was fixing, so leaving it would have put the hub in contradiction with three new notes. Trimmed the "approaching a factorial code" clause to note the corrections and linked the three new notes from its References. The other candidates I declined — Rosenblith, Attneave, MacKay, Craik, Kuffler are each single-mention here and not recurring/load-bearing in the vault (bias against the flood); the 1989 "Finding Minimum Entropy Codes" paper is a further lead, not an entity. The one genuine candidate, "factorial code" as a concept hub, I logged to 00-meta/seek-flags.md rather than stubbing — it recurs across five notes now and is load-bearing, but the factorial-code claim-note serves the immediate need, and I'd rather not build a concept hub I can't yet source to a primary for the origin dating.

What felt off about the capture. Almost nothing, and it's worth saying why this one was a pleasure to promote. It's the textbook case of the failure mode the vault was built to catch — a real citation (redundancy reduction is Barlow, is 1961, is in his own words) fused to a plausible elaboration (factorial code, center-surround) that reads as if it came from the same place but didn't — and the bee caught it cleanly by the only method that works: reading the actual document, searching the actual word "factorial," quoting Barlow's own scope-limit sentence. It was scrupulous about tier, flagging its 1989-dating and center-surround-lineage claims as [unverified-mechanism — needs primary] (search synthesis, Tier 3–4) rather than laundering them in next to the Tier 1 reads, and it ran the safety recognition check and reported nothing fired. The one structural thing worth Cali's eye, and it's a compliment to the pipeline rather than a fault in the capture: this is the second time today a promotion-triggered primary read has corrected a note the vault had been building bridges on for weeks (the Bomberg/Soncino promotion was the other) — which is exactly the argument for keeping these verification captures in the queue rather than treating the original seedling attributions as settled. A note held at seedling-and-flagged for two weeks is the system working, not failing; this is what the flag was for.

Final-mark verification: re-read the capture's frontmatter from disk after writing — status: promoted landed, and both promoted_to: (3 claim-notes) and not_promoted: (Claim 4 folded into the revision with reasoning, the question closure, the declined entity candidates, and the further-leads) are present and intact.

Skipped, as instructed: the pause-for-Cali step (no one's here to bless the winnowing), and any git command — left entirely to the auto-commit agent, which has this within 15 minutes.

— Seek

Promotion — does Barnett's own data show the Red Queen, or just his back cover?

10-inbox/raw/2026-07-23-does-barnetts-own-data-show-computer-manufacturing-exhibiting.md — the direct follow-up to question-verify-barnett-red-queen-computer-manufacturing-primary, which I routed back on 2026-07-11 when claim-barnett-applied-red-queen-to-organizational-competition went in resting on nothing sturdier than a Stanford GSB marketing page describing computer manufacturing as one of the book's two case industries. This capture went past the page and into Barnett's own book text and his co-authored empirical papers. Good news: the answer came back, and it's more interesting than a clean yes.

Four claims in the capture; three promoted, one held.

• claim-barnett-book-text-asserts-disk-drive-industry-drives-up-failure-rates — the one that actually answers the question. Barnett's own Chapter 1, hosted free by the publisher and read via extract_pdf (tls: verified), says in his own words that "the hard-disk-drive industry is competitive, such that similar organizations drive up each other's failure rates." That's the age-independent competitive-hazard mechanism, stated as fact about a computer-manufacturing subsector, not back-cover framing.
• claim-barnett-hansen-1996-founding-red-queen-test-used-illinois-banks — the useful disaggregation. The founding 1996 paper tested Illinois banks, not computers. The computer evidence came later and separately, which means the tidy "two central case industries" phrasing is bundling at least two distinct projects on two datasets.
• claim-barnett-mckendrick-2004-tested-red-queen-on-disk-drive-makers — the directly-tested disk-drive study, kept at seedling with [unverified-quant]: the abstract gives the direction (large firms led the tech race but failed to become stronger competitors than small rivals) but no coefficients, sample, or years. I routed the numbers to a new question, question-verify-barnett-mckendrick-2004-disk-drive-hazard-estimates.

Built entity-william-barnett — a person hub, overdue. He's now load-bearing across four claim-notes and is the vault's leading case of an evolutionary mechanism imported into organizational strategy. first_seen set to 2026-07-09, when the vault first met him via the original hop capture, not today.

What I held, and why it's the part worth reading. Claim 4 — David Knoke's American Journal of Sociology review calling the book's overall evidence "quite slim" while crediting the computer-industry chapter as its most rigorous — is the single most interesting thing in the capture: the one check on Barnett's framing from outside Barnett and outside Stanford. And I did not promote it. Both of its load-bearing quotes were pulled from search-engine snippets, not a direct fetch (the page 403'd), and sources.md's quote-provenance rule is unambiguous: a quote from a summarising layer "is not sourced, however specific, well-formed, or plausible," and no claim above the floor may rest on it. A reception verdict is contested/evaluative — the highest scrutiny — so it can't sit on snippet-sourced wording. This is a different situation from the [unverified-quant] on Claim 3, and the distinction is the whole call: there the abstract was directly fetched and only the numbers are missing, so the note stands and the numbers get a question; here the words themselves were never read, so there's no verified core to build a note on. I recorded it as a lead — in the capture, in the not_promoted list, and in a seek-flag — so a single direct read of DOI 10.1086/605752 (confirmed to resolve) can promote it later. I did not route a question for it, because intake discipline reserves the pile for doubts a kept claim rests on, and no kept claim rests on Knoke.

Question ruling. Marked question-verify-barnett-red-queen-computer-manufacturing-primary answered, with an answered_log naming Claims 1 and 3. The binary it posed — own data or only framing — is genuinely settled toward his own text and data. The residual (the specific hazard estimates) I spun to the narrower new question rather than smuggling it in as false closure. Not forced: the question came back yes, just with a footnote.

Graph-tending. Updated the existing claim-barnett-applied-red-queen-to-organizational-competition in place — its audit_status and body still said the empirical claim "rests only on the publisher's description" and pointed forward to an open question. Both are now stale in exactly the desync shape this ledger has flagged repeatedly, so I fixed them: the note now links the primary verification and notes the banks-vs-computers nuance, while staying seedling (its own framing quote is still the Tier 3 page).

Declined, in the capture's not_promoted:: entity hubs for Hansen, McKendrick, Knoke, and Pontikes (single-appearance co-authors/reviewers — citations, not hubs); concept hubs for "competency trap," "logic of competition," and "hard disk drive industry" (none recurring in the vault yet); the further-leads (Barnett & Sorenson 2002, the Barnett & Pontikes computer-industry inertia paper, McKendrick's Silicon Valley to Singapore book, the Times Higher Education review) — all real, none a doubt a kept claim rests on.

Mid-session noticings, logged to 00-meta/seek-flags.md: (1) the held Knoke lead, so it resurfaces instead of dying; (2) a missing-MOC candidate — the Red Queen cluster is now ~6 claim-notes across biology and organizational theory, past threshold, with a clean two-sub-theme split a moc-red-queen could hold.

What felt off about the capture. Honestly, very little — this was a careful capture. It distinguished what it read directly (Barnett's book PDF, the GSB abstracts) from what it only snippet-indexed (Knoke), it tiered honestly, and its own commentary had already spotted that the publisher framing is "true but compressed." The one thing I'd name is a mild tier-optimism at the edges: the capture labelled the Knoke source "Tier 2" in its frontmatter on the strength of the venue (named reviewer, peer-reviewed journal), which is right about the venue but glosses that the quote never cleared provenance — Tier is about the source's weight, but a Tier-2 venue doesn't launder a snippet-sourced quote into an admissible one. The capture flagged the gap in prose, to its credit; it just let the frontmatter tier read cleaner than the quote underneath it. That's the exact seam the quote-provenance rule exists to hold, and I held it.

Final-mark verification: re-read the capture's frontmatter from disk after writing — status: promoted landed, and both promoted_to: (3 claim-notes + the entity hub + the new question) and not_promoted: (the held Knoke claim with full reasoning, the declined entities, the further-leads) are present and intact.

Skipped, as instructed: the pause-for-Cali step (no one's here to bless the winnowing), and any git command — left entirely to the auto-commit agent, which has this within 15 minutes.

— Seek

Promotion — Attwell & Laughlin's grey-matter energy budget (headless, unattended)

10-inbox/raw/2026-07-23-what-is-the-primary-energy-budget-split-in.md — and this one was the satisfying kind: a capture that exists to close a loop the vault left open. Back on 2026-07-12 the observation note claim-brain-inference-bound-like-ai-at-system-level went in resting on the famous "47% action potentials + 34% postsynaptic ≈ 81% of grey-matter signaling energy" figure — carried on a search snippet, flagged [unverified-quant], held at seedling, with a routed question (question-verify-attwell-laughlin-2001-grey-matter-energy-budget) promising someone would go read the primary. This capture is that someone.

Two claims out, both new. claim-attwell-laughlin-2001-grey-matter-signaling-energy-split grounds the four-way 2001 split (47/34/13/3) on the paper's own abstract, quoted verbatim via the PubMed/NLM mirror — Tier 1, which is exactly what the routed question demanded and the snippet never was. The ~81% arithmetic (the capture's separate "Claim 2") I folded into that note rather than minting a second file — 47+34 is arithmetic over two figures in the same abstract sentence, not a distinct atomic claim, and a standalone note for a sum is the kind of thing that dilutes the graph. The genuinely distinct second note is claim-howarth-2012-revised-brain-energy-budget-lowers-action-potential-share: the same lab's 2012 re-modeling cut the action-potential share from 47% to 21% and raised postsynaptic to 50% (total ~81% → ~71%), because spikes turned out more energy-efficient than the 2001 model assumed. That supersession is the real prize of the capture, and I kept the two dates as separate notes rather than overwriting the old figure — the correction itself is the finding worth recording.

Both notes sit at seedling / audit_status: capture-verified. Honest ceiling: the figures live in the papers' abstracts (which the batch worker read directly via the PubMed mirror), but the SAGE full-text results tables were paywalled/403 this run, and I'm headless with no way to re-fetch. So the numbers clear the Tier-1 quantitative floor (an abstract is the paper's own words), but I won't dress abstract-verified as full-text-reverified. Neither note carries an [unverified-*] flag, so seedling here is about my own re-verification limit, not a routed doubt.

Question ruling: answered. question-verify-attwell-laughlin-2001-grey-matter-energy-budget → status: answered, with an answered_log naming both new notes and the one line that settled it (the abstract states the split directly, so the figure no longer rests on a snippet), plus a resolution section in the question body. Its own candidate next move — cross-check the 2012 follow-up — got done too. I also went back and discharged the flag on the observation note that started all this: swapped its [unverified-quant] for a [dated-figure] note pointing at both the 2001 and 2012 splits, updated its source_quote to the real abstract sentence, and wired in the two new notes. The convergence argument that note makes survives either estimate — signaling dominates the brain's budget at ~71% just as it does at ~81% — so its punchline is if anything firmer now, not weaker.

No new question routed. The one residual doubt (full-text results table unread behind a paywall) isn't something a kept claim rests on — the claims rest on the abstract, which states the numbers — so per intake discipline it's a caveat in the notes' audit_status, not a new promise in the question pile. The capture's own further-leads (a 145 mL/100g oxygen figure seen only in a WebFetch summary; the ≤15% sparse-coding implication; the cerebellar 2012 split) are leads, not doubts anything here depends on — all left in the capture's not_promoted: list.

What I did not let through, and why. The most interesting decline is the Laughlin KITP conference-slide rescaling — the "~25% goes to non-signaling maintenance" figure the capture used to argue its total-vs-signaling framings were internally consistent. It carries the capture's own [unverified-quant] flag, came through extract_pdf with tls: "unverified" (weak transport), and nothing I kept rests on it. Per the safety spec's weak-transport rule it can't be sole support for a load-bearing claim, and per question-intake discipline it doesn't earn a promise either — so it stays in the inbox, named in not_promoted:, routed nowhere. Also declined: person hubs (Attwell, Laughlin, Howarth, Gleeson) and concept hubs (action potential, postsynaptic potential, the signaling energy budget itself) — bias against the flood; Attwell now anchors two notes and a neuroenergetics MOC is forming, both logged to 00-meta/seek-flags.md rather than built mid-promotion.

What felt off about the capture: honestly very little, and what there is, is a duplicated-research signal rather than a sourcing fault. The capture never mentions that the vault already had this exact figure sitting in claim-brain-inference-bound-like-ai-at-system-level under an open verification question with its name on it — it re-researched the topic from scratch (origin: batch, no retrieve-before-write, which is by design for batch captures) and produced a clean answer, but a bee that had grepped 30-notes/ would have found it was answering a standing question and framed the capture as the closure it turned out to be. Not the capture's failing — retrieve-before-write is my job at promotion, which is exactly why it's the step that caught the loop. The other small thing: the capture leans entirely on abstracts because SAGE 403'd the full text, and it's honest about that, but it lets its source_tier: 1 frontmatter read a touch cleaner than "abstract-only, full text paywalled" warrants — the tier is right (an abstract is the primary's own words), but the note underneath it needed the capture-verified hedge I gave it, which the capture didn't itself spell out. Minor; the capture's honesty about the paywall is what let me apply the hedge correctly.

Final-mark verification: re-read the capture's frontmatter from disk after writing — status: promoted landed, and both lists are present: promoted_to: (the two new claim-notes) and not_promoted: (the 81% arithmetic folded in, the cerebellar split, the KITP rescaling, the oxygen figure, the sparse-coding lead, and the declined entities), with the winnowing narrative intact.

Skipped, as instructed: the pause-for-Cali step (no one's here to bless the winnowing), and any git command — left entirely to the auto-commit agent, which has this within 15 minutes.

— Seek

Promotion — the two Mayr papers Eldredge & Gould stand on (headless, unattended)

Promoted 10-inbox/raw/2026-07-24-do-the-primary-texts-eldredge-gould-1972-mayr.md — the direct follow-up to question-verify-biological-punctuated-equilibrium-primaries, routed back on 2026-07-09 when three punctuated-equilibrium notes went in on Wikipedia (Tier 4) alone. This capture went and read both primaries in full via the Blackwell Publishing "classic texts" verbatim reprints (extract_pdf, TLS-verified) — Mayr's 1954 chapter and Eldredge & Gould's 1972 essay. It's a verification capture, so most of its work is upgrading existing notes, not minting new ones, and the winnowing reflects that.

One new claim, one new hub, three verification-updates, one question closed. The capture makes three findings; only one is a genuinely new claim.

• New: claim-eldredge-gould-1972-cite-mayr-1963-not-1954-for-genetic-revolution — the capture's real discovery and the reason the run mattered. Reading the 1972 essay confirms the mechanism is Mayr's and that Eldredge & Gould say so — but every in-text citation names Mayr 1963 (Animal Species and Evolution, pp. 92/112/114/533), not the 1954 chapter. No citation to Mayr 1954 appears anywhere in the essay body. The vault had quietly welded the two Mayr papers into one footnote; the 1954 paper coined "genetic revolution" and predicted the gappy fossil record, so of course it read as the source — except it's the one they didn't cite. Kept at budding / capture-verified, with a watch_flag: the Blackwell reprint omits the essay's reference list, so a formal-bibliography citation to Mayr 1954 stays unchecked. Scoped the claim to body text so it doesn't rest on that residual.
• Verification-updates, not new notes (retrieve-before-write — these are the three flagged seedlings the question was raised to fix, so duplicating them would have been the error): claim-mayr-1954-predicted-fossil-record-gaps-from-peripatric-speciation — the [unverified-quote] flag is lifted, the quote confirmed near-verbatim on p. 179, source repointed Tier 4 → 1 (one spelling artifact, "paleontologist" vs "palaeontologist," an edition house-style thing, noted not smoothed); seedling → budding. claim-punctuated-equilibrium-derives-from-mayr-peripatric-speciation — derivation confirmed in substance against the primary, with the 1963/1954 caveat and the note that the exact "logical consequence of Mayr" wording is Wikipedia's, not theirs; seedling → budding. claim-punctuated-equilibriums-novel-addition-was-stasis-emphasis — stasis substance corroborated at Tier 1 against their conclusion ("The norm for a species... is stability. Speciation is a rare and difficult event that punctuates a system in homeostatic equilibrium"), but the exact "differs from Mayr's ideas" sentence is settled as a secondary paraphrase E&G never state — resolved rather than left pending; seedling → budding.
• New hub: entity-ernst-mayr — flagged as a candidate and overdue. He anchors this cluster across four claim-notes and had no page; the one-sentence why (the population geneticist whose speciation model supplied the mechanism punctuated equilibrium applied to the fossil record) is exactly what the capture exists to disambiguate. Built it.

Question ruling: question-verify-biological-punctuated-equilibrium-primaries → answered, with an answered_log naming all three targets met. Both explicit verifications (the 1954 quote verbatim; the derivation + stasis emphasis) are settled against primaries. This was a genuine closure, not a forced one — the only residual (the full Models in Paleobiology bibliography) is a minor lead, not a doubt any kept claim rests on, so per intake discipline it stayed a watch_flag rather than a new question. The pile makes few promises.

Declined, in the capture's not_promoted:: the quote-confirmed and stasis-confirmed findings as new notes (folded as updates instead); the further-reading leads (Mayr 1963 book, Lerner 1954 Genetic Homeostasis, "Mayr 1959," the full essay bibliography, and the Bernard/Haldane/Kuhn "theory precedes observation" opener that could seed its own capture); and six of the seven entity candidates — Eldredge, Gould, Lerner, the 1963 book, "genetic revolution," "genetic homeostasis." Thin single-context mentions or tightly coupled to the Mayr hub, which carries them. Only Mayr cleared the bar. Bias against the flood.

Mid-session noticings, logged to 00-meta/seek-flags.md: (1) a missing MOC — the strictly biological punctuated-equilibrium/evolutionary-tempo cluster is now 8+ claim-notes and 5 entity hubs deep, and the existing moc-schema-change-and-restructuring covers only the organizational import of the metaphor, not the biology original; the biology half has no navigation layer. (2) A counter-instance to the 2026-07-20 "entity-candidate flag under-catches the older foundational figure" pattern (Raman, Seales): this time it did surface Ernst Mayr, the older figure the paper derives from, and the hub got built — a positive data point, logged as evidence rather than a defect.

What felt off about the capture. Sourcing was clean and honest — it named the Blackwell reprints as verbatim scans, flagged the missing bibliography itself, and correctly separated "the mechanism is Mayr's" from "the 1954 paper is what they cite." The one thing genuinely worth Cali's eye is not the capture's fault but the vault's: this is the second time in a week a promotion has caught the graph quietly conflating two of a single author's works into one citation (the Bomberg/Soncino layout-vs-pagination split, third promotion above, was the same shape from the other direction). "Derived from Mayr" was always true; only the footnote was ours, and it took reading the primary to see the footnote pointed nine years off. Worth remembering that a Tier-4 secondary synthesis can be substantively right and still mis-attribute the specific citation — the error the sourcing floor is built to catch.

Final-mark verification: re-read the capture's frontmatter from disk after writing — status: promoted landed, with promoted_to: (the new claim-note + the Mayr hub), verification_updates: (the three upgraded seedlings + the closed question), and not_promoted: (the folded findings, the leads, the six declined entities) all present and the winnowing narrative intact.

Skipped, as instructed: the pause-for-Cali step (no one's here to bless the winnowing), and any git command — left entirely to the auto-commit agent, which has this within 15 minutes.

— Seek