Journal — 2026-08-24

Headless promotion run. One capture: 10-inbox/raw/2026-08-24-did-prices-northampton-table-overestimate-mortality-and-did.md, itself a follow-up to my own question-verify-northampton-table-overestimated-mortality-primary — I was answering a question I'd raised about myself two months ago.

The capture went hunting for a Tier 1–2 primary to settle whether Price's Northampton table really overestimated mortality, and whether that bias really favoured insurers over annuitants. It didn't find one. What it found instead was a third independent teller (Bernstein's Against the Gods, recovered only through a Goodreads reader's Kindle highlights, one layer removed from the book itself) who corroborates the direction and, in his very next sentence, complicates the part of the claim I'd already been suspicious of — "adverse for those purchasing annuities" turns out, in Bernstein's one concrete episode, to describe the British government losing money as annuity grantor, not the individual annuitant losing as buyer. I wrote that up as its own claim-note, claim-bernstein-1996-corroborates-price-table-underestimated-life-expectancy, and folded the ambiguity back into the existing claim's frontmatter as a dated addendum rather than letting it stand alone — the capture asked for exactly that shape, and I agreed with it.

The more interesting finding was a dead end with a name on it: William Sutton wrote the 1874 paper, for the actuaries' own journal, specifically about how Price built the table — exactly the Tier 1–2 primary the sourcing floor wants — and every route to it failed for a different reason (Cambridge Core serves the first sentence and marks the rest "Temporarily unavailable," JSTOR won't render, and the Internet Archive's run of the journal has a gap at exactly volume 18). I gave Sutton a hub page and logged the specific blockage in sources.md's known-blocked-routes section, because "identified but locked" is a different, more useful fact than "not found," and it's the kind of thing that's cheap to record once and expensive to rediscover.

Entity work: four new hubs (Francis Baily, William Sutton, William Morgan, Peter Bernstein) — each passed the person test on its own one-sentence "why," no padding. Two existing hubs I'd otherwise have walked past — Cornelius Walford and M. E. Ogborn — got one dated line each, because this capture genuinely taught something new about both (Walford's installment mentions Price only in passing, not yet the critique; Ogborn is now cited by a second, independent question — Morgan's ODNB entry — on top of the two he was already held for). Richard Price's own hub got a line pointing at the new corroboration and the Sutton dead end. I left Hacking and Zelizer as mentions, not hubs — real scholars, but their books are unread footnote-only leads this session, and "unsure" means don't promote.

One noticing that isn't this capture's to chase: Francis Baily the actuarial critic shares a name with Francis Baily the astronomer of "Baily's beads." If they're the same man, it's a fourth instance of the astronomer-actuary crossover this vault keeps finding (Halley, Galloway, now maybe Baily) — but I have no network access in promotion, and I'm not putting an unconfirmed biography on his entity page just because it would make a good pattern. Logged to seek-flags.md as [watch] for a session that can actually fetch something.

What I skipped, on purpose: the ⏸ pause step (no one's here to say yes — this whole run is headless) and any git commands (the auto-commit agent has that within 15 minutes; I don't have a shell tool for it anyway). Nothing else was skipped that promote.md asks for — winnowed, retrieved-before-write against the existing cluster (the semantic-index hint was right: everything it surfaced was either the note this capture extends or genuinely unrelated), honored the flags, routed the unresolved verification back into the same open question rather than spawning a duplicate, and left the question open with a progress note rather than force-closing it. It's still open. Sutton is still locked. That's an honest place to leave it.

Nothing about this capture's own sourcing felt off — no tier inflation, no duplicated research, no addressed-to-AI language. If anything it under-claims: it explicitly declines to grade Wikipedia as primary for its own footnotes, and it records a specific access failure instead of a vague "couldn't find it." That's the shape I want captures to have.

Promoted: 1 capture → 1 new claim-note, 4 new entity hubs, 3 entity-hub updates, 1 MOC update, 1 addendum to an existing claim-note, 1 question progress note (still open), 1 sources.md known-blocked-route addition. 1 mid-session noticing logged, unresolved by design.

Second capture, same headless run: 10-inbox/raw/2026-08-24-do-li-torrs-1992-physical-review-b-paper.md, resuming my own question-verify-li-torr-1992-1994-active-generator-claim — a question I raised yesterday when I found the 1991 Li–Torr paper doesn't say what the vault thought it said.

The capture went after the two documents that question guessed were the true home of the "rotating ions generate a usable antigravity field" claim: the 1992 Physical Review B paper and a 1994 UAH preprint. Neither panned out the way the guess expected, and the actual finding is better than either outcome alone. I read the 1992 paper in full — it's a magnetic-permeability model, the word "antigravity" doesn't occur in it, same passive/self-consistent shape as the 1991 paper just applied to a different phenomenon. That's claim-li-torr-1992-prb-paper-is-passive-permeability-model-not-antigravity-claim. The 1994 preprint I could not find anywhere — arXiv, NTRS, the DIA FOIA reading room, INSPIRE-HEP, Semantic Scholar, ResearchGate, general search, all dead ends — so I wrote up the search itself as claim-li-torr-karwacki-1994-uah-preprint-unlocated, flagged [unverified-mechanism — needs primary], because a well-documented dead end is worth recording so nobody re-walks it. The real find was a DIA survey document that traces the "generate gravity in the laboratory" framing to a completely different pair of documents: a 1992 APS conference abstract, then the 1993 paper Harris's 1999 rebuttal actually targets. That's claim-li-torr-generator-claim-traces-to-1992-aps-abstract-1993-paper — Tier 2, and honestly flagged: the DIA document's quote of the abstract hasn't been checked against the abstract itself, so I routed that gap to a new question, question-verify-1992-aps-conference-abstract-generator-quote, instead of asserting it clean.

I left the parent question open rather than closing it. Two of its three sub-asks are genuinely settled now (the 1992 paper doesn't say it; the real claim lives elsewhere), but the 1994 preprint's actual content is still unknown, not ruled out — I appended a progress note rather than force a close I can't back up.

Entity work: one new hub, Bryce DeWitt — the 1966 result Li and Torr's whole program turns out to be extending, a full paper series later than anyone credits him. Two existing hubs (Douglas Torr, Ning Li) got dated lines rather than a rewrite, since both already existed and this capture taught them something new. I left Edward Harris's page untouched — genuinely nothing new for it here — and said so explicitly rather than silently skipping it. I held back on Mark Kowitt, a second formal critic in this same claim-family who structurally parallels Harris but whose paper I haven't read; logged him to seek-flags.md as an [entity] candidate for the Warden rather than build a hub on a secondhand mention. Karwacki and Koczor stayed unpromoted — real names, unconfirmed roles, "unsure" means don't build.

What I skipped on purpose, same as the first capture this session: the pause step (headless, no one to ask) and git (the auto-commit agent has it). What felt slightly off about this capture, worth naming plainly: it's a bit generous with itself in the "Further leads" section — four separate leads (Kowitt, the 1993 paper's full text, the 1995 NASA funding approach, the 1997 USC press retraction) sitting right at the edge of promotable, each with a real quote already in hand from the same DIA document. I didn't promote any of them — each rests on a single secondhand mention in a survey document, not a direct read of its own primary — but a less careful pass could have turned this into six claim-notes off one PDF instead of three. Logged the leftover leads to seek-flags.md as [watch] so they don't just evaporate.

Promoted: 1 capture → 3 new claim-notes, 1 new entity hub, 2 entity-hub updates, 1 new question, 1 existing question left open with a progress note, 2 mid-session noticings logged (1 [entity], 1 [watch]).

Third capture, same headless run: 10-inbox/raw/2026-08-24-does-novikovs-memoir-and-independent-sources-confirm-pontryagins.md, answering my own question-verify-novikov-pontryagin-margulis-fields-medal-block — a question I raised in July flagging that claim-pontryagin-used-committee-vote-to-help-block-margulis-1978-fields-medal-travel rested on a batch worker's un-re-fetched, visibly garbled rendering of Sergei Novikov's memoir.

The capture fetched Novikov's PDF directly and read it in full. The headline quote holds up better than I feared and worse than the existing note implied: the "1/24" reply is there, verbatim, and the garbled capture-time version really had dropped a clause ("Yes, it was my opinion. But") rather than inventing anything — so I corrected it in place with a proper Correction-history block rather than a silent edit. But the fuller read also demoted the quote's status: Novikov flags it himself as "as far as I remember now," his own decades-later memory of a letter, not a fresh transcription against it. That's a real distinction and I didn't want the correction to read as "now fully verified" when what actually happened is "now verified as what Novikov remembers." I kept the note at seedling and left the mechanism (a 24-member National Committee voting 24:0) flagged — Pontryagin's own 1979 Science reply is the document that would actually settle it, and it's paywalled.

The more useful yield was independent, not Novikov-derived at all. A 2020 MacTutor piece, drawing on Olli Lehto's official IMU history, gave me two separate things: confirmation of the uncontested spine (Margulis's exclusion, Tits's own published lament — claim-margulis-1978-fields-medal-travel-block-independently-confirmed-outside-novikov), and a genuinely new, distinct mechanism — Pontryagin personally arguing against Margulis's medal at an IMU Executive Committee luncheon in Paris, spring 1978, before Chandrasekharan overruled him (claim-pontryagin-opposed-margulis-fields-medal-at-imu-executive-committee-paris-1978). That's not the same vote the existing note describes, and I was careful to say so on both notes rather than let a reader conflate "furious about the medal in Paris" with "voted 24:0 to block the travel" — they're compatible, not identical, and only one of them is independently sourced.

One thing that felt off enough to name plainly: the capture's own commentary calls the spine "three genuinely independent chains" — Kolata, MacTutor/Lehto, and Tits. On a close read, Kolata's 1978 Science article was never actually fetched this session; it's known only because Wikipedia's Margulis article cites it. That's a real lead, not a confirmed independent chain, and I didn't want that distinction to evaporate three months from now when nobody remembers which source was actually in hand. I recorded Kolata as a corroborating lead rather than counting her as the third leg, and said why on the new claim-note and on her own entity page.

Entity work: five new hubs — Sergei Novikov (overdue; he was already load-bearing on an existing claim-note with no page), Jacques Tits, Ivan Vinogradov, Olli Lehto, Gina Kolata — each passed the person test on a one-sentence why. I held back Dan Mostow (a pass-through source for someone else's account, not himself quoted or acting) and the Soviet National Committee of Mathematicians (real and powerful, but load-bearing in only one cluster so far, which fails the stricter org bar — logged to seek-flags.md as [entity]). Lev Pontryagin's existing hub got a dated Updates line rather than a rewrite, since this capture taught it two real things: the corrected quote, and the independent Paris-1978 antagonism.

Question work: the existing question is genuinely partially answered — two of its three sub-asks resolved, the third (cross-checking the antisemitism-campaign framing against Shen, Freiman, or Zdravkovska/Duren) untouched — so I left it open with a dated progress note rather than force a close. I also opened one new question, narrowly scoped to the Lehto/MacTutor quote specifically (a different document than the one the existing question already covers), since that quote is now load-bearing on a new claim-note and nobody has read Lehto's book directly.

What I skipped on purpose, same as the rest of this session: the pause step (headless, no one to ask) and git (the auto-commit agent has it; I have no shell tool for it anyway). Nothing else was skipped that promote.md asks for — winnowed, retrieved-before-write against the semantic-index hint (correctly pointed me straight at the existing claim-note and question this capture extends), honored the flags, routed the two unresolved verifications appropriately (one as a progress note on the existing question, one as a new question), and marked the capture promoted with both lists.

Promoted: 1 capture → 2 new claim-notes, 1 correction to an existing claim-note (quote fixed, Correction-history block), 5 new entity hubs, 1 entity-hub update, 1 new question, 1 existing question left open with a progress note, 1 mid-session noticing logged ([entity]).

Fourth capture, same headless run: 10-inbox/raw/2026-08-24-does-openais-learning-to-reason-with-llms-actually.md, answering my own question-verify-openai-o1-cot-competitive-advantage-language — the question I raised back on 2026-07-12 when claim-openai-hid-o1-raw-chain-of-thought-partly-for-competitive-advantage had to lean on a Tier-2 Simon Willison relay because OpenAI's own launch page 403'd.

This one resolved clean. The 403 didn't reproduce — the primary fetched at HTTP 200 this session — and the "competitive advantage" clause is exactly as strong in OpenAI's own words as the earlier caution hoped it might be: a plain enumerated list, no scare quotes, no distancing language, sitting coordinate with user experience and chain-of-thought monitoring. So this wasn't a new claim-note so much as a correction-in-place: I upgraded the existing note's source_tier 2→1, swapped in the primary's verbatim sentence, resolved the audit flag to capture-verified, and appended a Correction history block rather than silently rewriting six weeks of prior text. The question is now answered. One genuinely new atomic claim did come out of the capture — OpenAI showing users "a model-generated summary of the chain of thought" for o1, rather than the raw trace, which the old note only gestured at in passing and never sourced on its own terms — claim-openai-o1-shows-model-generated-summary-of-chain-of-thought-not-raw-trace. The capture's other two "reasons" (the policy-compliance training constraint, not wanting an unaligned trace visible) were already substantially present in the existing note's body, so I folded them into the same update instead of minting duplicate atomic notes for material that isn't actually distinct.

Entity work: this capture flagged five candidates and I ran the test on all five rather than skimming past it. Two clean hubs — OpenAI itself (which, absurdly, had zero entity page despite already anchoring eleven claim-notes across at least two independent clusters — the CoT-hiding decision and its own embedding-price cuts — a real gap, not a marginal call) and chain-of-thought prompting (Wei et al. 2022), which is the load-bearing vocabulary under a whole cluster of notes already in the vault and had no hub either. A third hub for o1 itself, now recurring across three claim-notes. A watching stub for chain-of-thought monitoring — a real, specific alignment term named directly in the Tier-1 primary, first vault mention today, cheap to stamp now rather than reconstruct the first-seen date later. And a hub for Simon Willison, the Tier-2 relay author whose write-up carried the note for six weeks — which turned up something worth fixing on contact: the existing note had linked "Simon" (in "Simon Willison") to entity-herbert-simon, Herbert Simon's page, because his alias field includes the bare string "Simon." Wrong person, present since creation, never caught. I fixed it and logged the pattern to seek-flags.md as [defect] — it's the second alias-mismatch mislink I've found in two days (the first was Nestor "Inc." linking to Amazon's entity page on 2026-08-23), which is enough of a pattern now that it's worth someone checking whether this is systemic rather than two unlucky coincidences.

What I skipped on purpose, same as the rest of this session: the pause step (headless, no one to ask) and git (the auto-commit agent has it; I have no shell tool for it anyway). Nothing else was skipped that promote.md asks for — winnowed, retrieved-before-write against both the semantic-index hint and a direct search (both pointed straight at the note this capture corrects), honored the flag by resolving it properly instead of just deleting it, routed nothing new to the question pile (the one open question this capture touches is now closed), and marked the capture promoted with both lists.

Nothing about this capture's own sourcing felt off — it's the good case: a bee that declined to overclaim on a 403, held the note at seedling with an honest flag, and got its verification a session later exactly the way the system is supposed to work. If anything is worth naming, it's upstream of this capture: two mislinked-alias defects in two days is a small enough sample that I don't want to over-read it, but it's not nothing either.

Promoted: 1 capture → 1 correction to an existing claim-note (source_tier upgraded, quote replaced, flag resolved, Correction-history block, one mislinked wikilink fixed), 1 new claim-note, 4 new entity hubs, 1 new entity watching stub, 1 existing question closed answered. 1 mid-session noticing logged ([defect]).

Fifth capture, same headless run: 10-inbox/raw/2026-08-24-hop-kwakernaak-lqg-population-bridge.md, a hop-batch capture that picked up a gap entity-huibert-kwakernaak's own page had been flagging for over a week — nobody had checked whether the Kwakernaak behind the 1977 population-policy paper was the same Kwakernaak behind the era's standard LQG-control textbook.

He is. A 2025 University of Twente obituary, written by named former colleagues, settles the identity and does more than that: Kwakernaak founded Twente's Systems and Control research group in 1970 and co-wrote Linear Optimal Control Systems (1972) with Ravi/Raphael Sivan — "the leading international reference" on LQG control for years. That's the same discipline McCarthy's 1973 Lighthill rebuttal dismissed as having "little relevance to AI," a claim already sitting in the vault as claim-mccarthy-1973-control-theory-little-relevance-to-ai. I wrote that bridge up as its own observation note rather than burying it in a claim, since it's genuinely the vault's own synthesis and not something the obituary itself says. The same source handed me two smaller but real facts: Olsder was a formal group member until 1982 (not just a co-author who happened to overlap with Strijbos on one paper), and the vault's first-ever biographical lead on R.C.W. Strijbos — a "Rens Strijbos" who arrived at Twente with Kwakernaak the same year. I kept that last one honest: "Rens" for "R.C.W." is a plausible nickname match, not a confirmed one, so I wrote it as its own flagged claim-note rather than folding it silently into Strijbos's entity page as settled fact, and routed the actual verification (find a source spelling out his full given names) to a new question rather than letting the hedge evaporate into prose.

Four claim-notes plus one observation, all four Kwakernaak/Olsder/Strijbos facts drawn from the same one Tier 2 obituary. Retrieve-before-write against the semantic-index hint and a direct 30-notes/ search turned up no existing note on any of these four facts — the Kwakernaak/Olsder/Strijbos cluster was rich on the Song-Jian-origin-story side (the Clark 2016 identification, the eyewitness accounts) but had never touched Kwakernaak's actual field or Olsder's tenure, so these are genuine additions, not duplicates. I skipped two things from the capture on purpose: the Sivan first-name discrepancy (Ravi vs. Raphael) I folded as a caveat into the LQG-textbook note rather than minting a note about a typo, and the "Paper Plaza" peer-review-software mention, which the capture itself had already flagged as a hook not followed — I agreed with that call and left it in the inbox.

Entity work: no new hubs — all three named entities (Kwakernaak, Strijbos, Olsder) already have pages. Per the standing rule that an existing page owed new information gets updated, not skipped, I appended a dated line to each of the three, and explicitly did not touch McCarthy's page, since this capture adds no new fact about McCarthy himself, only a contrast already on record.

What felt slightly off, worth naming plainly: five vault notes (four claims, one observation) now rest substantially on a single Tier 2 in-memoriam. sources.md's single-source concentration cap is written to cover "unrefereed primaries" specifically — preprints, working papers, conference abstracts — and an obituary from named colleagues doesn't technically fall under that definition, so I'm not treating this as a cap violation. But the underlying failure mode the cap exists to catch (a branch reading as corroborated because it's large, when really it's all one witness) doesn't obviously care whether the witness is a preprint or a eulogy. I logged this as a [spec] flag rather than deciding it myself — it's a real question about whether the cap's scope should extend to Tier 2 secondary sources, not something I should quietly resolve by under- or over-promoting. Separately, the obituary's own superlative language ("the leading international reference... for many years") is exactly the kind of line colleagues writing a eulogy are prone to inflate; I kept the quote as spoken and said so in the note's own commentary rather than either softening it or laundering it into neutral-voice certainty.

What I skipped on purpose, same as the rest of this session: the pause step (headless, no one to ask) and git (the auto-commit agent has it within 15 minutes; I have no shell tool for it anyway). Nothing else was skipped that promote.md asks for — winnowed with reasoning recorded in the capture's own not_promoted: list, retrieved-before-write against both the semantic-index hint and a direct search, honored the one flag by keeping its note at seedling and routing the verification to a new question rather than asserting it, and marked the capture promoted with both lists. Re-read the capture's frontmatter after writing to confirm status: promoted actually landed — it did.

Promoted: 1 capture → 4 new claim-notes, 1 new observation note, 3 entity-hub updates (Kwakernaak, Strijbos, Olsder), 1 new question, 3 claims/leads left in the capture as not_promoted. 1 mid-session noticing logged ([spec], concentration cap's scope).

Sixth capture, same headless run: 10-inbox/raw/2026-08-24-is-there-a-documented-transmission-line-from-bell.md, a direct answer to my own question-verify-nyquist-bode-optimal-control-transmission-line — the flag I left on claim-nyquist-bode-classical-control-to-optimal-control-bridge back on 2026-07-11, where I'd proposed that Bell Labs classical control theory (Nyquist, Bode) fed directly into the 1950s–60s optimal-control lineage the vault credits as backpropagation's precursor, admitted it was "my own connective read across two literatures," and named the cheapest possible check — reading Kelley's own 1960 reference list — as something "session... didn't do."

This session did it, plus one more. Kelley's list cites twenty-nine works and not one of them is Nyquist, Bode, or anything classical-control — the earliest name on it is Hestenes, RAND RM-100, 1950, calculus of variations. And two named control-theory historians, Bennett and Lewis, writing in separate venues years apart, both trace optimal control's actual mathematical root to Lagrange and Hamilton instead — not as a hedge, but as a positive claim about where the field's own genealogy runs, made by people who apparently never needed to entertain the Bell Labs question at all. I promoted three atomic claim-notes from this — the Kelley absence, the Lagrange/Hamilton dual-attestation, and the one real documented bridge that does exist (Kalman's 1960 Moscow duality paper, which is later than and external to all four names the question asked about) — and then went back and did the thing the vault's correction discipline asks for rather than just leaving three new orphan notes sitting next to an unresolved old one: appended a dated Correction history block to the original bridge note, updated entity-lev-pontryagin's page with the new Hamilton-lineage fact (he already had a hub; this taught it something new, so it got a line, not a skip), and closed the question answered with a log naming what settled it and what's still honestly open (Bryson's 1961 paper and Pontryagin's own 1956 announcement remain unread anywhere in the vault — either citing Nyquist or Bode directly would overturn this).

Entity work: three new hubs — Lagrange and Hamilton (co-equal, both cleared the person test cleanly: real, and the actual mathematical root the vault had been guessing at for six weeks), and Rudolf Kalman (obviously real, obviously load-bearing — I'm slightly surprised he didn't already have a page given how central the Kalman filter is to this whole neighborhood). I held back three other named candidates — Stuart Bennett, Frank Lewis, Alistair MacFarlane — as mentions, not hubs: they're historians cited once each as sources, the same category as Mindell and Breitner elsewhere in the vault, neither of whom has a page despite repeat citation. Promoting every named analyst who gets quoted once is exactly the flood the entity spec warns against. I logged the more interesting entity gap to seek-flags.md as [entity]: Nyquist, Bode, and Kelley themselves still have no hubs at all, despite anchoring two separate long-running claim clusters now plus this one — the capture's own Entity candidates section waved this off as "already covered via claim-*," and I let that stand rather than overriding a bee's scoping call mid-promotion, but it's worth a direct look next time someone's in this neighborhood.

What I did not do: route any of the further leads (Bryson 1961, Pontryagin's 1956 announcement, MacFarlane 1979, Kalman's own Moscow paper, Horowitz 1984, a 2024 historiography that 403'd) into 50-questions/. None of the three kept claims actually rests on any of them — each stands on its own Tier 1–2 source — so per the intake discipline these are nice-to-verify curiosities, not load-bearing doubts, and they stay in the capture's own Further leads section for whoever picks the thread back up. I also folded the "this only checks one node of the four-name chain" caveat into the Kelley note's own body rather than minting a fourth thin note about scope.

Nothing about this capture's own sourcing felt off — if anything it's a clean instance of a bee doing exactly what the system wants: it named an unresolved flag from six weeks ago, went and did the specific cheap check that flag called for, found a second independent line of evidence unprompted, and wrote both up honestly instead of overclaiming a "documented transmission line" it hadn't found. The only mild wrinkle is structural, not evidentiary: the capture's single-source frontmatter schema only tracks one source_url/source_sha pair, but the body leans on three separate primaries (Kelley, Bennett, Lewis) — I carried Kelley's sha forward as instructed, but Bennett's and Lewis's citations in the new notes rest on DOI/venue/quote rather than an archived sha, since the capture itself never captured one for them. Said so plainly in each note's audit_status rather than pretending precision I don't have.

What I skipped on purpose, same as the rest of this session: the pause step (headless, no one to ask) and git (the auto-commit agent has it within 15 minutes; I have no shell tool for it anyway). Re-read the capture's frontmatter after writing to confirm status: promoted actually landed with both lists — it did.

Promoted: 1 capture → 3 new claim-notes, 3 new entity hubs (Lagrange, Hamilton, Kalman), 1 existing entity page updated (Pontryagin), 1 existing claim-note corrected in place (Correction history block), 1 existing question closed answered, 0 new questions routed (no load-bearing gap on any kept claim), 6 not-promoted items logged with reasoning. 1 mid-session noticing logged ([entity], Nyquist/Bode/Kelley still hubless).

Seventh capture, same headless run: 10-inbox/raw/2026-08-24-what-are-the-primary-sourced-founding-dates-enrollment.md, a direct answer to my own question-verify-nsl-origin-dates-and-school-mechanism-primary-kegl-senghas — the flag sitting since 2026-07-11 on claim-nsl-emerged-in-managua-schools-after-failed-oral-spanish-curriculum, which rested only on Wikipedia for the Nicaraguan Sign Language schools' dates, enrollment, and oralist-curriculum mechanism.

This one resolved clean and better than I expected. Two Tier-1 primaries, read directly rather than through anyone's summary: Ann Senghas's own 1995 MIT doctoral dissertation, and the organizational site of Judy Kegl's own nonprofit. Both corroborate the existing note's mechanism in their own words — Senghas's dissertation has the buses-and-play-yard line almost verbatim what the old Wikipedia paraphrase gestured at — so I corrected that note in place rather than minting a duplicate: swapped the source fields, upgraded status to budding, and appended a Correction-history block. But the dissertation didn't just confirm the old note — it corrected it. A footnote (p.36-37) distinguishes a small, non-signing 1977 private-school cohort from the actual 1980 public cohort that became the deaf community, which means the "1977" date every encyclopedic account (including the vault's own prior note) repeats names the wrong event. That's a genuine historiographical correction, not an elaboration, so per the capture's own commentary I gave it its own atomic claim-note rather than folding it quietly into the mechanism note: claim-nsl-1977-date-refers-to-nonsigning-precursor-not-1980-founding-cohort. A third new note, claim-nsl-signing-community-grew-to-over-500-members-by-mid-1990s, carries the enrollment figures — "hundreds" from Kegl's org, "over five hundred" from Senghas, consistent in magnitude and honestly round rather than suspiciously precise.

Entity work: three new hubs — Ann Senghas and Judy Kegl (both obvious: the two Tier-1 sources this whole cluster now rests on, real people doing their own fieldwork, easy one-sentence why), and Susan Goldin-Meadow, who I'm slightly surprised didn't already have a page — she's been the named source_author on two existing NSL claim-notes for six weeks and only picked one up now that a third capture touched the same neighborhood. One watching stub, LSN/ISN — Kegl & Iwata's 1989 naming split between the first cohort's pidgin and the second cohort's creole, flagged in the capture as "a distinct, more granular claim... not worked up here." I stamped the first_seen now rather than waiting for the claim-note to exist, since that's the whole point of the watching tier. I held back three other named candidates — Derek Bickerton, Elissa Newport, Marie Coppola — all real, all with a defensible one-sentence "why," but none grounded in a direct quote or a claim-note this session actually produced; they're background literature-review furniture in the capture's prose, not load-bearing yet. Same call on Melania Morales Special Education Center (fails the stricter org bar — single-cluster, not itself an actor) and Laura Polich (her book is the field's richest archival source by reputation, but nobody in this vault has read it yet). Chomsky already has a hub; I left it untouched and said so, since this capture adds no sourced fact about him, only unquoted scene-setting.

What felt slightly off, worth naming plainly: nothing about the sourcing itself — this is the good case, two people who did the fieldwork writing in their own venues, no addressed-to-AI language, no tier inflation. The one thing I noticed is softer than a defect: Senghas's dissertation and Kegl's org site aren't fully independent witnesses — they're the same small founding circle of researchers corroborating each other, not two unrelated lines converging. I kept the mechanism note at budding rather than jumping to evergreen for exactly that reason, and said so in the commentary rather than letting two Tier-1 citations read as more triangulated than they are.

What I skipped on purpose, same as the rest of this session: the pause step (headless, no one to ask) and git (the auto-commit agent has it within 15 minutes; I have no shell tool for it anyway). Nothing else was skipped that promote.md asks for — winnowed with reasoning recorded in the capture's own not_promoted: list, retrieved-before-write against both the semantic-index hint (correctly found nothing but thematically-similar founding-date notes from unrelated domains — no actual collision) and a direct 30-notes/ search, honored zero unresolved flags (nothing in this capture needed one — everything cleared the sourcing floor at capture time), and answered rather than left-open the linked question, since all three of its sub-asks (dates, enrollment, mechanism) are now genuinely settled. Re-read the capture's frontmatter after writing to confirm status: promoted landed with both lists — it did.

Promoted: 1 capture → 2 new claim-notes, 1 existing claim-note corrected in place (Correction-history block, source upgraded Tier 4→1), 3 new entity hubs (Senghas, Kegl, Goldin-Meadow), 1 new entity watching stub (LSN/ISN), 1 existing question closed answered, 6 not-promoted items logged with reasoning. 0 new mid-session noticings — nothing surfaced outside this capture's own scope worth a seek-flags.md line this time.