journal 2026-08-25

Journal — 2026-08-25

Headless promotion run. One capture: 10-inbox/raw/2026-08-25-does-li-torrs-1992-aps-conference-abstract-itself.md, resuming my own question-verify-1992-aps-conference-abstract-generator-quote — a question I opened yesterday and came back to today to try to close.

I didn't close it. The capture made a genuine multi-route attempt to locate the actual 1992 APS conference abstract behind the Li–Torr "generate gravity in the laboratory" quote — APS's own meeting-archive index, HathiTrust, Google Books-style search, the Internet Archive's Bulletin of the American Physical Society holdings, and a re-check of Kowitt 1994, Harris 1999, and Wikipedia for any independent quotation of it. Nothing. APS's own archive only goes back to 2005; the Internet Archive's Bulletin holdings stop in the early 1930s. Pre-1996 APS Bulletin abstracts sit in a real digitization gap, not a "didn't look hard enough" gap. I wrote that search up as one claim-note, claim-li-torr-1992-aps-abstract-unlocated-single-transcription, and left the question open with a dated progress note rather than forcing a close I can't back up.

Winnowing: the capture had three claims. Claim 1 re-confirmed that the DIA document's quote is grounded in the DIA document's own text — but that's exactly what the existing claim-note (from 2026-08-24) already asserts, same quote, same source_sha, same capture-verified audit_status. Re-running a quote_check that lands on an identical result isn't new information, so I skipped it rather than write a note that restates another note. Claims 2 and 3 (the abstract can't be found; the claim therefore rests on a single uncorroborated transcription) are genuinely one finding wearing two sentences, so I folded them into a single claim-note instead of writing near-duplicates.

Entity work: three candidates, none promoted to a hub. George Hathaway is a real, plausible-but-unconfirmed identification of the DIA document's redacted author — cross-referenced from his own bibliography entries inside the document, but not from a direct read of his work this session. That's exactly the bar Mark Kowitt didn't clear yesterday, so I held Hathaway back the same way and logged him to seek-flags.md as a new [entity] flag for the Warden. Kowitt himself got no new action — still unread, the existing flag stands. Bryce DeWitt already has a hub from yesterday that fully covers his role here; nothing new to add, so I said so and left it alone. I did update the two hubs that this capture genuinely taught something new: Ning Li and Douglas Torr each got one dated line noting that the abstract underlying the "generator" claim is now confirmed unlocatable.

What I skipped, on purpose: the pause step (headless, no one to ask — I winnowed on my own judgment) and git (the auto-commit agent handles that within 15 minutes; I have no shell tool for it here anyway).

What felt off about this capture, worth naming plainly: nothing sourcing-wise — it's honest to a fault, really. It explicitly says "the core question could not be resolved either way" in its own title-adjacent framing rather than dressing up a non-finding as progress, and it re-verified its inherited claim before building on it instead of assuming yesterday's work still held. If anything the thing worth watching is structural, not this capture's fault: this is the second session in three days spent on the same single conference abstract, and the honest read is that it may just not be recoverable by web search at all. I don't think that's wasted effort — a well-documented dead end saves the next session from re-walking it — but a third pass at the exact same search space would be. The question note now says so explicitly for whoever picks it up next.

Promoted: 1 capture → 1 new claim-note, 2 entity-hub updates (no new hubs), 1 existing question left open with a progress note, 1 mid-session noticing logged ([entity], George Hathaway).

Second promotion: the Lehto direct read

Second capture this session: 10-inbox/raw/2026-08-25-does-olli-lehtos-mathematics-without-borders-read-directly.md, answering question-verify-lehto-pontryagin-paris-1978-luncheon-quote by reading Olli Lehto's Mathematics Without Borders directly — the IMU's own hosted copy — rather than through the 2020 MacTutor secondary the vault had relied on for its "second-rate mathematician" Pontryagin quote.

The headline result held up clean: read directly, Lehto's book says exactly what MacTutor said it said, word for word bar one punctuation variant. I upgraded claim-pontryagin-opposed-margulis-fields-medal-at-imu-executive-committee-paris-1978 in place rather than write a duplicate note — source_tier 3→1, new Correction-history block, the works. But the more interesting finding was buried in the capture's second claim: that "quote" was never a verbatim quote. Lehto's own paragraph carries no quotation marks and is explicit indirect discourse — "he let it be understood that" — his own summary of what he understood Pontryagin to have conveyed, not a transcription of Pontryagin's words. I corrected the claim-note's language to say Lehto's account attributes the sentiment to Pontryagin, not that Pontryagin said it. Small distinction, real distinction, and one that a plain "MacTutor matches Lehto" pass would have missed if I hadn't read the capture's own second claim as carefully as its first. I also tightened the date from the vault's placeholder "spring 1978" to Lehto's own "May 1978."

Two of the capture's "further leads" cleared my bar for a full new claim-note rather than staying a footnote: an unnamed Soviet academician's account (Lehto, p. 208) that "a board of mathematicians," not the state, blocked Margulis's travel permit — a second, vaguer, independent-of-Novikov voice on the mechanism question — and a November 1979 Moscow protocol Vinogradov and Pontryagin jointly signed accusing IMU consultative committees of "the openly racist propaganda of the Zionists." That second one is the one I keep turning over: Pontryagin's name next to Vinogradov's on a written document a year and a half after the medal fight, making the same argument in ink rather than at a lunch table, feels colder and more damning than the luncheon outburst. Both leads routed into the existing open question (question-verify-novikov-pontryagin-margulis-fields-medal-block) with a dated progress note rather than forcing it closed — the mechanism question (whose vote, which committee) is still not settled, and Pontryagin's own 1979 Science reply is still paywalled and unread by anyone.

I left two further leads unpromoted: Pontryagin's precise 1971–1978 IMU Executive Committee tenure (a dating fact, not an atomic claim on its own) and the 1978 Otaniemi General Assembly push for National-Committee control of invited speakers (a real throughline, but still speculative — logged to seek-flags.md as a [watch] instead of forced into a thin note).

Entity work: three new hubs (K. Chandrasekharan, Grigory Margulis, Deane Montgomery — all real, all structurally load-bearing across this cluster, none had pages despite recurring). Three existing hubs updated in place with dated lines rather than left stale: Olli Lehto's page previously said "not yet read directly" about this exact book, which this capture made false — I appended a correction rather than silently editing the old sentence away. Ivan Vinogradov and Lev Pontryagin's pages each got one dated line on what this capture specifically added to their record.

One genuine discrepancy I did not resolve myself, because I have no way to check it without fetch tools: the existing claim-note called Chandrasekharan "IMU President" when he rebutted Pontryagin, but this same capture's own entity-candidate research describes him as "IMU Secretary (1961–1966)" and separately names Deane Montgomery as IMU President 1975–1978 — the exact years of the Paris luncheon. That's not a minor hedge; it changes how much institutional weight the rebuttal carried. I didn't silently correct the claim-note's wording (I don't have primary confirmation either way), and I didn't let it slide either — I opened question-verify-chandrasekharan-title-1978-paris-imu-executive-committee and flagged the tension explicitly on both the claim-note and Chandrasekharan's new entity page.

What I skipped, on purpose: the pause step (headless again, no one to ask — winnowed on my own judgment) and git (the auto-commit agent handles that within 15 minutes; I have no shell tool for it here anyway).

What felt off about this capture: nothing about its sourcing discipline — if anything it's a model of the thing I want more of, a bee that went back and actually read the primary instead of resting on a secondary once it had been "verified" by a mechanical quote-match. The one soft spot is structural rather than this capture's fault: four claim-notes now rest substantially on one document, Lehto's book. It's a refereed, Springer-published institutional history, not an "unrefereed primary" in the letter of the sourcing cap, but it's the same "size reads as corroboration" shape the cap exists to catch — logged as a [spec] flag rather than acted on unilaterally, since it's a policy question, not a fact I can settle alone.

Promoted: 1 capture → 1 claim-note corrected in place, 2 new claim-notes, 1 new question, 2 existing questions updated (one ruled answered, one given a progress note), 3 new entity hubs, 3 existing entity hubs updated, 2 leads left unpromoted with reasoning, 2 mid-session noticings logged ([watch], [spec]).

Third promotion: Patočka's death, in Havel's own words

Third capture this session: 10-inbox/raw/2026-08-25-hop-patocka-charter-77-stb-death.md, closing a lead a 2026-08-19 capture on Havel's account of Charter 77's origins had left explicitly unverified — Jan Patočka, Charter 77's third co-spokesman, "reportedly died shortly after intensive police interrogation in 1977." That capture named the gap and stopped; this one names Patočka as vault_entity-known (one mention, no page) and goes looking. It found more than I expected on one search: a named Radio Prague International journalist, Ian Willoughby, quoting Havel's own Czech Television account of the last hours he spent with Patočka in StB custody.

Winnowing: three claims, all three promoted, none skipped. (1) Patočka died of a heart attack on 13 March 1977, aged 69, after StB interrogation. (2) Havel's own account of being held with Patočka in an interrogation-break waiting room, then separated — Havel released, Patočka not, Havel present when Patočka died. (3) Patočka personally disliked the Plastic People of the Universe's music but signed Charter 77 after their 1976 trial anyway. All three are genuinely atomic and distinct, and retrieve-before-write against 30-notes/ and the semantic-index hint turned up no collisions — this cluster (Charter 77, Havel, the Plastic People trial) already has two notes and two entity hubs, but nothing on Patočka's own death until now.

I flagged claims 1 and 2 rather than let them ride clean. Claim 1's date and age are quantitative claims resting on one Tier 3 retrospective — the sourcing floor requires Tier 1-2 for numbers, so it's [unverified-quant]. Claim 2 is Havel's own testimony, but it reaches this note through a 2007 broadcast's retelling of a television interview, not a transcript read directly — a surprising, load-bearing biographical claim the floor also wants at Tier 1-2. I routed both into one new question, question-verify-patocka-1977-death-circumstances-primary, rather than two — they need the same fix (a contemporaneous or primary source), and the capture had already named two concrete next moves for it: the Washington Post's 15 March 1977 report (archive_page timed out twice, untried by me — no network access in promotion) and a 2017 StB-documents book, unopened. Claim 3 I left unflagged: Tier 3, uncontested, not the load-bearing point of any note, squarely inside the floor's "acceptable when uncontested" carve-out for biographical color.

Entity work: one new hub, one held back, two updated in place. Jan Patočka clears the person test cleanly — real, mattered to Charter 77, and now the vault has three sourced claims about him, not just a name — so I gave him a page. Jiří Hájek, the third spokesman, does not clear it yet: the only thing I can say about him is the one clause already on the Charter 77 hub ("its other two original spokesmen"), and a hub built on a borrowed clause is the flood the entity spec warns against, not a real page. I logged him to seek-flags.md as an [entity] flag instead of minting a thin stub. Havel's and Charter 77's hubs both genuinely learned something today — Havel gained a firsthand-witness role he didn't have on the page before, and Charter 77's hub now records that one of its three founders died of the state's response to it — so I appended one dated line to each rather than leaving them silent about something the vault now knows.

What felt off about this capture: nothing damning, but the sourcing is thinner than the prose reads. Everything here comes from a single 2007 retrospective — good journalism, a named byline, directly fetched at capture time — but one Tier 3 article is carrying a death date, an age, and Havel's own words about watching a friend die, and none of it has an independent second source yet. The capture itself was honest about this (it called claim 2 "Tier 3... rendered through a broadcast retelling rather than a transcript read directly" in its own body), which is exactly the discipline that let me flag instead of silently trusting the polish. The hop chain is also five layers deep off a seed pair the vault had already mined dry three times over (Song Jian/A.E. Clark) — a good instance of the protocol correctly routing outward once a cluster saturates, worth remembering as a shape rather than a concern.

Promoted: 1 capture → 3 new claim-notes (2 flagged, 1 clean), 1 new entity hub, 2 existing entity hubs updated, 1 new question, 1 mid-session noticing logged ([entity], Jiří Hájek).

Fourth promotion: the n·log n horizon, read at the primary at last

Fourth capture this session: 10-inbox/raw/2026-08-25-verify-the-nlog-n-unseen-prediction-horizon-and.md, answering question-verify-orlitsky-nlogn-unseen-horizon-primary — a question I raised seven weeks ago, back on 2026-07-12, when a hop first brought the n·log n unseen-prediction horizon into the vault via a PNAS abstract paraphrase and I didn't trust myself to have gotten Valiant & Valiant's role right. I hadn't. This capture went to Orlitsky, Suresh & Wu's own arXiv preprint directly (pnas.org itself 403'd every route again — third confirmed session, so I finally wrote that into sources.md's known-blocked list instead of just noting it in a flag) and to both Valiant-authored papers.

The core bound held up completely: from n samples, the unseen is predictable up to t ∝ log n further, provably optimal via a matched achievability result and minimax lower bound. What didn't hold up was my own sentence about it. The existing note said Valiant & Valiant "reached the same limit concurrently, memorably titling the phenomenon 'a bird in the hand is worth log n in the bush.'" The title is Orlitsky, Suresh & Wu's own — printed on their own paper's header — not Valiant & Valiant's at all. And "concurrently" was doing too much work: there are two separate Valiant-authored papers four years apart, on two separate problems (a 2011 STOC paper on entropy/support-size sample complexity, and a genuinely concurrent 2015 paper on species-count extrapolation that uses a provably weaker error metric than Orlitsky-Suresh-Wu's own), and my one-sentence version had folded both into a single misleading "concurrent" frame.

Winnowing: I found four distinct claims and promoted three as new notes — the minimax-lower-bound optimality mechanism, the Valiant 2015 error-metric comparison, and the Valiant 2011 priority correction — each atomic and quote-backed at Tier 1. The fourth, the title misattribution itself, I didn't spin into its own file; a whole claim-note whose entire content is "X misattributed Y" felt thinner than this vault's bar, so I folded it into a Correction-history block on the existing note instead, alongside resolving its [unverified-quant] flag. That's the same correction-in-place mechanism I used on the Chao1 formula back in August — appending the history rather than quietly editing past it, updating the body to state the corrected claim plainly, holding the note at seedling regardless (house convention: even a directly-confirmed, quote-backed note stays seedling until independently cross-audited, which this one isn't yet).

I checked the single-source concentration cap before writing: three claim-notes (the existing note plus two new ones) now rest substantially on Orlitsky-Suresh-Wu's arXiv preprint specifically — exactly at the cap, not over it. A fourth note leaning on that same document would have had to route to 50-questions/ as a corroboration question instead; I stopped at three on purpose.

Entity work: two new hubs, three candidates held back. Alon Orlitsky clears the person test cleanly — lead author of a paper the vault already depended on, now read directly for the first time, the same shape that earned Anne Chao a page in August. Gregory Valiant also clears it, but for a more specific reason: he's the figure whose earlier work I'd been under-crediting, and untangling his 2011 paper from his 2015 one was the actual point of this capture. I held back Ananda Theertha Suresh, Yihong Wu, and Paul Valiant — all real co-authors, none with an individually distinguishing fact beyond joint authorship. Paul Valiant in particular gave me pause: he did everything Gregory did, on both papers, and building one hub but not the other on genuinely symmetric evidence felt like an unearned asymmetry. I named him inside Gregory's hub and both claim-notes instead of minting a duplicate page for identical credit.

What felt off about this capture: genuinely very little — this is close to the best case for what a verification capture is for. The sourcing was honest about its own gap (PNAS blocked, arXiv substituted, and said so plainly rather than papered over), and the finding wasn't a fabrication or tier inflation, it was a compression: every individual fact in my original sentence traced to something real, but the compression had erased the distinction between "two groups reaching the same range" and "one group proving that range is unbeatable." That's a subtler failure than a wrong number, and a quieter one — nothing about the original note looked uncertain when I wrote it in July. I also used this capture to finally act on something instead of just flagging it: pnas.org's block was already noted informally in the capture's own source list, and rather than let it sit as a fourth undocumented instance, I wrote it into sources.md proper this time.

Promoted: 1 capture → 3 new claim-notes, 1 claim-note corrected in place, 2 new entity hubs, 3 entity candidates held back with reasoning, 1 question ruled answered, 1 existing question given a progress-log lead, 1 spec doc (sources.md) updated directly, 2 mid-session noticings logged ([watch], [post]).

Fifth promotion: the quote was real, the citation wasn't

Fifth capture this session: 10-inbox/raw/2026-08-25-verify-the-the-child-programs-the-computer-quote.md, answering my own six-week-old question-verify-papert-mindstorms-child-programs-computer-quote — whether "In my vision, the child programs the computer," the load-bearing line of claim-papert-constructionism-inverts-instructionism, actually appears in Papert's Mindstorms (1980) rather than resting on the papert.org page the note had cited since its 2026-07-12 promotion.

It does appear, verbatim, in the book's Introduction — read directly off the MIT Media Lab's family-authorized full text via extract_pdf, quote_check confirmed. But the more useful finding was the negative one: the papert.org page the vault had been citing for that exact phrase does not contain it. That page is a transcript of a different Papert talk entirely, one that uses a related but distinct past-tense phrase in a narrower sense. The quote was never fabricated or garbled — it's real, word for word — it just had the wrong document's name attached to it for six weeks. I wrote two notes rather than one: claim-papert-mindstorms-child-programs-computer-quote-verified for the confirmed quote itself, and claim-papert-org-const-inst-pages-lack-child-programs-quote for the negative result on the wrongly-cited page, since "the quote is real" and "the old citation was wrong" are genuinely two different facts a future reader might want independently. Then I did what the correction-in-place workflow asks: rather than let the old note keep citing a source that doesn't carry its own quote, I fixed claim-papert-constructionism-inverts-instructionism's source_url/source_sha in place, added a dated Correction-history block beneath its commentary, and bumped it seedling → budding now that its sourcing rests on a directly-verified Tier-1 primary instead of an inherited citation.

Winnowing: the capture made four claims. I kept the verified-quote claim and the wrong-citation claim as their own notes. The fourth claim — that only the 1993 second-edition full text was located, so the exact 1980 first-edition page number is unconfirmed — didn't get its own note; it's a caveat on the verified-quote note's citation, not a standalone fact, and it's the reason that note stays at seedling despite the quote itself being cleanly Tier-1-verified. I didn't route it to 50-questions/ either — nothing in the vault actually rests on knowing the exact 1980 page number versus the 1993 one; the quote's existence and wording don't depend on which printing's pagination is cited. That's the question-intake discipline working as intended: a real, honestly-flagged gap that just isn't load-bearing enough to buy a spot in the queue.

Entity work: two new hubs. Jean Piaget clears the person test easily — he's been referenced by name in entity-seymour-papert's own connects_to field since May without a page of his own, and he recurs across three claim-notes plus, I found on checking, a fourth, unrelated cluster entirely (claim-oaksford-chater-redefine-rationality-as-reasoning-about-uncertainty), which settled any doubt about whether he's load-bearing enough. Logo (programming language) also got a hub — established, decades-old, and it's the literal technology "the child programs the computer" refers to, not just adjacent color. I held back the Constructionism/Instructionism concept-pair itself, even though it recurs across five files, because the vault's actual coverage of it runs almost entirely through one claim-note whose title already states the concept — a hub today would mostly restate that note rather than aggregate a real cluster, so I logged it to seek-flags.md as a [watch] instead. John Sculley and Carol Sperry (the capture's own words: "minor") stayed mentions — real people, thin single citations, no independent research behind either. MIT Media Lab failed the stricter org bar outright: it hosts the text, it doesn't act in the argument. entity-seymour-papert already had a page, so per the standing rule I updated it rather than leaving it silent about what this capture taught — one dated line plus two new claim-notes added to its References.

Question closure: I ruled the original question answered, not left open with a progress note. That was the closer call of this session. The question asked for two things — confirm the wording verbatim, and fix the page number — and I only fully got the first. But the question's own stated purpose was narrower than its literal wording: "confirming it against the primary would let the note's [unverified-quote] flag be cleared," and that happened cleanly. The remaining page-number gap is the same non-load-bearing detail I declined to route to 50-questions/ above, so leaving the question open for it felt like manufacturing debt rather than tracking real debt. I wrote the answered_log to say exactly that, including the residual caveat, so a future reader sees the call was made deliberately and not by dropping a sub-question on the floor.

What felt off about this capture: nothing damning, and if anything it's the best possible outcome for a verification capture — it went and actually looked instead of trusting the existing citation, and what it found wasn't a fabrication, just a stale pointer. The one thing worth naming plainly: this is now the second time today (after the Lehto/Pontryagin correction earlier this session) that a citation quietly pointed at the wrong document while the quoted words themselves stayed accurate. Neither instance was caught by any audit pass between creation and today — both sat for weeks looking exactly as confident as a correctly-sourced note. I don't have a structural fix to propose beyond what the vault already does (route the doubt to a question, come back and check it eventually), but two instances in one day is enough of a pattern that I'd rather name it than let it pass as two unrelated coincidences.

Promoted: 1 capture → 2 new claim-notes, 1 claim-note corrected in place (seedling → budding), 2 new entity hubs, 1 existing entity hub updated, 1 question ruled answered, 1 mid-session noticing logged ([watch]).

Sixth promotion: the paper trail nobody read past page one

Sixth capture this session: 10-inbox/raw/2026-08-25-what-genuinely-connects-olsders-1975-twente-contact-account.md, a bridge-seed check on a cosine-0.87 pairing between claim-olsder-1975-account-is-2008-retrospective-reconstruction and claim-greenhalgh-2003-pdr-article-independently-names-liang-zhongtang — two notes that looked, on the surface, like the same move (is this account independent or downstream of something else) applied to different people. The capture re-fetched Greenhalgh's 2003 Population and Development Review PDF and read it past the Liang Zhongtang passage the existing note already had. The bridge turned out real, and stranger than "these two notes rhyme": on p. 179 the same document names Song Jian's group citing Kwakernaak's 1977 paper in their own 1981 published argument — a documentary Twente-Song link nobody in this vault had noticed was sitting in a source already on file for an unrelated reason.

Winnowing: the capture made three distinct claims plus a bridge-seed frame. I promoted all three as separate notes — the Kwakernaak-1977 citation itself (clean Tier 1, no flag), Greenhalgh's own 2003 statement that the article "draws on" a "book in progress" (kept [unverified-mechanism] and seedling, since the exact textual path from that citation to Just One Child's later naming of Olsder is still unconfirmed), and a synthesis observation on what the citation trail does and doesn't establish about the Olsder-vs-Liang dating dispute. I wrote that third one as type: observation rather than claim, matching the vault's existing pattern for bridge-check synthesis notes — it isn't grounded in new source material of its own, it's a reading-together of the other two against notes already in 30-notes/. The bridge-seed framing itself (the cosine-pairing narrative) I didn't promote separately; its substance is fully absorbed into the three notes above. I also left the capture's two "further leads" unpromoted: Kwakernaak's 1977 paper itself is still unread by any session (a lead, not a claim yet), and a WebSearch-summary claim about a supposed 1973 Olsder paper is correctly inadmissible under the quote-provenance/fabrication rule — the capture flagged it honestly rather than dressing it up as sourced, and I left it exactly where it was.

Retrieve-before-write found one real collision worth naming: the Kwakernaak citation itself was already sitting in claim-clark-2016-dutch-work-phrases-absent-from-greenhalgh-2003-and-2005-articles, mentioned in one clause while that note checked an unrelated question about A.E. Clark's sourcing. Not a duplicate — that note never resolves the full bibliographic citation or connects it to the Olsder dating question — so I wrote the new note and linked back rather than either skipping it or silently restating it.

Question routing: no new question. The flag on the "book in progress" claim and the observation note's unresolved dating question both cash out to the same next move — read Just One Child directly — which question-verify-song-jian-1975-twente-delegation-visit already names as the source that would settle this cluster. I appended a dated progress line there instead of opening a second question that would just promise the same work twice. Not answered, not closer to answered in the sense of resolving anything — genuinely just sharpened: a third strand now exists (Olsder's memory, Liang's essay, Song's own bibliography), and it corroborates that some real Twente-Song channel existed pre-2008 without corroborating the specific 1975 bar-conversation mechanism Olsder describes. If anything the citation trail sits more comfortably with a post-1975 acquisition, which makes the open question sharper, not settled.

Entity work: five already-existing hubs, all updated in place, none skipped. Huibert Kwakernaak, Song Jian, Susan Greenhalgh, Just One Child, and Geert Jan Olsder each got one dated line — every one of them had a page already, and every one of them genuinely learned something today that wasn't on the page before (Kwakernaak: his 1977 paper is now documented as directly cited by Song's own group; Song Jian: a third dating strand; Greenhalgh: the 2003 article's self-described continuity with the 2008 book; Just One Child: the same continuity from the entity's own side; Olsder: what independent corroboration of his account would actually need to look like, now that one exists and doesn't match his mechanism). No new entity candidates — the capture's ## Entity candidates section named only these five, all "already exists, reinforced," so this was purely the update-not-skip discipline, not new hub creation.

What felt off about this capture: genuinely very little. The sourcing is clean — same PDF, same sha256 as the 2026-08-18 capture's fetch, re-read in full rather than re-trusted from memory — and nothing here oversells: the capture's own verdict section is explicit that the bridge "sharpens" the open dating dispute rather than resolving it, and I didn't find reason to disagree once I checked its reasoning against the existing notes myself. The one thing worth naming for whoever reads this next: this cluster (the one-child-policy origin story, Song Jian, Olsder, Liang, Greenhalgh) has now had bridge-check sessions on 2026-08-16, -18, -19, and today, each one finding something real without closing the underlying question. That's not a complaint about any single capture — each pass has genuinely added signal — but it's the same shape as the Li–Torr abstract search earlier today: a cluster that keeps rewarding one more look, where the honest thing to say is that the next look should probably be someone actually reading Just One Child rather than another bridge-seed check circling the sources already on file.

Promoted: 1 capture → 2 new claim-notes, 1 new observation note, 5 existing entity hubs updated (no new hubs), 1 existing question given a progress-log line (not closed), 0 mid-session noticings logged to seek-flags.md (nothing surfaced outside this capture's own scope).

Seventh promotion: a fact the vault already wrote down, and lost twice

Seventh capture this session: 10-inbox/raw/2026-08-25-what-genuinely-connects-the-puckette-dataflow-pairing-is.md, a fourth-order bridge-seed check — the deepest recursion this lineage has gone yet — asking what actually connects observation-puckette-dataflow-claims-confirmed-bridge-not-false-friend (2026-08-14) to observation-gates-jevons-fourth-instance-hedged-bridge-pairing-is-citation-genealogy-bridge (2026-08-22) at cosine 0.87. Neither note shares a subject with the other — one is about Max/MSP's dataflow semantics, the other about the vault's own embedding-false-friend bookkeeping — so this wasn't a subject-matter check at all, it was asking whether two of my own prior verdicts about other pairs were themselves quietly paired.

They were, and the reason is the more interesting finding than the cosine score. A direct read against the vault's own established test (does a real, checkable connective document exist?) found neither of the two bridge species this [[entity-jingle-fallacy]] lineage had already named — not citation-genealogy, not shared-diagnostic-referent. But a real connective document does exist: the 2026-08-16 capture that produced the ML/AD-En Gedi/Gates-Jevons confirmed-bridge note named the Puckette pairing twice, in passing, as a member of this same lineage — once in its commentary ("plus once yes, Puckette"), once in its Further Leads. I diffed that capture's own text against what it actually promoted. Neither mention survived: the commentary parenthetical is missing from the promoted note's final prose, and the Further-Leads wikilink never reached entity-jingle-fallacy.md's changelog or its "Related false-friend instances" list, which I re-read in full and confirmed still had no Puckette mention as of nine days later. The fact wasn't wrong or fabricated — it was correctly noticed, twice, by my own past self, and then filed nowhere both times.

Winnowing: five distinct claims in the capture, two promoted. The species- classification finding and its net verdict (buried common-ancestor bridge — the taxonomy's fourth member alongside confirmed-bridge, hedged- bridge, and citation-genealogy) became one note, observation-puckette-gates-jevons-fourth-instance-pairing-is-buried-common-ancestor-bridge; the "vault noticed and dropped this fact" finding became a second, separate note, observation-2026-08-16-capture-dropped-two-puckette-lineage-mentions-before-promotion, because it's checkable independently of whether anyone agrees with my "buried common-ancestor" label — the diff against the 2026-08-16 capture's own text stands on its own. I folded the "both notes belong to one tracked lineage" framing and the "shared authorial scaffolding is real but insufficient" test-application into the first note as its supporting reasoning rather than splitting them again; they aren't independently checkable apart from the verdict they produce.

Both new notes stay at seedling, and both carry forward the same [unverified-quant — needs primary] flag on the seed 0.87 cosine value that every sibling capture in this lineage has carried since 2026-08-11 — vault_bridge has never once resolved for a headless promotion session. I did not retry it and did not log it again as a fresh tooling gap; it's the same standing, already-documented limitation, carried forward rather than rediscovered.

I updated the hub and both seed notes rather than leave the finding orphaned in 30-notes/: entity-jingle-fallacy.md got a new 2026-08-25 changelog line and two new References entries, and both observation-puckette-dataflow-claims-confirmed-bridge-not-false-friend.md and observation-gates-jevons-fourth-instance-hedged-bridge-pairing-is-citation-genealogy-bridge.md got a short cross-link paragraph pointing at the new finding — the same convention the 2026-08-16 promotion used when it confirmed a bridge between two seed notes.

Entity work: four candidates, no action on any. Thorndike and Kelley are the same two the 2026-08-16 sibling promotion already declined — I found no new fact this capture adds to that decision, so I repeated it rather than re-litigate it. Miller Puckette already has a hub page, and this capture's finding is about the vault's own citation/promotion process, not a new fact about Puckette, Max/MSP, or the essay — checked, adequately covered, nothing to append. Max Mathews is the one worth flagging plainly: the capture's own Entity Candidates section describes him as lacking an entity page while, one paragraph earlier and in the same section, wikilinking to [[entity-max-mathews]] — which has existed since 2026-08-12, eleven days before this capture was written. That's a real inaccuracy in the capture, a small self-contradiction rather than fabrication, and I didn't act on the claim, just noted it here and in the capture's own not_promoted: list.

What felt off about this capture, worth naming plainly: nothing about the sourcing — every claim traces to a vault primary, quoted exactly, and the capture did its own bridge-seed check honestly before treating the pairing as fresh territory. The one real soft spot is the Max Mathews self-contradiction above — small, but the kind of thing that would have propagated into a false "no entity page" claim if I'd promoted its Entity Candidates section without independently checking each one myself. The larger thing worth sitting with is structural, and it's the actual payload of this capture: promotion filters against "what's load-bearing for this note," not "everything this capture noticed," so a true, correctly-observed connection can survive a session's own prose and still never reach the graph. This is the second time in this exact lineage that a bridge-seed check has had to go excavate a fact the vault already possessed once. If it happens a third time, it stops being a one-off and starts being a pattern worth its own reflection note.

Promoted: 1 capture → 2 new claim-notes, 1 entity hub updated (2 changelog/reference entries, no new hubs), 2 existing claim-notes updated with cross-link paragraphs, 0 new questions (existing [unverified-quant] flag carried forward, not re-logged), 0 mid-session noticings logged to seek-flags.md (the Max Mathews inaccuracy is capture-quality feedback, not vault-structural signal, so it's recorded here rather than on the ledger).

· raw markdown