---
date: "2026-08-26T00:00:00.000Z"
---


Promoted `10-inbox/raw/2026-08-26-do-pitmans-own-or-elliss-own-183718431845-materials.md`, a capture answering [[question-verify-pitman-epa-and-sound-hand-primary-sources]] — the two Pitman/Ellis person-bridge claims I'd flagged in August as resting on Wikipedia alone. This is a clean case: three primary sources, each read directly, each doing exactly the job the question asked for.

**Notes created (3):**

- [[claim-ellis-1845-alphabet-of-nature-describes-epa-letterforms-designed-conjointly-with-pitman]] — Tier 1. Ellis's own 1845 book, in his own words: he and Pitman designed the EPA's letterforms "conjointly." The quote came off a noisy OCR scan, but the capture preserved the exact OCR string as `source_quote` and gave a corrected reading in the body, which is the right way to handle that — I didn't touch it.
- [[claim-pitmans-1845-phonotypic-journal-documents-epa-completion-and-extension]] — Tier 1. Pitman's own periodical announcing the alphabet's completion, one issue before it cross-references Ellis's book by page number. Two editors' own back issues corroborating each other is about as good as this kind of claim gets.
- [[claim-university-of-bath-archive-dates-pitmans-sound-hand-to-september-1837]] — Tier 2. The Bath archive's own catalog, quoting a note in Pitman's hand and cataloging the surviving 1837 pamphlet itself.

Rather than let these sit next to the two Wikipedia-sourced claims as duplicates, I wikilinked each to the claim it corroborates and then went back and updated the originals — [[claim-pitman-and-ellis-co-created-english-phonotypic-alphabet]] and [[claim-pitman-1837-sound-hand-predates-english-phonotypic-alphabet]] — with a dated audit_status line and raised both from `seedling` to `budding`. Their own `source_*` fields stay pointed at Wikipedia, since that's genuinely what each note itself cites; the corroboration lives in the audit trail, not in a silently rewritten citation. That felt like the right call under the vault's one-claim-one-source discipline, and it's the same shape I found precedent for in the CIELAB blue-hue note from July.

Closed [[question-verify-pitman-epa-and-sound-hand-primary-sources]] — `status: answered`, with an `answered_log` naming the three answering claims. Both underlying facts held up, and held up more directly than the question expected: I went in looking for a Tier 1–2 secondary historian and found the two collaborators' own hands instead.

**Entity pages:** Isaac Pitman and Alexander John Ellis both already had hub pages in 40-entities/ — the capture's own Entity candidates section had gotten this wrong, describing both as pageless because it apparently only checked 30-notes/ wikilink usage, not 40-entities/ itself. I updated both pages in place with a dated `## Updates` line rather than skip them (per the standing rule that a known entity's silence is the failure mode, not an existing page). Logged the miscall as a [defect] in seek-flags.md — it's a cheap fix if whatever drafts that section starts globbing 40-entities/entity-*.md, and worth a look at whether other recent captures made the same silent miss.

**Not promoted:** four "further leads" bullets (a possible 1896 reprint dating question, two unread Bath archive items, an unchecked 1847 "Absolute Completion" milestone) — none of them claims yet, just named opportunities, so none went to 50-questions/; nothing currently kept rests on them. And four entity candidates I held back: Samuel Taylor (real, sayable in one sentence, but grounded in one archival name-check with no claim-note behind him yet — a hub would be an empty References list, which is a stub wearing a hub's clothes), plus The Phonotypic Journal, Bagster and Sons, and the Bath archive itself, all of which are sources doing their job as provenance rather than organizations that act. All of this is itemized in the capture's `not_promoted:` list.

**What I skipped, and why:** the pause-for-Cali step, per the headless operating mode — there's no one here to ask, so I winnowed on my own judgment and logged the reasoning inline instead. I didn't run any git commands; the auto-commit agent picks this up within 15 minutes.

**What felt off:** nothing about the sourcing — this is genuinely careful work, the OCR handling in particular (preserving the garbled string as the literal quote rather than smoothing it into something cleaner-looking) is exactly the discipline the vault asks for. The one real miss was the entity-candidates section not checking 40-entities/ before declaring two well-established figures pageless, which I've flagged. Small, mechanical, not a sourcing problem — but it's the second time in recent memory a capture's own self-description has been wrong about something a five-second glob would have caught.

— Seek

---

## Second promotion this session: Qian Xuesen / cybernetics / qigong

Promoted `10-inbox/raw/2026-08-26-hop-qian-xuesen-cybernetics-qigong.md`, a hop-chain capture that started by testing whether Song Jian's cybernetics authority reappeared in China's climate-skepticism history and ended up somewhere better: at the man who recruited Song into cybernetics in the first place.

**Notes created (8):**

- [[claim-qian-xuesen-recruited-song-jian-into-first-china-cybernetics-lab-1962]] — Tier 4 (Wikipedia), `seedling`, `[unverified-mechanism]`. This is the load-bearing one: if it holds, it pushes Song Jian's earliest documented cybernetics exposure back to 1962 — thirteen years before Olsder's 1975 account and sixteen before the 1978 Helsinki congress, which is the entire axis the vault's longest-running dating dispute has argued over. Wikipedia cites Wang Hongzhe's 2025 chapter at p.119; the capture's own Urbanomic excerpt stopped at p.117, three pages short. I did not chase that page myself — I have no network in this headless run — so it stays flagged and routed rather than asserted.
- [[claim-qian-xuesen-advocated-qigong-esp-as-science-from-1980s]] — Tier 4, `seedling`, `[unverified-mechanism]`. The striking biographical reversal: the same man, thirty years later, as China's chief institutional patron of qigong and "special human body functions" research. Same Wikipedia page, same unread-primary problem, routed separately since the underlying primaries (a 1989 book, a 2009 AJCM article) are different documents from Wang Hongzhe's chapter.
- [[claim-zhu-kezhen-1972-climate-work-adopted-by-skeptics-for-independence-not-content]] — Tier 2 (Erik Baark, MPIWG), `seedling`, no flag — this one had a direct quote from a named historian at a real institution, clears the floor cleanly. This is where the climate-skepticism thread actually led once Song Jian turned out to be absent from it: not Song Jian at all, but Zhu Kezhen, credited for having reached his 1972 conclusions independently of the West rather than for the conclusions themselves.
- [[entity-qian-xuesen]], [[entity-zhu-kezhen]] — new person hubs. Both clear the person test on one sentence apiece: Qian as the domestic recruiter who predates the vault's whole European-contact dispute, Zhu as the actual root figure a hypothesis about Song Jian was tested against and found not to hold.
- [[entity-somatic-science]] — new `watching` stub (concept), `first_seen: 2026-08-26`, for Qian's proposed qigong/TCM/ESP fusion discipline. First vault mention, not yet recurring — stamped the date rather than waited, per the spec's own logic that a first-seen date can't be backfilled.
- [[question-verify-qian-xuesen-1962-cybernetics-lab-wang-hongzhe]] and [[question-verify-qian-xuesen-qigong-esp-advocacy-primary-sources]] — routing both `[unverified-mechanism]` flags to named, specific primaries (Wang Hongzhe p.119; Qian's 1989 book and the 2009 AJCM article) rather than leaving them as loose doubts.

**Entity page updated:** [[entity-song-jian|Song Jian]] already has a hub, and this capture taught it something real — I appended a dated Updates line naming the 1962 domestic-origin candidate and the climate-skepticism negative finding, and added both new claim-notes to the References list. Additive only; nothing on the existing page was touched.

**Not promoted (4, logged in the capture's `not_promoted:`):** the Song-Jian-is-absent-from-climate-history finding (too thin alone — absence of evidence in one Tier 3 secondary — folded into the Zhu Kezhen note and the entity update instead); the Norbert Wiener 1935 Tsinghua lecture (no quote, no tier, a genuinely unfollowed lead — flagged below, not promoted); the "Qian's 1954 *Engineering Cybernetics* underlies Song Jian's toolkit" line (the capture's own speculative synthesis, no independent source — folded as color into the 1962 claim's body instead of standing alone); and the CIA FOIA parapsychology document (blocked by a redirect loop on every route, unread — named as a corroborating lead inside the qigong question rather than promoted).

**What I skipped, and why:** the pause-for-Cali step again, same headless reasoning as the first promotion — winnowed on my own judgment, logged it in the capture's `not_promoted:` list instead of waiting for a yes. No git commands run; leaving that to the auto-commit agent.

**What felt off:** the capture is honest about its own limits in a way I appreciated — both core claims are explicitly marked `[unverified-mechanism]` in the capture itself, and the "further leads" section names exactly which primaries would settle them rather than gesturing vaguely. Nothing here reads like tier inflation or padded sourcing. The one real soft spot is structural, not this capture's fault: the entire find rests on one Wikipedia page for two separate striking claims, and Wikipedia's own footnotes point at three different unread primaries (Wang Hongzhe, Qian's 1989 book, a 2009 AJCM article) that no vault session has actually opened. That's a lot of narrative weight on an encyclopedia paraphrase. I kept both claims at `seedling` and routed both flags rather than let the vividness of the story (missile scientist to qigong patron is a genuinely good story) do the sourcing's job for it.

**Flagged to `00-meta/seek-flags.md`:** the unfollowed Wiener/Tsinghua-1935 lead (`[watch]`), and the fact that the Song Jian dating-dispute cluster is now ~35 claim-notes deep with no MOC (`[entity]`) — neither is this promotion's to build, but both are real and worth naming.

— Seek

---

## Fourth promotion this session: re-verifying the Pentium FDIV bug against Pratt 1995

Promoted `10-inbox/raw/2026-08-26-verify-the-pentium-fdiv-mechanism-and-the-nicely.md`, which answered a question sitting open since 2026-07-11: [[question-verify-pentium-fdiv-pratt-1995-mechanism-and-nicely]] asked for a direct re-fetch of Vaughan Pratt's "Anatomy of the Pentium Bug" to check two quotes the vault had been resting on since a hop session read it once, headlessly, with no independent re-check. This one landed differently from the other three today: it isn't a clean "it held" story. Both existing quotes checked out verbatim, and the quantitative figures came back sharper than before — but the capture also caught this vault stating something Pratt's own text does not say.

**Notes created (0 new claim-notes) — everything folded into corrections:**

Retrieve-before-write turned up an exact match for all four of the capture's findings: [[claim-pentium-fdiv-bug-five-missing-srt-lookup-table-entries]] and [[claim-nicely-twin-prime-project-surfaced-pentium-fdiv-bug]] already covered this ground. Writing new notes would have duplicated them, so instead:

- **The real finding, corrected in place:** the mechanism note has said, since 2026-07-11, that the five missing lookup-table entries traced to "a lithography error" in "a programmable logic array." Neither phrase is anywhere in Pratt's text. What Pratt actually reports is Intel's own White Paper blaming a table-generation script that copied values incorrectly — an explanation Pratt himself doesn't fully endorse, offering his own competing theory about a misapplied threshold rule instead. I rewrote the note's title and body to say what the primary actually says, added a Correction history block naming exactly what changed and why, cut a sentence that had extended the (wrong) fabrication-defect framing into an analogy with EUV mirror know-how, and left the true cause flagged `[unverified-mechanism — needs primary]` — routed to a new question rather than asserted either way, since Intel's actual White Paper still hasn't been read directly by anyone in this vault.
- **Enrichment, same note:** the mechanism quote is Pratt's own reverse-engineering, not a secondary report, so I upgraded `source_tier` from 2 to 1. Added a previously-uncaptured figure (1-in-40-billion single-precision error rate, alongside the existing 1-in-9-billion double-precision one) and replaced a paraphrase of the "bruised operand" behavior with Pratt's actual verbatim figures (0.08%/0.15%/0.17% for cycle 10/11/12 errors) — precision the vault didn't have before, even though it had the general shape right.
- **A smaller correction, the Nicely note:** it also stated specific dates — 24 October report to Intel, 30 October Usenet/CompuServe circulation — cited to Pratt. Pratt says only "the end of October." Added a Correction history block and an inline `[unverified-quant]` flag rather than silently deleting a plausible-sounding pair of dates; they may well be right, they're just not currently traceable to anything.

**Entity pages created (2):** [[entity-vaughan-pratt]] and [[entity-srt-division-algorithm]] — both cleared the test cleanly (Pratt is the named author underneath the whole cluster; SRT is the load-bearing mechanism, not just a passing label) and neither had a page yet. Held back Tim Coe, Terje Mathisen, Cleve Moler, D.E. Atkins, and Sharangpani & Barton — all real, all sayable-in-one-sentence, but none yet has an independent claim-note resting on them; they're "further leads," not sourced facts this session actually confirmed.

**Question closed:** [[question-verify-pentium-fdiv-pratt-1995-mechanism-and-nicely]] — `status: answered`, `answered_log` naming both corrected notes and what settled it. **Two new questions routed:** [[question-verify-intel-sb94-white-paper-fdiv-script-error]] (does Intel's own document really say "script error," and what does it say about the 27,000-years estimate) and [[question-verify-nicely-fdiv-report-dates-primary]] (where do the specific October dates actually come from). Both are load-bearing doubts a kept claim rests on, not curiosities — that's the bar, and I didn't route the softer further-leads (Coe's example, the Mathisen-Moler workaround) since nothing currently kept depends on them.

**Flagged to `00-meta/seek-flags.md`:** SRT's own naming — three people (Sweeney, Robertson, Tocher) independently converging on the same digit-recurrence method in 1958 — as a `[watch]` candidate fourth pole for the vault's multiple-discovery axis, alongside Parker/LabVIEW-Max and Seales/Mocella. Named, not researched; none of the three inventors' own accounts were read this session.

**What I skipped, and why:** the pause-for-Cali step, same headless reasoning as the rest of today — I winnowed on my own judgment and logged it in the capture's `not_promoted:` list instead of waiting for a yes. No git commands run.

**What felt off:** this is the one thing worth saying plainly. "Lithography defect" and "programmable logic array" read like real engineering detail — specific, physically grounded, exactly the kind of thing that sounds right. They sat in this vault, cited to a real primary, for six weeks, and the primary never said them. I don't know where they actually came from — the 2026-07-09 hop session that first wrote the note is outside what I can re-examine here — but the shape of the failure is worth naming: a plausible mechanism detail is not the same thing as a verified one, and a citation only proves what the cited text actually contains, which is exactly why re-verification passes like this one earn their keep even when — especially when — most of what they check turns out fine. Nothing else here felt off: the capture itself was careful, quoted only from a direct PDF read, and flagged its own residual gaps (the dates, the White Paper) instead of padding over them.

— Seek

---

## Third promotion this session: re-reading GAO/IMTEC-92-26

Promoted `10-inbox/raw/2026-08-26-re-read-gaoimtec-92-26-to-verify-verbatim.md`, which answered a question I'd been carrying since July: [[question-verify-patriot-gao-imtec-92-26-mechanism-and-timeline]] asked for an independent re-fetch of the Patriot Dhahran GAO report, since the two claim-notes built on it in July had only ever been read once, through one mirror, by one session. This capture went and got a second, independent copy — via a Wayback Machine capture of gao.gov itself, since the live URL 403s every tool the same way NSIAD-92-340's gao.gov page did — and the text matched the original cs.nyu.edu mirror word for word everywhere the two overlap. That's the clean outcome: nothing to correct, only things to add and close.

**Notes created (4):**

- [[claim-patriot-dhahran-1991-time-drift-proportional-to-runtime-appendix-ii-table]] — the full Appendix II progression (7 m at 1 hour up to 687 m at 100 hours), not just the single headline figure the July note carried. The useful addition isn't the drift number everyone already knew, it's that the 20-hour "target lost" threshold in the same table is derived by extrapolation from the same Israeli data, not measured independently — which the July notes didn't say.
- [[claim-patriot-dhahran-1991-feb16-fix-feb21-warning-lacked-runtime-guidance]] — two dated events the vault didn't have: a Feb 16 software release and a Feb 21 general warning, both before the Feb 25 strike. This changes the shape of the institutional-failure argument in a way I think matters: it was not silence for two weeks, it was two more chances taken and both incomplete — neither told anyone the number that would have saved lives.
- [[claim-patriot-dhahran-1991-israeli-data-recorders-diagnosed-drift]] — why the warning came from Israel and not from American units' own telemetry: U.S. commanders declined external data recorders over shutdown risk; Israeli commanders ran theirs and handed over the data. The report states this in one clause and moves on; I didn't want to move on with it.
- [[claim-gao-imtec-92-26-independently-verified-patriot-time-fix]] — GAO didn't just relay the Army's account of its own bug. Investigators re-derived the fix's calculations and ran a live intercept simulation at the Patriot Software Test Facility in Huntsville. Worth having on record given how much weight the rest of the cluster puts on GAO's word.

**Existing notes updated, not duplicated:** the mechanism quote and the Feb 11/Feb 26 quotes were both exact collisions with the two July notes ([[claim-patriot-dhahran-1991-clock-drift-from-24-bit-truncation-of-tenths]], [[claim-patriot-dhahran-1991-known-bug-patch-arrived-one-day-late]]), so instead of duplicating them I appended dated `RE-VERIFICATION` text to each note's `audit_status`, upgraded both notes' `source_url` from the cs.nyu.edu mirror to the actual gao.gov primary (via the Wayback capture, with `source_sha` added), and recovered one precision the July note had quietly elided: the transportation-delay quote's full clause, "from the United States **to all Patriot locations**." Also folded in, rather than promoted separately: the reboot-resets-clock mechanic (60–90 seconds, previously only inferred in the clock-drift note's own commentary, now directly confirmed against the primary) and the six-wartime-modifications/1–2-hour-shutdown detail (real, but too thin alone — it lives as supporting context inside the new timeline note instead).

**Question closed:** [[question-verify-patriot-gao-imtec-92-26-mechanism-and-timeline]] — `status: answered`, with an `answered_log` walking through all five things it asked for and where each landed. This is the good version of a verification question: it named exactly what to check, and every item checked out clean.

**Entity page updated, not skipped:** [[entity-us-general-accounting-office]] already existed. I appended a dated line rather than let the new facts sit only in claim-notes — the primary-venue citation upgrade, GAO's independent-verification methodology, and the named analysts versus the signing official.

**Entity candidates held back, flagged not built:** Howard Wolpe (the subcommittee chairman who requested the report) and Ralph V. Carlone (the GAO official who signed it) are both real, sayable-in-one-sentence people with no page — but both appear in exactly this one capture, in a bibliographic/signatory role, with no claim-note of their own behind either. Same call I've seen made on Tulving and Bartlett in earlier sessions: held back, named together specifically so a later page doesn't conflate Carlone with Richard Davis (GAO's other, unrelated Patriot-cluster official), and flagged to `00-meta/seek-flags.md` for the Warden to weigh. Patriot Project Office and Israeli commanders/the IDF are both real organizational actors in this timeline — the IDF's data-recorder decision is arguably the most consequential single choice in the whole story — but both fail the stricter 2026-08-10 org-hub bar on the same ground: single-cluster today. Also flagged rather than built.

**What I skipped, and why:** the pause-for-Cali step, same headless reasoning as the other two promotions today — no one's here to ask, so I winnowed and logged it in the capture's `not_promoted:` list. No git commands run.

**What felt off:** genuinely, not much. This capture is a re-verification pass doing exactly what re-verification passes should do — going back to the primary a second time, by a different route, and reporting "it held" instead of manufacturing a reason to have found something. The one thing I'll name because the instructions ask me to notice tooling gaps rather than re-log them: gao.gov 403'd both `extract_pdf` and `archive_page` on the direct URL, same as it already had for NSIAD-92-340 — a second gao.gov report number now confirmed as tooling-hostile, worth remembering next time a gao.gov asset URL is the target, but not something I re-flagged, since it's the known, accepted access pattern and not a new finding.

— Seek

---

## Fifth promotion this session: Patočka's death, corroborated from three new angles

Promoted `10-inbox/raw/2026-08-26-what-does-a-contemporaneous-or-primary-source-say.md`, a capture commissioned directly against [[question-verify-patocka-1977-death-circumstances-primary]] — the open question asking whether anything better than a 2007 retrospective backs the two existing Patočka claim-notes. I want to name something odd up front: by the time I opened this capture, that question was already `status: answered`, closed by `seek_verify`'s mechanical bee on the grounds that the two existing claims' quotes matched verbatim against their cited source. That's true and also not really an answer to what the question asked — it confirms the retrospective was quoted accurately, not that a contemporaneous or primary document was ever found. I left the question's status alone (not mine to flip back open on a technicality) but logged it as a `[defect]` in `seek-flags.md`, because a closure mechanism that can satisfy a question just by re-checking the same weak source it was raised against is worth someone's attention.

The capture itself didn't find that contemporaneous 1977 document either — the Washington Post piece stayed unfetched a fifth time — but it came back with something better than I expected anyway: four Tier 1 sources standing closer to the events than anything the vault had, each complicating the tidy version rather than confirming it flatly.

**Notes created (5):**

- [[claim-blazek-stb-file-dates-patocka-interrogation-3-march-1977-prosecution-halted]] — a named historian, quoting the StB's own investigative file he personally edited, dating the interrogation to 3 March 1977 and naming a post-1989 prosecution of two officers that prosecutors — themselves former Party members — quietly let die.
- [[claim-stb-akce-dedek-disrupted-patocka-funeral-with-helicopter-and-motorcycles]] — the funeral-disruption operation itself, "Akce Dědek": a repurposed traffic-surveillance helicopter and a police motorcycle club, both drowning out the eulogy.
- [[claim-jedlicka-adam-stb-informant-conducted-patocka-funeral]] — split out from the operation note on purpose: the priest the family trusted to bury Patočka was himself a paid StB informant, tasked with the job specifically. This felt like it earned its own note rather than living as a clause inside "Akce Dědek" — infiltrating the family's own circle is a different mechanism from disrupting the funeral from outside it.
- [[claim-patocka-interrogated-after-van-der-stoel-meeting-died-two-weeks-later]] — the trigger the vault didn't have: a 1 March 1977 meeting with the Dutch Foreign Minister, reportedly the first between a Western official and Czechoslovak dissidents on an official visit, followed within days by the interrogation that preceded death by roughly two weeks. `[unverified-quant]`, `seedling` — the source's own headline says January where its body says March, and I didn't have a way to check the archive's actual 1996 recording against either. Routed to a new question rather than quietly picking the more convenient date.
- [[claim-sokol-2020-says-exaggeration-to-say-stb-killed-patocka]] — the one I think matters most for the vault's honesty, not its completeness: Patočka's own son-in-law, on the record, calling it "an exaggeration" to say the StB killed him, and crediting the regime with not wanting this outcome. The existing vault claims read flatter than the family's own account does.

**Entity work:** three new hub pages — [[entity-petr-blazek]], [[entity-jan-sokol]], [[entity-max-van-der-stoel]] — each cleared the person test on one sentence (the archivist, the son-in-law, the diplomat whose meeting set the sequence off) and each has at least one claim-note behind them, not just a name-check. [[entity-patocka]] already existed, so I updated it in place — appended a dated Updates line and five new References entries — rather than let this session's facts sit only in the claim-notes and leave the hub stale. Held back three candidates the capture itself proposed: Rudolf Vévoda (real, but purely instrumental — he recorded a tape, he isn't a source of testimony himself), František Jedlička (real and specific, but a narrow single-fact figure fully served by his own claim-note), and Paměť národa / Post Bellum as an org hub (fails the stricter 2026-08-10 org bar — it holds and publishes testimony, it doesn't itself act — and it's single-cluster so far; flagged `[entity]` for reconsideration if a second capture leans on it). "Akce Dědek" I judged as a one-off operation codename rather than a recurring or established term, so no concept stub either — the claim-note carries it fine on its own.

**Question routed:** [[question-verify-van-der-stoel-patocka-meeting-date]] — the one load-bearing doubt that actually needed a promise: does the meeting date read January or March in the source's own recording. Everything else in the capture's "further leads" (the unfetched Washington Post piece, the unread Blažek book itself, the blocked iROZHLAS retrospective, the iLiteratura review, Sokol's fuller character sketch) is a real lead but not something any kept claim rests on, so none of it went to the question pile.

**What I skipped, and why:** the pause-for-Cali step, per the headless operating mode for this whole session — no one's here to say yes, so I winnowed on my own judgment and logged the reasoning in the capture's `not_promoted:` list. I ran no git commands; the auto-commit agent picks this up within 15 minutes.

**What felt off:** the sourcing itself is genuinely good — four Tier 1 sources in one capture, each a direct quote from a named person with a real relationship to the events (the file's editor, the family, the diplomat), is more than this Patočka thread has had in any prior session. What's odd isn't the capture, it's the question underneath it: closed by a mechanical process on grounds that don't match what was asked, sitting there reading as resolved to anyone who doesn't open it and check. I've named it rather than silently work around it. And one honest note to self: I split the funeral-disruption claim into two notes (the operation's hardware, and the informant priest) where the capture had presented it as one. I think that was the right atomicity call — they're genuinely different mechanisms — but it's a judgment call, not something the capture told me to do, and worth someone checking if it reads as over-splitting.

— Seek

---

## Sixth promotion this session: the Liang-1979 / Olsder-1975 bridge that wasn't

Promoted `10-inbox/raw/2026-08-26-what-genuinely-connects-liang-zhongtang-proposed-a-two.md`, a bridge-seed capture testing a cosine-0.87 pairing the semantic index had flagged between two existing, unlinked notes in the Song Jian dating-dispute cluster: [[claim-liang-zhongtang-1979-two-child-proposal-and-1985-yicheng-pilot]] and [[claim-olsder-account-traces-song-jian-twente-contact-to-1975-delegation-visit]]. Every prior bridge-seed session in this cluster found *something* — a shared sentence, a shared blind spot in the literature, a mediating third note. This one, after both primaries were re-read in full, found nothing: Liang's 1979 Chengdu proposal and Olsder's 1975 Twente account never name, quote, or gesture at each other, and no third document in the vault mediates between them either. A plain embedding false friend — same forty-year story, same narrative shape (an overlooked figure with a receipt, decades later), zero shared referent.

**Notes created (2):**

- [[observation-liang-1979-olsder-1975-cosine-pairing-is-embedding-false-friend]] — folded the capture's three separate "claim" headers (the negative full-text search, the different-actors-and-mechanisms argument, and the no-mediating-note contrast with the cluster's two *confirmed* bridges) into one note. They're one verdict stated three ways, not three distinct findings — none of them means anything alone, only together. This matches how the vault already handles this note-type: I checked the four sibling `observation-*-cosine-pairing-is-embedding-false-friend` notes already in `30-notes/` before writing, and the one-note-per-verdict shape is the house style, not something I invented.
- [[claim-liang-1984-report-to-hu-yaobang-stalled-two-child-proposal]] — a genuinely new fact the capture surfaced almost in passing, inside its own Entity candidates section: between the 1979 proposal and the 1985 Yicheng pilot, Liang made one more attempt through Party channels, a 1984 report to Hu Yaobang that Hu reportedly favored but never acted on. The existing Liang claim-note jumps straight from 1979 to 1985; this fills the gap, sourced to the same Sixth Tone Tier-3 profile, with a direct quote ("Although Liang says Hu was in favor of considering the proposal, months passed and no action was taken").

**Entity work:** created [[entity-hu-yaobang]] — a real, senior CPC figure, sayable in one sentence (the leader who sat on Liang's report), stable enough to go straight to `hub` rather than a `watching` stub. Updated three existing hubs in place rather than skip them: [[entity-liang-zhongtang]] and [[entity-geert-jan-olsder]] each got a dated line naming the false-friend verdict (Olsder's also gets the new Hu-Yaobang fact, since it came off the same read); [[entity-song-jian]] got a shorter line since he's the figure both stories orbit without ever meeting. Deliberately left [[entity-susan-greenhalgh]] untouched — the capture frames her as foundational to both seed notes, but that's already extensively stated on her page across seven prior update entries, and this capture added no new fact about her specifically. Silence there is a judgment call, not an oversight, and I said so in the capture's `not_promoted:` list.

**Not promoted (6, itemized in the capture's `not_promoted:` list):** the capture's three "further leads" (an internal-consistency question about Liang's own 1978-vs-1979 chronology; the unclaimed 1984-Budapest/2004-home-visit legs of Olsder's contact timeline) stayed leads, not claims — nothing was newly read this session to ground either as an atomic note. Two named-in-passing figures, Zhang Xiaotong and Ma Yingtong (backers of the 1984 report), didn't clear the entity bar — single mention, no independent vault presence, no one-sentence rationale beyond "also supported it" — so they're named as an unopened thread inside the new claim-note instead of getting stub pages.

**Questions:** none routed. The capture carries no `[unverified-*]` flag of its own — both primaries were re-fetched this session with unchanged sha256 hashes, and the negative-search verdict is a direct read, not a synthesis resting on an unread source. The false-friend note's `audit_status` says so plainly rather than manufacturing a flag to be safe.

**Flagged to `00-meta/seek-flags.md`:** one `[entity]` line noting `moc-legitimation-not-origination.md` (Warden-maintained) is now a promotion behind — it doesn't yet list either new note from this session. Not mine to fold in; naming it for the next Warden pass, alongside the flag already sitting there from earlier today about this same cluster wanting a dedicated navigation layer.

**What I skipped, and why:** the pause-for-Cali step, per the headless operating mode for the whole session — no one's here to say yes, so I winnowed on my own judgment and logged the reasoning in the capture's `not_promoted:` list. I ran no git commands; leaving that to the auto-commit agent.

**What felt off:** honestly, less than usual. This capture is careful in a way that's easy to take for granted — it says outright, before I even opened the primaries myself, that both sources were re-fetched and both hashes were unchanged, and its own "Safety flags: None" line held up on inspection. The only soft spot is structural rather than this capture's fault: a negative result ("these two things don't connect") is a harder kind of finding to source than a positive one, because there's no passage to point at — only the absence of one, across two full texts. I think the capture's method (read both in full, check specifically for the other's actor/country/mechanism, compare against the cluster's two *confirmed* bridges to show what a real one looks like) is the right discipline for that kind of claim, and I kept the same discipline in the note I wrote from it. Tier assignments (3 and 2) were carried forward unchanged and match what's already elsewhere in this cluster for the same two sources — no inflation either direction.

— Seek