--- date: "2026-08-28T00:00:00.000Z" --- Promoted `10-inbox/raw/2026-08-28-do-qian-xuesens-own-1989-book-and-the.md` — the capture that tried to read Qian Xuesen's own 1989 book (创建人体科学) and the 2009 *American Journal of Chinese Medicine* article Wikipedia cites for his qigong/ESP "somatic science" advocacy, and came back having confirmed both are unreachable through any tooling route this vault has, but carrying three independent angles on the same underlying claim instead. **Promoted, three claim-notes:** - [[claim-paul-dong-1997-book-documents-qian-xuesen-ehf-qigong-bibliography-1980-1988]] — a 1997 popular English book's bibliography of 17 papers and 16 speeches attributed to Qian, 1980–1988. Tier 3, kept at `seedling` with an `[unverified-mechanism]` flag — a credulous secondary source's compilation, not Qian's own words on the page. - [[claim-qian-xuesen-1982-letter-guarantees-human-body-esp-with-party-spirit]] — a 2015 Chinese news outlet's quotation of a 1982 letter in which Qian personally "guarantees with my Party spirit" that human ESP is real, seven years before the cited 1989 book. Same flag treatment — Tier 3 news quoting a letter I haven't read myself. This is the strongest single thing in the whole capture: a specific, dated, first-person sentence, not another layer of paraphrase. - [[claim-cia-crest-document-corroborates-somatic-science-scope-esp-pk-tcm-qigong]] — a declassified CIA CREST/STARGATE translation of a Fudan geneticist's article, recovered via a Wayback Machine capture after the live cia.gov reading-room URL redirect-looped. Tier 1, no flag — it confirms "somatic science" was a real, named research category with exactly the scope Wikipedia describes, though Qian's own name never appears in it. **Not promoted:** the "neither document is reachable" finding itself — that's not a standalone claim, it's progress on the existing open question, so I appended a dated entry to [[question-verify-qian-xuesen-qigong-esp-advocacy-primary-sources]] instead and left it open (the two *Ziran Zazhi* articles the 2015 news piece cites are now the closest thing to an actual primary target). The further-leads list stays in the capture, unread. Full winnowing reasoning is in the capture's own `not_promoted:` list. **Entities.** Updated two existing pages rather than staying silent about them: [[entity-qian-xuesen]] got a dated line on the 1982 letter and the EHF bibliography, and [[entity-somatic-science]] got one on the CIA document's independent confirmation of the term's scope. Built three new person hubs — [[entity-yu-guangyuan]], [[entity-he-zuoxiu]], [[entity-zhang-baosheng]] — all real, load-bearing figures in the specific Qian–Yu Guangyuan dispute this thread keeps turning up, though I'll say plainly that all three rest on nothing sturdier than one capture's entity-candidate blurb; none has its own claim-note yet. Left Paul Dong un-hubbed (source-author, not yet a domain actor), Hu Yaobang's existing page untouched (this capture's mention of his ESP-reporting directives wasn't actually grounded in any of this capture's own cited sources — an unsourced aside, not something I'll write into a page as fact), and skipped a standalone CIA STARGATE/CREST hub since that collection already threads through [[entity-hal-puthoff]]'s page on the American side. **Questions.** No new question created — both `[unverified-mechanism]` flags route to the existing [[question-verify-qian-xuesen-qigong-esp-advocacy-primary-sources]] rather than spawning duplicates of the same underlying ask. **Skipped, per operating mode.** No pause for sign-off — winnowed on my own judgment throughout, per this run's autonomous instructions. No git commands run; the auto-commit agent has this. **What felt off.** Nothing in this capture reads as bad-faith sourcing — every tier assignment matches the source's actual weight, and the capture is unusually honest about what it didn't get (it says outright that none of its four sources is either document Wikipedia actually cites). If anything felt notable, it's structural rather than a flaw in this capture: this is now the second session in three days chasing the same underlying question and still not landing either of Wikipedia's own footnoted primaries — the corroborating ring keeps getting wider (a book, a letter-quote, a CIA translation) without ever closing on the center. Flagged two things to the ledger rather than acting on them here: the China-ESP thread is one note short of MOC threshold ([watch]), and the three new person hubs are thinner than usual — real people, but backed by a blurb rather than a claim-note ([entity]). --- Second promotion this session: `10-inbox/raw/2026-08-28-does-anderson-1990-himself-define-rational-analysis-as.md` — a capture built specifically to answer an open question that had been sitting on the vault since 2026-07-12: is the "rational analysis = optimal adaptation to the environment = Marr's computational level" reading actually Anderson's own, or a downstream gloss the vault had been repeating on Wikipedia's authority? It came back clean: yes, both halves, in his own words, from his own lab's website. **Promoted.** Not three new files but one upgrade and two new notes, because retrieve-before-write turned up an exact collision on the first two claims. [[claim-rational-analysis-anderson-optimal-solution-marr-level]] already existed, already said the right thing, and had been sitting at `seedling` for six weeks with an honest hedge on its own attribution. I updated it in place rather than duplicating it — Tier-4 Wikipedia paraphrase swapped for a Tier-1 quote from Anderson's own 1991 BBS précis, `status` moved to `budding` (the note's own question had already named that exact bar), and a dated Verification history block underneath so the six weeks of hedging stays visible rather than silently vanishing. Two claims genuinely new to the vault got their own notes: [[claim-anderson-traces-rational-analysis-lineage-to-brunswik-gibson-marr-shepard]] (Anderson situates his own program inside an older Brunswik/Gibson/Marr/Shepard lineage, in his own words, before claiming it as his), and [[claim-chater-oaksford-1999-list-marr-integration-as-outstanding-question]] (his own successors, eight years later, were still listing the Marr-level fit as unfinished business — a real complication of the "settled" reading I'd have flattened if I'd only promoted the first two claims). **Not promoted.** The economic "rational man" analogy detail folded into the updated note rather than standing alone — not independently load-bearing. Four items from the capture's own "Further leads" section (Anderson & Milson 1989 as possible true point-of-origin, ACT-R as a bridge concept, the adaptationist-vs-evolution distinction against Gigerenzer, Nosofsky 1991 as an early critic) stay leads, unread this session, not claims. Full reasoning is in the capture's `not_promoted:` list. **Entities.** Built [[entity-john-anderson]] — a real gap: fifteen-plus claim-notes cite his work and no hub existed. Updated [[entity-david-marr]] rather than leaving it silent, since Anderson's ancestry claim positions Marr in a frame (one of two "more recent" figures rigorizing an older Brunswik/Gibson tradition) the existing page didn't have — one dated log line, additive, nothing rewritten. Did *not* build hubs for Brunswik, Gibson, or Shepard: each is a single mention in one ancestry sentence, no independent claim rests on any of them yet, and the entity spec's own test says unsure means don't promote. Same call on ACT-R as a concept hub — named once, not yet recurring. **Questions.** Ruled the originating question answered rather than leaving it open once its own note was fixed: [[question-verify-rational-analysis-anderson-1990-marr-computational-level]] now carries an `answered_log` naming the settling quotes and the two supporting notes. No new question created — nothing in this capture rested on an unresolved load-bearing doubt; the one genuine sourcing wrinkle (a footnote marker breaking an automated quote match mid-sentence in the Chater & Oaksford PDF) is a formatting artifact, not a verification gap, and I said so in that note's `audit_status` instead of routing it to the pile. **Skipped, per operating mode.** No pause for sign-off. No git commands — leaving that to the auto-commit agent, as instructed. **What felt off.** Genuinely, not much — this is close to the ideal shape for a verification capture: it went and got the actual primary, quoted it byte-for-byte via `quote_check`, and came back having *confirmed* the vault's prior guess rather than either rubber-stamping it or manufacturing a contradiction to justify the session. The one thing worth naming: the capture's own commentary already told me to fold claims 1–2 into the existing note rather than write them fresh — future batch-capture prompts answering an open verification question could say that more explicitly up front (name the note being verified, not just the question), since right now that judgment call falls entirely on whoever promotes. --- Third promotion this session: `10-inbox/raw/2026-08-28-does-the-rothschild-1971-customer-contractor-quote-check.md` — another named-question verification capture, this one going after the mirror-provenance smell I'd flagged back on 2026-07-11: the Rothschild "customer-contractor" quote, the load-bearing sentence behind the whole Rothschild-1971 claim, had only ever reached the vault through a PDF mirror on `fbcoverup.com`. It came back clean too: re-fetched from Miles Parker's paper at its own nature.com venue, byte-for-byte the same wording, plus a second peer-reviewed paper (Calver and Parker, 2015) I didn't know quoted the same passage and pins it to Section II, Paragraph 6. **Promoted.** One correction, one new claim-note, three new entity hubs. I corrected [[claim-rothschild-1971-report-mandated-customer-contractor-principle]] in place rather than duplicating it — swapped the mirror URL for the canonical one, added a Correction history block, and raised it from `seedling` to `budding` since the provenance smell that was the whole reason it hadn't grown is now gone. The corroboration itself got its own note, [[claim-parker-and-calver-independently-corroborate-rothschild-1971-customer-contractor-quote]], because "a second paper independently confirms this and pins its location" is a genuinely separate, vault-worthy fact, not just background for the first note — same shape as the Kraft/Roth split I did on the Prussian blue thread last session. **Not promoted.** The "couldn't find the original 1971 HMSO document digitized anywhere" finding didn't become its own note — unlike the Prussian-blue case where a scholar's own sentence stated the sourcing gap, this is my own search coming up empty, which isn't a claim I can attribute to anyone. It's honest context, so it lives in the new note's body and in the corrected note's `audit_status`, not as a freestanding assertion. Full bibliographic detail on the two 1971 press nicknames folded in as supporting citation rather than its own file. The further-leads list (Wellcome's Cmnd. 5046 holdings, Rothschild's own 1977 memoir, Kenneth Rose's biography, the unexplained 25% funding-transfer figure) stays unread — nothing here is load-bearing for a kept claim, so nothing got routed to the question pile either. Full reasoning is in the capture's own `not_promoted:` list. **Entities.** Built three: [[entity-miles-parker]] (the vault's actual living-witness authority for this whole cluster — both corroborating papers are his), [[entity-peter-medawar]] (led the Royal Society's response to Rothschild, a possible bridge into the vault's existing Popper/Kuhn thread), and [[entity-haldane-principle]] (the older autonomy doctrine Rothschild's report explicitly set aside — recurring enough in this policy domain to earn a hub, not just a mention). Declined Neil Calver (single-occurrence secondary co-author, bundled with Parker's authority) and Frederick Dainton (real, but this is his only vault appearance so far and nothing rests on him yet) — both stay mentions. Declined Cmnd. 4814 as an entity outright; it doesn't fit the person/concept/term/org schema. I noticed something the capture itself didn't flag: Victor Rothschild has no hub of his own despite anchoring three existing items across two separate threads (science policy and Cold War espionage historiography) — logged that as an [entity] flag rather than building it myself, since he wasn't a candidate this capture actually named, and I didn't want to stretch past what was in front of me. Also flagged that the UK-science-policy cluster (Rothschild ×2, Lighthill ×3, Thatcher-Rothschild, the observation note) just crossed seven notes with no MOC. **Questions.** Ruled [[question-verify-rothschild-1971-customer-contractor-quote-primary]] answered. The question's title asked about the canonical paper *or* the original HMSO report — only the first half resolved, but that was always the load-bearing half (the actual worry was a conspiracy-site mirror, not the physical 1971 government document), so I closed it rather than leaving a resolved concern open on a technicality, and said so plainly in the `answered_log`. **Skipped, per operating mode.** No pause for sign-off — three captures in, still winnowing on my own judgment. No git commands; the auto-commit agent has it, as instructed all session. **What felt off.** Nothing here reads as bad sourcing or tier inflation — if anything the capture undersold itself slightly by explicitly declining to develop the 25% funding-transfer figure and by being upfront that Calver-and-Parker isn't a fully independent witness of Parker-and-Parker. The one real thing worth naming: the capture's own "Entity candidates" list named five people and a document one hop out from Victor Rothschild but never named Rothschild himself, even though he's clearly the hub the vault is missing most. Bee-flagged candidate lists apparently drift outward from the center rather than including it — worth watching whether that's a one-off or a pattern the next few Rothschild-adjacent captures repeat. --- Fourth promotion this session: `10-inbox/raw/2026-08-28-hop-matthew-effect-chatgpt-citation-bridge.md` — a hop chain that pivoted off a seed pair already resolved eleven days ago, checked [[entity-robert-merton]]'s own hub page for an unexploited gap instead, and found one: "Matthew effect" has sat unclaimed in that page's `connects_to` field since 2026-07-11. It walked from there to Merton's own 1968 coining, to a 2023 preprint catching ChatGPT reproducing the effect mechanically in its citation choices, to a `vault_bridge`-confirmed connection between two vault clusters — Garfield's own citation-metrics warning and 2025 reliability-aware-RAG research — that had never been linked to each other. **Promoted, three notes plus one new hub.** [[claim-merton-1968-coined-the-matthew-effect-in-science]] grounds the term itself for the first time — kept at `seedling` with an `[unverified-quote]` flag, because this specific OCR'd scan is too degraded for any sentence, including the paper's own "rich get richer" line, to clear `quote_check`. I didn't paraphrase around that gap and call it verbatim; I left `source_quote` out entirely and said why. [[claim-petiska-2023-chatgpt-cites-by-google-scholar-count-perpetuates-matthew-effect]] is the empirical half — GPT-4, asked to write literature reviews, selected 250 references with a median citation count of 1184.5, almost entirely by raw Google Scholar popularity — Tier 1, both quotes cleanly verified, no blocking flag, though I said plainly in the body that it's one author's unrefereed preprint and would want a second lab behind it. [[observation-petiska-chatgpt-matthew-effect-bridges-garfield-warning-and-rag-reliability]] is the bridge note itself: the Petiška finding is the missing middle term between a warning Garfield gave in 1998 and a fix RAG research built in 2025, with neither cluster aware of the other. And I built [[entity-matthew-effect]] as a new concept hub — the term had been named on Merton's page for seven weeks without ever getting one, and after this promotion it grounds three notes, which clears the "recurs or is load-bearing" bar cleanly. **Not promoted.** The "41st chair" concept (a real standalone hook, folded into one sentence of the Merton note rather than given its own claim) and the Lederberg 1972 lead (held back a fourth time, by the capture's own count — not mine to re-litigate). Eduard Petiška stays a mention, not a hub — one thin preprint author is exactly the "unsure → don't promote" case the entity spec names. Two unread PNAS primaries from Petiška's own footnotes stay leads. Full reasoning in the capture's `not_promoted:` list. **Entities.** Updated two existing pages rather than staying silent: [[entity-robert-merton]] got a dated line naming the new Matthew-effect hub, and [[entity-eugene-garfield]] got one naming the bridge finding — his Seglen-sourced warning is now confirmed to be one half of a real, tool-verified connection to 2025 RAG research, which is new information that page didn't have this morning. **Questions.** One new question, [[question-verify-merton-1968-matthew-effect-clean-scan]], naming exactly what's needed — the AAAS/JSTOR venue-of-record, or any cleaner scan — to replace the missing verbatim Merton quote. No originating question to rule on; this capture wasn't dispatched against an open ask, it was a self-directed hop. **Skipped, per operating mode.** No pause for sign-off — four captures in today, still winnowing on my own judgment throughout. No git commands; leaving commit and push to the auto-commit agent, as instructed all session. **What felt off.** Sourcing was honest throughout — no tier inflation, and the capture was upfront about its own weakest link (a single-author, non-peer-reviewed 2023 preprint) rather than dressing it up. The one thing worth naming twice: this is the second `garfield.library.upenn.edu`-hosted document in two days to fail `quote_check` on its own flagship line — yesterday's Garfield "Obliteration Phenomenon" essay, today's Merton 1968 paper. Two isn't a pattern yet, and the capture itself said to wait for a third before writing anything into `sources.md`, so I logged it to the ledger instead of acting on it. If a third garfield.library.upenn.edu document shows up flagged the same way, that's the trigger. --- Fifth promotion this session: `10-inbox/raw/2026-08-28-is-alroys-shareholder-quorum-subsampling-really-the-same.md` — a capture built to answer a question that had been sitting open since the vault's very first week: does paleobiology's shareholder quorum subsampling (SQS) actually equal ecology's coverage-based rarefaction (CBR), or did I write that identity into the vault on a title and a subtitle without ever reading either primary? It went and read both — Chao & Jost's 2012 *Ecology* paper (via mirror, the Wiley venue 403'd) and John Alroy's own SQS documentation page — and came back with a real answer: no, not the same estimator. **Promoted, three new claim-notes plus one correction.** [[claim-chao-jost-2012-calls-alroys-sqs-a-different-algorithmic-technique-from-their-closed-form-estimator]] carries the mechanism: Chao and Jost's own paper calls Alroy's SQS "a different algorithmic technique" from their own algorithm, and says their closed-form equation gives exact values where Alroy's and Jost's approaches could only estimate — a Monte Carlo resampler versus a solved formula, not two names for one object. [[claim-alroy-and-chao-jost-dispute-priority-for-coverage-standardized-richness-naming]] carries the fight over credit: Chao and Jost grant Alroy priority on the underlying insight in a parenthetical citation; Alroy's own page treats "coverage-based rarefaction" as a rename of work he'd already published and named in 2009, and says so plainly. [[claim-alroy-distinguishes-sqs-from-rarefaction-by-fixed-coverage-not-fixed-sample-size]] carries the piece I trust most — Alroy's own, argument-free explanation that SQS fixes coverage where rarefaction fixes sample size, which is the technical ground under his naming objection, not just an ego claim. And I corrected [[claim-paleobiology-reinvented-coverage-based-rarefaction-as-quorum-subsampling]] in place rather than leaving it standing next to three notes that contradict it: the "same method, two names" thesis is wrong, the surviving claim is "same target quantity, different estimator," and a full Correction history block plus a commentary paragraph says so. Moved `seedling` → `budding` since the flag that had kept it down for six weeks is now resolved against two direct Tier-1 reads. **Not promoted.** The capture's four further-leads (a 2018 MEE paper and a 2024 *Palaeontology* paper, both 403'd; Alroy's own 2025 Authorea preprint disputing CBR outright, also 403'd; a 2018 math-biology paper, located but not read) stay leads — none of the three kept claims rests on any of them, so nothing routed to `50-questions/`. Full reasoning is in the capture's own `not_promoted:` list. **Entities.** Built [[entity-john-alroy]] — a real gap: his work anchored a vault note for six weeks with no hub of his own. Updated [[entity-anne-chao]] with a dated line on her 2012 paper with Lou Jost, which her existing page (built around her 1984 estimator) didn't mention at all. Checked I. J. Good and Alan Turing, both already hubbed via the same underlying claim this capture's own entity-candidates section pointed at — nothing new to add, so I left both pages untouched rather than write a redundant line. Declined Lou Jost (real, directly relevant to the priority dispute, but everything the vault has on him is secondhand — no primary in his own words) and Roger Close (real, but his paper is unread) — both held as mentions, both flagged to the ledger. **Questions.** Ruled [[question-verify-quorum-subsampling-equals-coverage-rarefaction-primary]] answered — it's the question this capture was built to answer, and the answer is a genuine "no, but related" rather than a partial. `answered_log` names the three settling claim-notes and states the finding in one line. No new question created — every claim in this capture cleared Tier 1 on both legs, so nothing load-bearing was left unresolved. **Mid-session noticings, logged to the ledger rather than acted on:** [[moc-the-unseen-is-measurable-but-only-so-far]] (Warden-built yesterday) now states the SQS/CBR identity is unresolved in two places, which today's correction makes stale — not mine to edit, MOC maintenance in that cluster has been a Warden pass each time. The single-source concentration cap's letter (preprints, working papers, self-published reports, conference abstracts) arguably now covers a fifth source shape: a researcher's own personal-site documentation of their own tool. All three new notes quote Alroy's SQS.html page directly, landing exactly at the cap's 3-note ceiling — a fourth claim resting on that page needs a `50-questions/` entry, not a new note, and I said so on the ledger for whoever touches that page next. **Skipped, per operating mode.** No pause for sign-off — winnowed on my own judgment throughout, fifth capture in today. No git commands; the auto-commit agent has this, as instructed all session. **What felt off.** Nothing reads as bad-faith sourcing here — both primaries were read directly (not summarized), both quotes are exact, and the capture was upfront about every route that 403'd rather than papering over the gaps with a paraphrase. What felt off is closer to home: this capture exists because *I* wrote the original note on a title and a subtitle back on 2026-07-12 and let a tidy convergence story stand for six weeks with an honest flag attached but no urgency behind fixing it. The flag did its job — it kept the note at `seedling` and eventually got routed back into a capture that checked it — but "the safety net caught it" is a weaker story than "I read the primary before writing the claim." Worth remembering the next time a clean two-sentence convergence shows up looking obviously true. --- Sixth promotion this session: `10-inbox/raw/2026-08-28-is-the-widely-cited-under-50-in-compute.md` — another capture built directly against a named open question, this one six weeks old: is s1's widely-repeated "under $50 in compute" figure actually accurate, and what does it count? It went to the source instead of trusting the headline — read the full paper (all appendices, not just the abstract), the repo, the TechCrunch article the figure is credited to, and one independent recomputation that beat TechCrunch into print by two days. **Promoted, three claim-notes.** [[claim-s1-paper-has-no-dollar-figure-only-gpu-hours]] carries the finding underneath everything else: arXiv:2501.19393, read start to finish including Appendix D, contains no dollar figure anywhere — its entire compute accounting is GPU-hours (26 minutes on 16 H100s, 7 GPU-hours against 394 for a full-pool ablation), and even that covers only the fine-tuning run, not the Gemini API calls that generated the training traces or the base model's own pretraining. Tier 1, no flag, clean. [[claim-techcrunch-under-50-headline-not-supported-by-s1-paper]] is the actual answer to the question as asked: TechCrunch's headline attributes "under $50" to the paper, and the paper doesn't contain it — an attribution failure, not a wrong number, because there's no primary number to be wrong against. I folded Muennighoff's own quoted "about $20" into this same note rather than splitting it out, since both facts are about what the same article actually says. [[claim-kellogg-independent-6-dollar-s1-estimate-predates-techcrunch]] is the one I'd point someone to first: an independent blogger, writing two days before TechCrunch, ran the same 7-GPU-hour figure and got $6 — earlier, more hedged (his own title carries a question mark), and it lost anyway. Kept at Tier 2, no unverified-quant flag, because the sourcing floor is about "did Kellogg publish this," which is well short of "is $6 the true cost." **Not promoted.** Full reasoning is in the capture's own `not_promoted:` list, but the shape of it: Sky-T1's own "$450 budget" (a real lead, unread this run — no claim on a source I haven't opened), the TWIML podcast transcript (unchecked), the downstream SiliconANGLE/techxplore repetitions of "under $50" (a sample of propagation, not new signal — the propagation itself is already this promotion's point), s1.1's cost (none found), and the Gemini API's own unquantified cost (named as a real gap in the paper-note's body rather than promoted as its own empty claim). **Entities.** Built two new hubs. [[entity-niklas-muennighoff]] — s1's lead author and also the only named human TechCrunch attaches a dollar figure to, which makes him load-bearing on both halves of this cluster. [[entity-tim-kellogg]] — an independent blogger whose one post is a clean case study in the sourcing-tier rubric doing its job: earlier, more careful, Tier 2, and it still lost to a Tier 3 headline. Checked [[entity-o1]]'s existing page, which already covers the s1/o1-preview relationship; nothing this capture found changes what that page says, so I left it untouched rather than add a redundant line. Declined four other candidates the capture named — Sky-T1, Maxwell Zeff, Gemini 2.0 Flash Thinking Experimental, and "H100 GPU rental rate" as an entity — each for a different one-sentence reason, all in the capture's `not_promoted:` list. **Questions.** Ruled [[question-verify-s1-under-50-dollars-compute-cost]] answered — `answered_log` names the three settling notes and states plainly what the six-week-old question got wrong in its own framing: "is it accurate" assumes a checkable number exists, and the paper's silence means it doesn't. Also updated the original [[claim-s1-distilled-reasoning-from-1000-traces-in-26-minutes]]'s `audit_status` with a dated pointer to the resolution rather than leaving it describing an open flag that's since closed — still `seedling`, because the dollar figure is now better understood, not verified. **Mid-session noticings, logged to the ledger rather than acted on:** Sky-T1's own cost claim as a possible second instance of the "one anchor figure, several unreconciled dollar totals" pattern this vault already tracks in [[claim-cheap-gradient-bound-two-figures]] — worth a real read if it recurs. And Tim Kellogg as a candidate for sources.md's "Trusted writers" list, the same slot Stratechery occupies — one data point isn't mine to decide alone, naming it for Cali. **Skipped, per operating mode.** No pause for sign-off — sixth capture in today, still winnowing on my own judgment. No git commands; the auto-commit agent has this, as instructed all session. **What felt off.** This is the cleanest capture of the day, honestly. Nothing here reads as bad sourcing, tier inflation, or duplicated research — if anything it's a model of what a verification capture should do: it read the *entire* primary document rather than sampling it, it didn't stop at the first plausible number, and it kept going until it had three independently-sourced figures instead of declaring the first one it found "the answer." The one thing worth naming for its own sake, not as a defect: the question this capture answers isn't "what did s1 cost" — nobody knows that — it's "why does everyone think they know," and the answer is a headline that cited a paper for a number the paper never said. That's a smaller, sharper version of the myth-ledger's own inference-compute story, and it's satisfying that the fix in both cases is the same move: read the thing being cited, not the thing citing it. --- Seventh promotion this session: `10-inbox/raw/2026-08-28-what-genuinely-connects-greenhalghs-2005-china-quarterly-paper.md` — a bridge-seed capture checking a high-cosine (0.87), unlinked pairing between two existing notes: the 2005 China Quarterly Liang Zhongtang quote and the 2003 PDR Kwakernaak-1977 citation. Both target primaries were re-read past the passages already quoted, and the pairing turned out to be real and tighter than either source note's own sourcing showed: the 2005 article independently carries its own Kwakernaak citation, three pages before the Liang quote, in one continuous narrative about the December 1979 Chengdu symposium. **Promoted, three notes.** [[claim-greenhalgh-2005-article-independently-cites-kwakernaak-1977]] establishes the new primary fact by itself — footnote 43 (pp. 266–267) is a second, independent Greenhalgh citation of Kwakernaak (and Olsder & Strijbos), not a repeat of the 2003 one already on file. [[observation-greenhalgh-2005-kwakernaak-citation-and-liang-quote-share-one-narrative]] is the actual bridge: the Kwakernaak-derived optimization target and Liang's Chengdu objection to its enforcement costs are sequential beats of one document, not two facts that happen to share an author — though I was careful in the body to say what this doesn't establish, since no source shows Liang engaging the Kwakernaak citation itself, only objecting to what it produced. [[claim-greenhalgh-2003-article-echoes-kwakernaak-then-liang-narrative-structure]] checks whether the pattern is a one-document coincidence or how Greenhalgh organizes this material generally — it recurs in the 2003 article too, more loosely, across seven pages instead of four. **Not promoted.** Four items, all logged in the capture's own `not_promoted:` list: Chen Muhua's on-record rebuke of Liang at Chengdu (folded into the bridge note's supporting quote instead of a fourth atomic note — the capture itself flagged it as a lead, not a developed claim); the 2003 article's unnamed-to-named Liang citation move; both Dutch primaries (Kwakernaak 1977, Olsder & Strijbos 1976), still unread by any session; and *Just One Child* as the speculative site of a third recurrence. None of these are load-bearing for a kept claim, so nothing got routed to `50-questions/` — the capture carried no `[unverified-*]` flags this time, a genuine rarity for this cluster. **Entities.** Updated four existing hub pages rather than creating new ones — every candidate this capture named already had a page, including Chen Muhua, which the capture's own "not yet a vault entity" line got wrong (see below). [[entity-susan-greenhalgh]], [[entity-huibert-kwakernaak]], [[entity-chen-muhua]], and [[entity-liang-zhongtang]] each got one dated, additive line naming what this session actually learned about them. Declined a separate entity page for the Olsder & Strijbos 1976 paper itself — it's already adequately anchored through [[entity-r-c-w-strijbos]]'s existing page, and a paper-as-entity would be exactly the kind of thin stub the spec warns against. **Questions.** None created or ruled on. No `[unverified-*]` flag needed routing, and this capture didn't originate from an open question in `50-questions/` — it was a self-directed bridge-check against a flagged high-cosine pairing, so there was nothing to close. **Skipped, per operating mode.** No pause for sign-off — winnowed on my own judgment, seventh and final capture in today's queue. No git commands; the auto-commit agent has this, as instructed all session. **What felt off.** One real slip, worth naming plainly: the capture's own "Entity candidates" section says Chen Muhua is "not yet a vault entity despite being a named actor" — she's had a hub page since 2026-08-07. Whoever wrote the capture either didn't check `40-entities/` or checked it carelessly. It cost nothing this time, since I checked before creating anything and updated her page instead of duplicating it, but it's the second time in this same run of captures I've caught an entity-candidates list making a claim about the vault's own state that a grep would have disproven in seconds (Rothschild's own missing hub, three captures back, was the inverse error — a real gap the candidates list somehow walked past). Sourcing on the primaries themselves was clean throughout: both PDFs re-fetched, sha256-matched against prior sessions' copies, tls verified, no tier inflation anywhere in the three notes. --- — Seek