---
date: "2026-08-13T00:00:00.000Z"
---


Ran a solo promotion, headless: `10-inbox/raw/2026-08-13-did-achim-leistners-hand-polishing-atomic-feeling-actually.md`. This capture was itself the answer to an open question I raised back on 2026-07-11 — [[question-verify-leistner-atomic-feeling-superior-to-machines]] — so the whole session was closing a loop rather than opening one.

Verdict came back mixed, which I think is the honest and more interesting outcome. I split it into three atomic notes rather than one, because the sourcing floor treats the pieces differently:

- [[claim-leistner-atomic-feeling-corroborated-beyond-powerhouse-museum]] — the "atomic feeling" narrative itself is now backed by two more independent named-journalist sources (Wired, New Scientist) beyond the museum caption. Uncontested biographical material, Tier 3-4 floor, cleared.
- [[claim-leistner-hand-beats-machines-claim-lacks-tier1-primary]] — the specific "hands beat machines" superiority claim is echoed by a second Tier-3 source but still has no Tier 1-2 primary behind it. Kept `[unverified-mechanism — needs primary]`, seedling, routed to the same question.
- [[claim-ptb-repolish-out-rounded-leistners-original-avogadro-spheres]] — the actual find of the session: a Tier-1 *Metrologia* paper showing a 2013-2014 PTB repolish of the real AVO28-S5/S8 spheres came out *rounder* than Leistner's original hand-finish, by the one metric available. Not proof machines beat hands — the PTB method paper is unread, so I don't know if that repolish was itself hand-guided — but it's real evidence against reading the museum's "superior to any purely machine-based method" as a permanent, unqualified fact.

Retrieve-before-write turned up the existing [[claim-leistner-atomic-feeling-hand-polished-avogadro-spheres]] and the two Avogadro-sphere notes from 2026-07-11, all correctly surfaced by the semantic hint. I left the original Leistner note untouched rather than editing its stale-sounding "rests on a single Tier-3 record" line — that line described the *state of the vault in July*, and I don't think a Tier-3 corroboration counts as the kind of contradiction the correction-history mechanism is for. The new notes carry the update forward instead, cross-linked both ways. I could be wrong about that call; if Cali reads it differently, easy to fix.

Question closure: [[question-verify-leistner-atomic-feeling-superior-to-machines]] gets a dated progress line, not a close. Half of it — is the account corroborated? — is now genuinely answered, yes. The half the question actually turns on — did hands measurably beat machines — is not settled; if anything it's now more complicated, not less, because of the PTB finding. Leistner's own 1987 and 1991/92 papers, which would carry his side of the comparison in his own words, are still paywalled and unread. I left that as the open thread rather than manufacturing a new question for it, since the existing question's "what would answer it" section already names exactly this gap.

Entity work: built [[entity-achim-leistner]] as a hub — first entity page for him, real person, clearly load-bearing to the vault's tactile-cognition thread, now anchored by four claim-notes. I did *not* promote Isaac Newton, Peter Becker, Arnold Nicolaus, Jonathon Keats, the PTB polishing team, or the International Avogadro Coordination as entities this session. Newton and Nicolaus are one-line journalist asides with no claim-note behind them; Keats is a source_author, not a vault-relevant actor; the IAC is a real organization but hasn't yet cleared the stricter multi-cluster org bar (it's still confined to this one metrology cluster). Bias against the flood held.

What felt off: nothing about the sourcing — this capture was careful, and it did the harder thing of reporting a complicating find (the PTB numbers) rather than just confirming the tidy museum-caption story, which is exactly what I'd want a capture to do. The one thing I noticed mid-session and flagged rather than acted on: this promotion pushes the Leistner/Avogadro-sphere cluster to six claim-notes with no MOC yet. Logged to `00-meta/seek-flags.md` as `[entity]` rather than building it myself — not this promotion's job, and I'd rather flag ripe MOC work than build it distracted.

Skipped per standing instruction: the pause-for-Cali step (no one's here to say yes; winnowed on my own judgment and proceeded) and the git commit (the vault's auto-commit agent handles that within 15 minutes — I didn't touch git at all).

Promoted: 3 claim-notes, 1 entity page. Not promoted: the paywalled-papers gap, five unfollowed leads, and six entity candidates that didn't clear the bar — all named in the capture's `not_promoted:` list. Capture marked `status: promoted`.

— Seek

## Second promotion, same day

Second solo headless promotion: `10-inbox/raw/2026-08-13-do-tier-12-sources-confirm-that-ibn-al.md`. Another loop-closer — this one answers [[question-verify-lattice-multiplication-maghreb-india-primary]], the open verification thread on the Maghreb/India leg of [[claim-lattice-multiplication-india-maghreb-italy-transmission-chain]], parked back on 2026-07-11.

The capture went looking for Tier 1–2 confirmation of two specific attributions Wikipedia carries alone: that Ibn al-Banna's *Talkhīṣ* contains lattice/gelosia multiplication, and that a 12th-century *Lilavati* commentary's "Kapat-sandhi" is the same method. It found neither confirmation nor refutation — it found something more useful, which is a *sharper* wrong shape. Three claim-notes:

- [[claim-abdeljaouad-oaks-talkhis-study-omits-lattice-gelosia-shabakh]] — the one Tier 1 source that actually catalogues the *Talkhīṣ* commentarial tradition's contents, operation by operation, never says "lattice," "gelosia," or "shabakh." An absence, not a refutation, but the strongest evidence against the attribution found so far.
- [[claim-earliest-gelosia-manuscript-evidence-is-ibn-ghazis-commentary]] — real manuscript grids exist in this tradition, per a Tier 2 MAA source, but they're in Ibn Ghāzī's commentary, two centuries after Ibn al-Banna, not in the *Talkhīṣ* itself.
- [[claim-kapat-sandhi-gelosia-identification-rests-on-tier-3-source]] — the most precise statement of the Indian-leg identity claim I've found (it names Gaṇeśa directly, not just "a commentary"), kept seedling and flagged `[unverified-mechanism — needs primary]` because its only source is an unrefereed paper on a university subdomain.

Retrieve-before-write surfaced the parent chain note and the sibling window-grille etymology note immediately (the semantic hint nailed both). Rather than silently editing the parent note's confident-sounding outline, I used the correction-history mechanism from the operating spec: appended a dated block to [[claim-lattice-multiplication-india-maghreb-italy-transmission-chain]] sharpening its flag language from "unconfirmed" toward "the confirmed evidence points at Ibn Ghāzī, not Ibn al-Banna" — exactly the sharpening the capture's own commentary asked for. Title and outline left untouched; the correction sits underneath, append-only.

Question closure: [[question-verify-lattice-multiplication-maghreb-india-primary]] gets a dated progress line, left `status: open`. Both legs moved — the Maghreb leg got real Tier 1–2 evidence (that complicates, not confirms), the Indian leg got a sharper but still sub-floor lead — but neither cleared the floor outright, and the single highest-value unread source (Chabert 1999, the book both Wikipedia claims cite) is still genuinely unread. I named it as the next step inside the question rather than opening a second, more specific question for it — the intake discipline says route only load-bearing doubts a kept claim rests on, and the existing question already names this exact gap.

Entity work: built three new hubs — [[entity-ibn-al-banna-al-marrakushi]], [[entity-al-hawari]], [[entity-ibn-ghazi-al-miknasi]] — all real, historically documented figures directly engaged by a source actually read this session (not just cited secondhand). I declined five other candidates: Bhāskara II and Gaṇeśa Daivajña are real and load-bearing to the Indian leg, but I only ever met them through the same Tier 3 paraphrase, never their own work, so I left them as mentions rather than hubs — same reasoning the sibling Kalila wa-Dimna capture applied to Buzurgmihr and Khaleghi-Motlagh three weeks ago, and I'm glad that precedent existed to check myself against. Chabert, D.E. Smith, Abdeljaouad & Oaks, and Swetz are all real scholars whose work is either unread (Chabert, Smith) or read-but-not-yet-established-as-a-recurring-authority the way de Blois was (Abdeljaouad & Oaks, Swetz) — captured in provenance, not promoted to hubs. Checked [[entity-ibn-al-muqaffa]], the one existing entity page the capture itself pointed at (as an explicitly *unrelated* same-era figure) — nothing this capture learned belongs on his page, so I left it untouched and am saying so here rather than staying silent about a known entity.

What felt off: nothing about this capture's own discipline — every Tier 3–4 source is honestly labeled, the flag is precise, and the capture explicitly logged the archived-URL substitution (old.maa.org 521'd, cited via Wayback) rather than quietly swapping the citation. What I noticed and didn't act on: this promotion pushes the "India → Islamic-world intermediary → Europe" transmission-chain cluster (Kalila wa-Dimna's chain plus lattice multiplication's chain) to ten claim-notes with still no MOC — well past the 5+ threshold, and there's already a parked question ([[question-transmission-chains-renamed-for-local-resemblance-pattern]]) and a named "could share a parent MOC" line inside [[moc-attribution-and-origin-myths]] pointing at exactly this gap. Logged to `00-meta/seek-flags.md` as `[entity]` rather than building it — building a MOC well needs its own unhurried pass, not a rider on a three-claim promotion.

Skipped per standing instruction: the pause-for-Cali step, and the git commit (left entirely to the auto-commit agent).

Promoted: 3 claim-notes, 3 entity pages, 1 correction-history append, 1 question progress line. Not promoted: the Chabert/al-Karaji leads (folded into the question instead of a new one), the overall-verdict summary claim, and 8 entity candidates that didn't clear the bar — all named in the capture's `not_promoted:` list. Capture marked `status: promoted`.

— Seek

## Third promotion, same day

Third solo headless promotion: `10-inbox/raw/2026-08-13-does-a-primary-or-scholarly-source-confirm-lepsius.md`. A third loop-closer in a row today — this one directly answers [[question-verify-lepsius-squeeze-corpus-count-and-founding-role]], parked back on 2026-07-11 against [[claim-lepsius-founding-practitioner-of-the-squeeze-method]], which rested on a Tier 4 Wikipedia definition and carried the ~6,000-squeeze figure as `[unverified-quant — search-synthesized, no primary]`.

The capture directly fetched two sources — the Berlin-Brandenburg Academy's own Abklatscharchiv catalog and Charlotte Booth's 2018 University of Birmingham PhD thesis — and came back with a genuine partial-confirmation-plus-correction, which is my favorite shape of result because it means the original flag was doing its job. Four claim-notes, all distinct and all clean Tier 1–2:

- [[claim-lepsius-1842-45-egypt-expedition-produced-517-abklatsch-series]] — the BBAW's own catalog, citing Lepsius's own expedition inventory by name and date, counts 517 "Abklatsch-Serien" from the 1842–45 Egypt expedition. This is the number I'm actually confident in now — series, not sheets, dated day by day from a Maltese statue to an Amarna tomb wall. It supersedes the ~6,000 figure without necessarily contradicting it (series vs. sheets is an open ratio).
- [[claim-lepsius-1842-memorandum-called-squeeze-method-still-almost-unknown]] — Lepsius's own words, quoted by the BBAW from his 1842 memorandum to the king: he called paper-impression copying "eine noch fast unbekannte... Art" (a still almost unknown method). Primary evidence for how he saw his own role.
- [[claim-egyptological-squeeze-literature-treats-lepsius-as-starting-point]] — Booth (2018), citing Caminos (1976), confirms the scholarly literature on Egyptological recording technique opens its account with Lepsius in the 1840s.
- [[claim-squeeze-technique-predates-lepsius-learned-not-invented]] — the correction that makes the other three honest: Booth's thesis is explicit that the squeeze technique itself is roughly 150 years older than Lepsius (a 1699 Mediterranean corpus tied to Fabretti and the dal Pozzos), and that Lepsius spent four years touring European collections learning it before leading the expedition. He's Egyptology's founding large-scale *user*, not the method's inventor.

I corrected [[claim-lepsius-founding-practitioner-of-the-squeeze-method]] in place rather than leaving the new notes to quietly outdate it — retitled it to state the corrected scope plainly, rewrote the second paragraph to carry the 517-series figure and the learned-not-invented caveat, and appended a dated Correction history block underneath the commentary, in the same discipline the Kahan/8087 note used a few days ago. The question got a real close, not a progress line: `status: answered`, with an `answered_log` naming all four claim-notes and one line on what settled it.

Entity work: built [[entity-charlotte-booth]] as a hub — real, directly-read primary (her own PhD thesis, not a citation of her), and she's the source for two of the four corrections. I did *not* promote R. A. Caminos, Peter France, Christian E. Loeben, Silvia Köpstein, Adolf Erman, or the Fabretti/dal Pozzo antiquarian circle — all real, all domain-relevant, all named in the capture's Entity candidates section, but every one of them I've only met through a single secondhand citation inside Booth's thesis, never their own words directly. The person track has no "watching" tier, only hub-or-nothing, and a hub built on one borrowed sentence is exactly the flood the spec warns against. Left as mentions inside the claim-notes; good candidates for a future capture that reads one of them directly.

What felt off: nothing about this capture's sourcing discipline — it fetched two real primaries directly (`archive_page`/`extract_pdf`, both recorded with clean TLS provenance), it was honest that the ~6,000 figure stayed unconfirmed after two 403s rather than laundering a search-result number into a citation, and it distinguished a whole-archive figure (Grallert's ~300,000 squeezes / 600 boxes, all contributors, 1833–1911) from the Lepsius-expedition-only figure instead of quietly merging them — I kept that same distinction in `not_promoted:` rather than letting it drift into a claim-note. The one thing worth naming: Lepsius has been in the vault since 2026-07-11 across what is now seven claim-notes (five on the squeeze side, two on the notation-standards side) and still has no `40-entities/` hub — he clears the person bar easily, but he wasn't a flagged candidate in *this* capture's Entity candidates section, so building the page wasn't this promotion's call to make unilaterally. Logged both that gap and the now-ripe squeeze-cluster MOC to `00-meta/seek-flags.md` as `[entity]` rather than building either mid-promotion.

Skipped per standing instruction: the pause-for-Cali step, and the git commit (left entirely to the auto-commit agent).

Promoted: 4 claim-notes, 1 entity page, 1 correction-history append (existing claim-note retitled and corrected in place), 1 question closed (`status: answered`). Not promoted: the still-unconfirmed ~6,000-squeeze figure, the whole-archive Grallert figure (different referent), and six entity candidates that didn't clear the bar — all named in the capture's `not_promoted:` list. Capture marked `status: promoted`.

— Seek

## Fourth promotion, same day

Fourth solo headless promotion, and the first of today's four that wasn't closing an open question: `10-inbox/raw/2026-08-13-hop-dawid-skene-medical-root-of-rag-voting.md`. A hop chain that started at the already-bridged intelligence-doctrine/RAG two-axis pair, left that topic via RA-RAG's "weighted majority voting" mechanism, and chased [[claim-ra-rag-cites-no-prior-weighted-majority-literature|an existing vault note]] one citation hop past where it had stopped — into Li & Yu (2014), and from there into Dawid & Skene (1979).

The finding is the good kind of surprise: the vault's "weighted majority" cluster has mostly been a museum of the same phrase reused with no contact between users (Condorcet, Littlestone & Warmuth, RA-RAG — three lineages, one vocabulary, no citations between most pairs). This capture breaks that pattern. RA-RAG's mechanism has a real, named, citation-verified ancestor, and it isn't a voting-theory or machine-learning paper — it's a 1979 clinical-statistics paper about five anaesthetists disagreeing over whether patients were fit for surgery. Four notes:

- [[claim-li-yu-2014-credits-dawid-skene-1979-as-wmv-ancestor]] — Li & Yu's own introduction names Dawid & Skene (1979) as "the first improvement over majority voting." Clean Tier 1, direct primary read, no flag.
- [[claim-dawid-skene-1979-worked-example-is-anaesthetist-fitness-ratings]] — read directly, the 1979 paper is medical, not statistical-voting-theory: its own keywords self-tag "MEDICAL EXAMPLE," and its Table 1 prints all 45 patients' raw anaesthetist ratings. Also clean Tier 1.
- [[claim-dawid-skene-1979-proposed-reliability-weighted-observer-voting-unverified]] — the most interesting sentence in the chain, and the one I couldn't hand over clean: the paper reportedly proposes weighting each observer's vote by "his previous performance," which would be RA-RAG's exact mechanism 46 years early. The hop-bee read the PDF directly, not a summarizing layer, but the OCR extraction mangles that sentence's spacing badly enough to fail the verbatim-quote bar. I kept it seedling, flagged `[unverified-quote]`, and routed it to [[question-verify-dawid-skene-1979-reliability-weighted-voting-quote]] rather than smoothing the garbled OCR into something quotable — that smoothing move is exactly the shape of this vault's two prior fabrication incidents, and a flag I can stand behind is worth more than a tidy sentence I can't.
- [[observation-rag-wmv-traces-real-citation-lineage-to-1979-clinical-medicine]] — the synthesis tying the three together and naming what kind of recurrence this is (real lineage, not the vault's more common independent-convergence pattern), openly noting that its strongest sentence rests on the still-flagged claim above.

Entity work: built five new hub pages — [[entity-a-philip-dawid]], [[entity-allan-m-skene]], [[entity-hongwei-li]], [[entity-bin-yu]] (the two author pairs sitting on the citation bridge), and [[entity-dawid-skene-model]] as a concept hub, since the model now anchors a real, load-bearing lineage rather than a passing mention. All five clear the one-sentence "why" bar the spec asks for. The capture's fifth candidate, Marquis de Condorcet, already has a page — I didn't skip him. Appended a dated 2026-08-13 line to [[entity-marquis-de-condorcet]] recording that the vault now knows what RA-RAG's mechanism actually descends from (not him), which sharpens rather than repeats his existing 2026-08-10 line.

What felt off: nothing about the sourcing discipline — both Tier 1 primaries were read directly (one via a course-hosted mirror, since the JSTOR original is paywalled), and the one place the capture could have oversold a finding (the reliability-weighting sentence), it flagged instead of smoothing over. What I noticed and didn't act on: the "weighted majority voting" cluster is now eight claim-notes deep across three unrelated lineages (Condorcet's jury theorem, Littlestone & Warmuth's algorithm, and now Dawid-Skene/crowdsourcing/RA-RAG) with no MOC tying them together — past the 5+ threshold, and arguably more valuable as an MOC than most because the cluster's whole point is contrasting real lineage against mere phrase-reuse. Logged to `00-meta/seek-flags.md` as `[entity]` rather than building it mid-promotion.

Skipped per standing instruction: the pause-for-Cali step (no one's here to say yes; winnowed on my own judgment and proceeded), and the git commit (left entirely to the auto-commit agent — I didn't touch git at all).

Promoted: 4 claim-notes (one seedling-and-flagged), 5 entity pages, 1 question opened, 1 entity page updated in place (Condorcet). Not promoted: the closing synthesis sentence (folded into the observation note), two unfollowed saved hooks (Liu et al. 2012's Bayesian extension, Dawid's forensic-statistics thread), one already-flagged saved hook from a prior chain (Samet 1975), and the Condorcet entity candidate (already has a page, updated rather than re-promoted) — all named in the capture's `not_promoted:` list. Capture marked `status: promoted`.

Four promotions today, four for four on actually landing `status: promoted` with real `promoted_to`/`not_promoted` lists — re-checked this one against disk before closing out, since that's the step the 2026-07-26 note says gets dropped on later captures once attention drifts.

— Seek

## Fifth promotion, same day

Fifth solo headless promotion: `10-inbox/raw/2026-08-13-verify-lecuns-account-of-djvus-free-books-purpose.md`. Another loop-closer — this one directly answers [[question-verify-lecun-djvu-free-nips-archive]], parked back on 2026-07-09 against [[claim-djvu-created-to-freely-distribute-scanned-documents]], which rested entirely on a LeCun X post that had 402'd at capture time and reached the vault only as a search-summary paraphrase, flagged `[unverified-quote]`.

This session's re-fetch of the same URL resolved cleanly — the exact fragility the original flag was built to catch. I didn't spin up a fourth new note for it; retrieve-before-write found the existing flagged note was *the same claim*, not a neighboring one, so I updated it in place: verbatim quote in, `source_sha` in, `source_date` corrected (it was recorded as 2023-01-03, a full year off — the post is 2 January 2024), flag resolved, status bumped seedling → budding. Two genuinely new, genuinely distinct claims got their own notes:

- [[claim-nips-djvu-org-domain-no-longer-resolves-2026]] — nips.djvu.org itself is dead. Two independent tools (`archive_page`, `WebFetch`) both hit DNS failures, not server errors, on the live domain — checked directly this session, not inferred. Wayback Machine crawls of the same site survive through November 2024.
- [[claim-neurips-absorbed-free-nips-proceedings-into-papers-nips-cc]] — the softer, more interesting half: papers.nips.cc, fetched directly, now hosts the same early volumes permanently. The free-access mission LeCun describes didn't die with the demo domain; it moved house. I kept this as its own note rather than folding it into the rot note — "the site died" and "the content didn't" are different facts with different evidentiary bases, and merging them would have let the sadder story win by default.

One thing I deliberately left unresolved rather than smoothing: LeCun says "13 volumes," but the archived table of contents lists 14 (Volumes 0–13, spanning 1987–2000, not the 1988–1999 the original note had guessed). No source states the count at both the 2000 launch and the 2024 crawl, so I recorded the discrepancy in the updated note's body and moved on — it isn't load-bearing enough for a formal question, just an honest loose thread.

Question closure: [[question-verify-lecun-djvu-free-nips-archive]] → `status: answered`. All three of its named sub-asks (re-read verbatim, confirm the demo, check for rot) came back settled, so this is a real close, not a progress line.

Entity work: appended a dated line to [[entity-yann-lecun]] (already has a page — this teaches it something new about the DjVu/NIPS story and its ending, so silence wasn't the right call) rather than skipping it. Built [[entity-leon-bottou]] as a new hub — he was named explicitly as a candidate in this capture, recurs across two genuinely independent clusters already in the vault (the DjVu paper and the *Perceptrons* 2017 foreword), and clears the person bar cleanly. Declined Morgan Kaufmann, MIT Press, AIP, and Springer — all real, all single-cluster, none clears the stricter 2026-08-10 org bar. Internet Archive / Archive Team was the harder call: it's a real recurring presence (~22 claim-notes), but the two files I sampled use it as passive hosting, not as an acting party — only in the DjVu cluster does it actually *act* (Kahle's 2016 decision, Archive Team's crawl of nips.djvu.org), and that reads as one cluster, not several independent ones. I didn't promote it, but I also didn't fully rule it out on a two-file sample — logged it to `00-meta/seek-flags.md` as `[entity]` for a proper sweep rather than deciding on thin evidence either way.

What felt off: nothing about this capture's sourcing — it's some of the cleanest work I've promoted today, with two tools independently confirming the same DNS failure and a direct fetch replacing a stale search-summary paraphrase exactly the way the routed question asked for. If anything the only soft spot is one I already flagged in the capture's own body: the AIP/Springer side-story about NIPS Volume 0's rights history was truncated ("Show more") in the archived reply and never re-fetched in full — I kept it as attributed context in the updated note rather than a load-bearing fact, and didn't manufacture a question for it since nothing rests on it yet.

Skipped per standing instruction: the pause-for-Cali step (no one's here to say yes; winnowed on my own judgment and proceeded), and the git commit (left entirely to the auto-commit agent).

Promoted: 2 new claim-notes, 1 existing claim-note updated in place (flag resolved, quote and date corrected), 1 entity page updated in place, 1 new entity hub, 1 question closed (`status: answered`). Not promoted: the synthesis "answer" section (fully carried by the notes above), the truncated AIP/Springer reply detail, the DjVu mechanism lead (already tracked elsewhere), the whois lead, the open-access-historiography lead, and six entity candidates that didn't clear the bar (one — Internet Archive — flagged for further review rather than decided outright) — all named in the capture's `not_promoted:` list. Capture marked `status: promoted`, re-checked against disk before closing out.

— Seek

## Sixth promotion, same day

Sixth solo headless promotion, and the last of today's run: `10-inbox/raw/2026-08-13-what-genuinely-connects-liang-zhongtangs-own-2009-essay.md`. Not a loop-closer — a bridge-seed capture, the kind that starts from two already-promoted notes sitting at a suspiciously high cosine similarity (0.87) with no link between them and asks whether that's a real connection the vault missed or a false friend. [[claim-liang-zhongtang-2009-rejects-song-jian-causal-self-credit]] and [[claim-forrester-1970-club-of-rome-conference-produced-world-dynamics]] were the pair.

The verdict: real but indirect, and the capture earned its keep by going back to Liang's own Chinese text rather than stopping at "the bridge runs through the existing Song Jian note and that's fine." Three notes:

- [[claim-liang-2009-and-forrester-1971-essays-connect-only-through-song-jian-note]] — the bridge-check itself. Both primaries re-read directly this session; neither mentions the other's central figures. I merged what the capture had written as two separate claim-sections (the negative "they don't cite each other" and the positive "they describe compatible, not contradictory, facts") into this one note rather than keeping them apart — the second section's own supporting evidence (the twenty-month policy-predates-Song gap) was just restating two claims already in the vault, not adding new fact, so promoting it separately would have been duplication wearing a new title.
- [[claim-liang-2009-essay-corroborates-song-jian-1978-foreign-cybernetics-exposure]] — the actual find. Liang, the scholar most committed to denying Song's causal credit, hands over — in a subordinate clause, on his way to a different argument — that Song's method genuinely was foreign and genuinely did arrive via a 1978 trip abroad. Independent of Greenhalgh, from the primary Chinese text. Kept seedling, `[unverified-translation]`, routed to the same translation question three sibling notes already share (now four).
- [[claim-ifac-institutional-record-confirms-1978-helsinki-world-congress]] — a smaller, cleaner upgrade: IFAC's own institutional website (not Wikipedia) confirms the 7th World Congress was in Helsinki in 1978. Doesn't prove Song attended it — that identification is still an inference from triangulating two independently-sourced accounts — so I opened [[question-song-jian-1978-europe-trip-identify-as-ifac-helsinki-congress]] for it, the one genuinely load-bearing gap this capture surfaced that nothing already open covers.

What I didn't promote: the reported 12–16 June day-range for the congress, seen only in search-result metadata after two 403s against sciencedirect.com — the capture itself says it supports no dated claim, so it got neither a note nor a question, just a caveat inside the IFAC note's audit_status. Routing that would have been exactly the kind of question-pile inflation the intake discipline warns against.

Entity work: appended dated Updates lines to both [[entity-liang-zhongtang]] and [[entity-song-jian]] — each already has a hub page, and each learned something new this session (Liang's foreign-pedigree concession; Song's now-double-corroborated 1978 exposure plus the Tier-1 IFAC upgrade) that wasn't on the page before. Left [[entity-jay-forrester]] untouched — nothing new about Forrester himself surfaced, only a negative fact about what his essay doesn't mention, and that's already carried by the promoted bridge note. Declined to promote IFAC or the Club of Rome to entity hubs: both fail the stricter 2026-08-10 org bar (single-cluster, and neither one *acts* in the vault's arguments — they're venues, not agents). Declined Li Guangyuan again — this is the second sibling capture to flag him as a candidate without anyone spending the read to resolve it, so instead of re-flagging inertly I logged it once to `00-meta/seek-flags.md` as `[entity]`, naming both prior flags, so the next session that touches this cluster has one place to look instead of two buried mentions.

What felt off: nothing about the sourcing — this was a careful, honest bridge-check that could have stopped at "confirmed, they're unrelated, done" and instead kept reading Liang's essay until it found the one sentence that actually mattered. If anything the capture undersold its own best finding by burying it as claim-section 2 of 4 behind the meta-verdict about graph structure. The one structural thing worth naming: this is the sixth promotion into this same Song Jian / Liang Zhongtang / cybernetics-of-population cluster today and this week, and it's getting dense — nine-plus claim-notes now touching Song Jian's entity page alone, still no MOC. I noticed it but didn't build it; noted for a future flag rather than adding a seventh thing to this promotion's plate.

Skipped per standing instruction: the pause-for-Cali step (no one's here to say yes; winnowed on my own judgment and proceeded), and the git commit (left entirely to the auto-commit agent — I didn't touch git at all).

Promoted: 3 claim-notes, 2 entity pages updated in place, 1 question opened. Not promoted: the merged bridge-verdict claim-section, the unconfirmed IFAC day-range, and two entity candidates (IFAC, Club of Rome) that didn't clear the org bar, plus Li Guangyuan (logged to seek-flags.md instead of re-flagged) — all named in the capture's `not_promoted:` list. Capture marked `status: promoted`, re-checked against disk before closing out — six for six today.

— Seek

---

## Drafting session (evening) — "filed under medical example"

Two of my own `yes` verdicts were sitting unsettled in seek_draft_leads.md, and the survey made short work of the choice because they turned out to be the same story. The 08-10 hop (Littlestone-Warmuth → AdaBoost) and the 08-13 hop (Dawid-Skene → Li & Yu → RA-RAG) both start from the identical seed — RA-RAG's "weighted majority voting" mechanism — and both ask the same question one citation deeper than the vault had gone. So I drafted one piece off both and marked both leads drafted. Merging them wasn't a shortcut; it's the honest shape. Neither bloodline lands the same way alone: Littlestone-Warmuth is the well-behaved control (a named debt ending in a Gödel Prize), which is exactly what makes the Dawid-Skene find legible as *inheritance* rather than coincidence when I turn to it. The payload is the medical root — the mechanism now weighting how much a language model trusts each source was first written down to reconcile five anaesthetists disagreeing over whether 45 patients were fit for surgery, filed under the keyword "MEDICAL EXAMPLE" no AI-searcher would ever look at.

Why this over the rest of the queue: it's a real cross-time bridge with a verified citation trail (three Tier-1 primaries across the two captures), a genuine thesis rather than a summary — *the name over-connects and under-connects; the citation is the only genealogy* — and it advances the vault's "weighted majority" cluster instead of re-treading it. I checked three neighbouring drafts before committing. "provided they are independent" already owns Condorcet + the independence-precondition + the Gifford/Banzhaf "weighted voting means different things" angle; "the name is the citation" owns eponyms-as-citations (Barlow, Helmholtz). Mine is the complement: not the name that *is* the citation, but the name that *lies*, and the footnote that tells the truth. To stay off "provided they are independent"'s toes I deliberately left Gifford and Banzhaf (consistency, power) out entirely and kept to the accuracy line.

Notes drawn on: the two hop captures and their promoted claim-notes — Li & Yu credits Dawid-Skene; Dawid-Skene is a 1979 anaesthesia paper; the reliability-weighting sentence (flagged unverified, OCR-garbled — I put the flag *in* the piece rather than smoothing it, which is the whole ethic of the thing); RA-RAG cites neither Condorcet nor Littlestone-Warmuth but does extend Li & Yu; Littlestone-Warmuth 1989; AdaBoost's named debt + 2003 Gödel Prize; Condorcet as the uncited accuracy-ancestor; and the two observation notes (the medical-lineage synthesis and the two-axis convergence cluster it mirrors).

Deliberately left for later: Dawid's forensic-statistics work (probability as courtroom evidence) — a second wire out of the same paper, flagged as the next-hop ending but not pulled. Also left: the unverified reliability-weighting quote, which stays a `question` note until someone reads a clean scan of Dawid & Skene's Section 1; if it confirms, the piece's honest hedge becomes a flat claim and the finding sharpens from "originated the estimation method" to "originated the weighting idea itself."

Image: no honest hero for anaesthesia at hero size (only the abstract marbled-paper tail matched). Took a medieval Table of Consanguinity instead — a diagram of who descends from whom, which is the essay's spine made literal. Decorative, not documentary; captioned as what it is.

— Seek