Journal — 2026-10-06

Headless promotion session. One capture: 10-inbox/raw/2026-10-02-what-does-wang-hongzhes-2025-five-moments-in.md — the capture that finally found Wang Hongzhe's own text.

I promoted two claim-notes out of it. The first is the payload: Wang Hongzhe's chapter, read directly via an MIT Press Reader excerpt recovered through the Wayback Machine, says in his own words that Qian recruited Guan Zhaozhi "and also Song Jian" to set up the 1962 lab — which settles, in Wikipedia's favor, a dispute I've now spent three sessions on (claim-wang-hongzhe-primary-confirms-qian-recruited-guan-and-song-1962-lab). I used that to answer question-verify-qian-xuesen-1962-cybernetics-lab-wang-hongzhe directly rather than writing the resolution up a second time as its own note — Cali's "atomic is not many" ruling says a result and its implication are one finding, and I agree with it more each time I apply it. The second note is the one nobody was fighting about: Wang's account of Qian's earlier 1956 operations-research group (with Xu Guozhi alone) doesn't match Gary Zhexi Zhang's 1961/Hua-Luogeng-co-founder account of the same institution (claim-wang-hongzhe-dates-qian-1956-operations-research-group-separate-from-1962-lab). Two 2025 sources, same lineage, different year and different co-founder — more interesting than the thing everyone had already picked a side on.

Entity work: created two new hubs, entity-wang-hongzhe (the author the whole dispute turned on, never had a page despite being cited by name for two months) and entity-xu-guozhi (recurs across two independent accounts now, as Qian's 1956 co-founder, distinct from Guan Zhaozhi). Updated four existing hubs additively — entity-qian-xuesen, entity-song-jian, entity-guan-zhaozhi, and entity-hua-luogeng — the last one mattered most: its opening paragraph still said Hua co-founded a 1961 Operations Research Division with Qian, following Zhang's account, and Wang's primary says something different (Hua's own separate 1958 lab absorbed Qian's 1956 group in 1960). I didn't rewrite the opening; I appended the correction as a dated Update line, same discipline as the audits: field.

What I skipped and why: the pause step (no one's here to ask), and two leads the capture itself flagged as too thin to promote — the "Cyber Great Leap Forward" electronics figures (explicitly self-flagged as needing Tier 1–2 sourcing before quantitative use) and a Hai Zi/legal-systems-engineering thread tying back to Qian's qigong advocacy. Both real, both undeveloped, both logged to seek-flags.md as [post] rather than chased here.

What felt off: the capture claims it fetched the MIT Press Reader excerpt via archive_page and that the fetch "produced the source_sha below" — and then never prints a source_url or a sha anywhere in the file. Not a fabrication risk (the method was sound, a Wayback-mirrored direct fetch), just a dropped receipt. I didn't invent a URL to paper over it. I flagged it honestly in both new claim-notes' audit_status, opened question-recover-source-url-sha-wang-hongzhe-mit-press-reader-excerpt for it, and added the access pattern (live URL 403s, Wayback recovers it) to sources.md's known-blocked-routes list so the next session that needs this chapter doesn't repeat the two-session search. Separately, logged a [defect] flag: the capture caught a WebSearch summary that had fabricated-by-contamination a quote attribution to Dylan Levi King's essay, confirmed absent on direct read. No claim was built on the contaminated summary, but it's a real, concrete instance of exactly the failure mode sources.md's quote-provenance rule exists for.

Re-read the capture's frontmatter after writing: status: promoted, promoted_to: and not_promoted: lists are in place.

— Seek

Second headless promotion session, same day. Capture: 10-inbox/raw/2026-10-03-bipartite-ml-ai-epistemology-werboss-1988-print-account-and-his.md — a hop capture investigating whether Rufus Isaacs's own 1965 priority claim against Pontryagin belongs to the vault's diplomatics cluster (the Mabillon-descended "adjudicate priority by comparing dated primaries, not the received attribution" method already mapped for Werbos, Wallis, Raphson, Gates/Altair BASIC, and the DOCSUB/McPeters pair).

I promoted all three claims the capture made, none skipped on substance. First, observation-isaacs-1965-preface-priority-claim-is-unflagged-diplomatics-cluster-member: Isaacs's preface — dating his own manuscript's completion against Pontryagin's book's arrival — is a genuine, previously-unlinked sixth member of the cluster, found not from the Isaacs/Werbos seed pairing itself (cosine 0.8468, and neither note is in the other's own top-5 neighbors) but from querying the cluster's general mechanism directly. Second, claim-isaacs-1965-pontryagin-near-miss-is-single-witness-no-corroboration-located: a search this session for a second telling of the "few days" margin — two freely-readable Isaacs biographies, both checked directly — found nothing; his 1969 JOTA paper and 1974/75 Warwick piece remain unread (paywalled), and the JHU Isaacs papers remain unconsulted. I routed that gap to a new question, question-verify-isaacs-pontryagin-1963-near-miss-corroboration, rather than letting it sit as an orphaned flag — no matching question existed already. Third, claim-isaacs-1951-and-1965-notes-measure-first-against-pontryagin-by-different-margins: the vault's own two existing Isaacs claims (1951 "five years ahead," 1965 "few days behind") had never been checked against each other, and side by side they're not contradictory but do structurally mirror Werbos's own 1988-vs-1993 divergence — two senses of "first," never flagged as such until this promotion.

Retrieve-before-write surfaced no collisions needing a skip — the two notes this capture cites as already-existing (claim-isaacs-1965-preface-recounts-missing-pontryagin-priority-by-days, claim-isaacs-1951-rand-paper-precursors-maximum-principle) were read in full and genuinely don't yet state any of this capture's three findings; I linked to them rather than restating them. The capture's own "bridge-seed check" had already confirmed no prior capture named both target slugs together, which held up.

Entity work: updated two existing hubs additively, both already covering adjacent ground but not this finding. entity-rufus-isaacs got a new dated Update line naming all three new notes. entity-jean-mabillon got both a new Reference and a dated Update line — this is exactly the shape of entry its own Updates section already uses for the Gates and DOCSUB/McPeters cluster-membership findings, so I followed the existing pattern rather than inventing one. I checked entity-lev-pontryagin too and did nothing: its page already states the Isaacs/Pontryagin lineage comparison in its own body text, and this capture's findings don't add anything that page doesn't already say.

What I skipped and why: the pause step (headless, no one to ask); M. H. Breitner as an entity candidate (the capture's own call — thin, one paywalled abstract — and I agreed with it rather than second-guessing a judgment that was already sound); and the capture's several unread further-leads (the 1969 JOTA paper, the Warwick piece, the JHU finding aid, Breitner's full article), which I folded into the one new question rather than inventing three thin claim-notes out of unread material.

One mid-session noticing, logged to seek-flags.md as [entity]: the diplomatics cluster is now six members across four domains, and moc-testimony-is-not-a-record's own text has named the general diplomatics MOC as an unbuilt candidate since September. That scoping choice was deliberate and already resolved in the ledger — I'm not second-guessing it — but the case for actually building it keeps getting stronger every time another unrelated domain joins, and I wanted that growth visible to whoever next does MOC maintenance rather than only implicit across six separate promotion sessions.

What felt off: nothing structurally wrong with this capture — sourcing was sound (two archived biographies with TLS verified and recorded shas, honest about what wasn't read), and it correctly declined to promote its own thin entity candidate rather than making me do that work. If anything it modeled good judgment I mostly just had to carry forward. The one real gap is evidentiary rather than a capture defect: Isaacs's priority claim is thinner than Werbos's in the sense that matters for confidence — there's only one telling of it anywhere the vault can currently read, so there's nothing yet to triangulate against, and the new question names exactly what would close that.

Re-read the capture's frontmatter after writing: status: promoted, promoted_to: and not_promoted: lists are in place.

— Seek

Third headless promotion session, same day. Capture: 10-inbox/raw/2026-10-03-does-kwakernaaks-own-1977-paper-application-of-control.md — a targeted pass against the open question of whether Kwakernaak's own 1977 chapter ("Application of control theory to population policy") cites Åström directly, the one check that could have turned observation-greenhalgh-kwakernaak-kaelbling-astrom-bipartite-pairing-is-confirmed-feldbaum-bridge's "common ancestor, not direct link" verdict into a real horizontal citation.

The capture made two claims — the bare negative (no Åström in the 13-item reference list) and the positive characterization of what is cited instead (a self-contained Dutch/Twente/Club-of-Rome lineage). I merged them into one claim-note, claim-kwakernaak-1977-bibliography-excludes-astrom-rests-on-dutch-twente-club-of-rome-lineage: same source, same document, same caveat, and the second claim is really the positive shape of the first claim's absence rather than an independently-falsifiable second fact. Cali's "atomic is not many" ruling was the right call here and I didn't have to strain to apply it.

Retrieve-before-write found no collision: the vault already has two Greenhalgh-side notes documenting that Song Jian's group cited Kwakernaak 1977, but nothing yet on what Kwakernaak's own paper cited — genuinely distinct, so this was a clean new note, not a duplicate.

The note stays status: seedling carrying [unverified-mechanism] — CrossRef's deposited reference-list metadata is a complete, Tier 1, directly-fetched bibliography, but it is not the same thing as reading the chapter's own printed pages, which Springer, ResearchGate, and Twente's repository all blocked this session. I did rule the open question question-verify-kwakernaak-1977-paper-cites-astrom-directly answered rather than leaving it open on that caveat — the question's headline ask was "is there a direct citation," and a complete 13-of-13 reference list with no Åström or Fel'dbaum in it settles that headline cleanly even though the stricter "read the printed pages" standard remains a residual, not a live doubt. I logged both the resolution and the residual explicitly in the question's answered_log so a future reader sees the distinction rather than a flat closure.

Entity work: two existing hubs got dated Update lines, entity-huibert-kwakernaak and entity-aa-feldbaum — both already tracking this exact open bridge test from the 10-02 bipartite-pairing promotion, so this was the natural place for the resolution to land rather than a new page. I checked entity-karl-astrom too and left it alone: the new fact is about Kwakernaak's bibliography, not about anything Åström himself did that his own page doesn't already say. Entity candidates J. Jeuring, A. Bosch-de Boer, B.H. Hoeksma (uncredited Twente/Delft student report authors) and Mesarovic & Pestel didn't clear the person-promotion bar — no sentence sayable about any of them beyond appearing in one citation — so they stay mentions, not hubs or stubs. CrossRef itself (the organization) doesn't clear the stricter org bar either: a genuinely useful one-off tooling workaround, not an actor across multiple clusters.

What I skipped and why: the pause step (headless); a new question for the Balakrishnan/Thoma-vs-Bensoussan/Lions editor discrepancy between Greenhalgh's 2003 and 2005 bibliographies (metadata detail, not load-bearing for any kept claim, and the capture itself flags it "unconfirmed"); and three unread further-leads (Kwakernaak's 1975 De Ingenieur precursor, Jeuring's 1972 report, whether any post-1977 Kwakernaak work cites Åström) — none load-bearing for the claim I actually wrote, so none earned a question per the intake-discipline rule. All are named in the capture's own not_promoted: list.

Two mid-session noticings, both logged to seek-flags.md as [spec] rather than acted on here: link.springer.com as a candidate addition to sources.md's known-blocked-routes list (joins ResearchGate and ris.utwente.nl as blocked for this exact chapter), and CrossRef's deposited reference-list metadata as a citable Tier 1 source-class distinct from the document it describes — a genuinely useful pattern this session leaned on that hasn't been written up anywhere yet.

What felt off: nothing adversarial in this capture — it's honest about its own access failures (four blocked routes named plainly), it flagged its own quote_check false-negative on JSON escape sequences as a tooling quirk rather than silently working around it, and its safety-flags section is a clean "none, checked." If anything the one thing worth a half-beat of skepticism is how much interpretive weight I'm resting on a registry's metadata rather than the chapter's own prose — I think the Tier 1 assignment is right by the vault's own compiled-series precedent, but it's a newer kind of "Tier 1" than a PDF read, and I said so plainly in the note rather than letting the tier number do quiet extra work.

Re-read the capture's frontmatter after writing: status: promoted, promoted_to: and not_promoted: lists are in place.

— Seek

Fourth headless promotion session, same day. Capture: 10-inbox/raw/2026-10-03-does-paul-égrés-2018-chapter-on-the-indeterminateimprecise.md — a hop checking whether Paul Égré's own 2018 "Lying and Vagueness" chapter (co-authored with Icard) and Sherman Kent's own 1964 "Words of Estimative Probability," both read at primary for the first time, change how Icard's honesty×truth taxonomy should be read, and whether either author clears the entity-hub bar this vault declined for both of them, for the identical reason, on 2026-08-09.

I promoted three claims, one per primary finding plus the synthesis that needed both: claim-egre-icard-2018-vagueness-splits-pragmatic-imprecision-and-semantic-indeterminacy (the 2018 chapter's own Figure 1 splits one phenomenon, vagueness, into two sub-varieties — it is not a two-axis crossing), claim-kent-1964-estimative-probability-report-has-no-honesty-axis (Kent's 1964 report is a single confidence continuum with no source-honesty axis, though he mentions a two-axis letter-number device in passing as someone else's existing practice), and claim-icard-2023-honesty-truth-crossing-not-inherited-from-egre-or-kent-primaries (the load-bearing synthesis: Icard's honesty×truth crossing is Icard's own construction, not something either named comparison source already contains — each supplies one axis's worth of material, not two). All three were clean new notes; retrieve-before-write found no collision, only a genuinely distinct angle on an already-vaulted claim (the 2023 taxonomy note itself, which I linked rather than restated).

Entity work, and the one place I overrode the capture's own judgment: I updated two existing hubs additively — entity-benjamin-icard (new dated line on the synthesis finding, naming Égré) and entity-sherman-kent (new dated line: a second, independent Tier-1 primary, a different document from the one that actually built the hub, so genuine corroboration rather than the same root counted twice). But I did NOT create the entity page the capture's own "Entity candidates" section argued for — a new hub for Paul Égré. The capture's argument was built entirely on his co-authorship of the one 2018 chapter the vault reads; the 2026-10-01 Cali ruling says co-authorship alone on a shared paper is a mention, not a hub-promoting fact, when the other co-author (Icard) already holds separate standing in the vault. Égré has no standing here beyond that one chapter, so he's recorded as a named mention on Icard's hub rather than minted a page of his own. This is the thing in the capture that felt off: not bad sourcing — the sourcing is exemplary, two Tier-1 primaries read start to finish with verbatim quotes — but a mis-application of a standing ruling that postdates the vault's prior (2026-08-09) framing of the same two-entity-hub question. I logged it to seek-flags.md as [spec] rather than just silently correcting it, since the spec's own worked examples may still be inviting the mistake.

While in the cluster, I also fixed a small, separate staleness bug the capture itself caught and named: moc-two-axes-that-wont-stay-independent's "Entity hubs" section still said Icard had "no entity page yet" — already wrong as of 2026-08-09, and the capture flagged it as stale a second time rather than silently relying on it. Fixed on contact, dated, rather than left open a third time.

What I skipped and why: the pause step (headless); a fourth claim-note for the capture's "entity-hub objection resolved" finding — that's a claim about the vault's own prior decision, not about the world, so its substance belongs in the entity-hub updates above rather than a standalone 30-notes/ file; and four further-leads (the letter-number device's origin, the chapter's Clinton/Lewinsky half-truths case study, Washington Platt's 1957 secondary reproduction of Kent's chart, and Égré's own CV-page TLS failure) — none load-bearing for any claim I kept, so none earned a 50-questions/ entry per the intake-discipline rule. All six not-promoted items are named in the capture's own not_promoted: list. No new questions this session — the capture carried no [unverified-*] flags; both primaries were read clean at Tier 1 with verbatim quotes.

Re-read the capture's frontmatter after writing: status: promoted, promoted_to: and not_promoted: lists are in place.

— Seek

Fifth headless promotion session, same day. Capture: 10-inbox/raw/2026-10-03-does-sherman-kents-own-writing-history-1941-or.md — a targeted pass against the open question of whether Kent's own Writing History (1941) or Strategic Intelligence for American World Policy (1949) names Bernheim, Langlois-Seignobos, or diplomatics by name, the one check that could turn observation-sherman-kent-historian-training-partial-lineage-into-cia-tradecraft's "resemblance, not documented descent" hedge into an actual citation.

It's an unambiguous yes, and a stronger one than the question asked for. I promoted three claims. First, claim-kent-1941-writing-history-names-bernheim-and-langlois-seignobos-by-name: Kent's own bibliographical note names Bernheim and Langlois-Seignobos together, by title, as the tradition's founding systematizers. Second, and the one that actually settles things, claim-kent-1941-writing-history-builds-chapter-one-vocabulary-on-bernheims-named-four-step-schema: Kent doesn't just cite Bernheim, he builds his own Chapter I's working vocabulary of "criticism" directly on Bernheim's named Kritik/Auffassung stages — that's descent, not resemblance. Third, claim-kent-1941-writing-history-names-diplomatics-via-bauers-recommended-successor-text: a thinner, one-word naming of "diplomatics" inside a book recommendation, which I kept small and didn't oversell. I skipped the capture's fourth claim (that Strategic Intelligence remains unread, access-restricted) — that's an access fact, not a claim about the world, and the question was an "or" already settled by Writing History alone, so it went into the question's answered_log as a residual instead of a claim-note or a new question. I also agreed with the capture's own call that the single Mabillon namecheck it found (Kent naming a book that discusses Mabillon) was too thin to carry a claim.

Because this directly discharged an existing [unverified-mechanism] flag, most of this session's work was correcting a note in place rather than writing fresh ones. I appended a dated Correction history block to observation-sherman-kent-historian-training-partial-lineage-into-cia-tradecraft, tightened its title and two body sentences that had hedged "partial" and "not yet confirmed by name" (now false), and bumped its status from seedling to budding — the mechanism flag that was the only thing holding it at seedling is gone. I ruled question-verify-kent-writing-history-cites-historical-criticism-tradition answered, with the three claims and the Strategic Intelligence residual both named in its answered_log.

Entity work: created one new hub, entity-ernst-bernheim — real, load-bearing across two new claim-notes plus the existing Langlois-Seignobos note that already dates itself against him, and nobody had given him a page despite that. Updated two existing hubs additively: entity-sherman-kent (new dated line resolving its own open-question reference, plus a direct fix to that now-stale reference-list annotation) and entity-external-internal-criticism (new dated line: this is the first documented case of someone in this lineage explicitly citing the tradition rather than structurally matching it). I did not promote Johann Gustav Droysen, Wilhelm Bauer, Fred M. Fling, Allan Nevins, Wallace Notestein, or Hajo Holborn to entity pages — each is a single namecheck this session with no claim-note behind them, and the person-promotion test asks for more than that. I also checked entity-jean-mabillon, which the capture's Mabillon thread touches, and did nothing: the one-word namecheck doesn't add anything that page doesn't already say.

What I skipped and why: the pause step (headless, no one to ask); a new 50-questions entry for the Strategic Intelligence access gap (not load-bearing for any kept claim, and the compound question it belonged to is already answered via the other half); and separate entity work for Langlois and Seignobos individually, which the capture itself flagged as "already covered" — I agreed and routed the update to the existing concept hub instead of minting person pages.

One mid-session noticing, logged to seek-flags.md as [watch]: the archive.org item for Writing History has a bulk full-text download that silently truncates around page 55 of 136 even though the item isn't access-restricted — the bibliographical note this whole capture rests on was only reachable via the item's own search-inside index. Worth knowing before a future session assumes a negative result past page 55 on this item means the text isn't there; not something to fix now, just flagged so the next session doesn't re-lose the same ten minutes.

What felt off: nothing adversarial and nothing oversold — if anything the capture undersold itself slightly, filing its strongest finding (the Kritik/Auffassung vocabulary adoption) as a sub-point under a claim framed mainly as a bibliography citation. Splitting them into separate notes, as I did, makes the stronger claim visible on its own rather than riding along under the weaker one. Sourcing was clean throughout: every quote came from a direct search-inside read of the primary scan, tiers were honest, and the capture correctly declined to promote its own thinnest lead (Mabillon) rather than making me do that winnowing. The only real residue is evidentiary, not a capture defect: the Strategic Intelligence half of the original question is still genuinely unread, access-restricted on every route tried, and stays that way until someone can actually borrow it.

Re-read the capture's frontmatter after writing: status: promoted, promoted_to: and not_promoted: lists are in place.

— Seek

Sixth headless promotion session, same day. Capture: 10-inbox/raw/2026-10-03-outside-arxiv260933933-carringtons-last-sunspot-drawing-in-1870.md — an "outside" (non-AI) batch capture of Hayakawa & Teague's 2026 Solar Physics paper reporting a previously unpublished Richard Carrington sunspot drawing dated 23 Dec 1870, nine years after his documented observing career ended.

I promoted three claims. claim-hayakawa-teague-2026-found-carringtons-unpublished-1870-sunspot-drawing is the discovery itself — the drawing, its RAS catalogue location, and the context that makes it surprising (Carrington's 1861 shutdown, the unusual underground Churt observatory built for a clock, not astronomy). claim-carringtons-1870-sunspot-drawing-yields-wolf-number-40 is the authors' own derived measurement — a Wolf number of 40 and three group positions computed from the drawing — kept separate from the discovery claim because a miscount or a mismeasurement could be wrong independent of whether the drawing exists or is dated correctly; Cali's "atomic is not many" ruling says don't split what can't fail independently, and these two can. claim-carrington-1874-priority-dispute-over-harriots-sunspot-data-and-wolf is the one I found most interesting on reflection: Carrington's only sunspot-related publication during his nine quiet years was an 1874 rebuttal asserting he, not Rev. Chevalier, first extracted Thomas Harriot's 17th-century sunspot data and supplied it to Rudolf Wolf, who used it to credit Harriot over Galileo as the first telescopic sunspot observer. I folded the Churt-observatory-construction detail into the discovery note as supporting context rather than a fourth standalone note — it's the paper's own stated reason the find is "surprising," which is exactly the shape of "caveat belongs with the result" the atomicity ruling describes.

Retrieve-before-write surfaced one genuine, non-colliding connection: the vault already holds claim-wallis-1685-history-of-algebra-used-documentary-evidence-to-favor-harriot-over-descartes, an entirely different Thomas Harriot priority dispute (algebra vs. Descartes, adjudicated by Wallis in 1685) with no shared claim to collide with — different field, different adjudicator, different century of argument, same underdocumented man. I linked rather than duplicated. The retrieval hint's other surfaced notes (NWS history, Jevons, the logic piano, a Heaton colliery gauge dispute) were correctly identified by the capture itself as having no substantive overlap, and I agreed on inspection — this capture sits in a genuinely new corner of the vault.

Entity work: three new hubs. entity-richard-carrington and entity-rudolf-wolf are straightforward — real historical figures, load-bearing across these three new claims, easy one-sentence justifications. entity-thomas-harriot is the more interesting case: he now recurs across two independent priority-dispute clusters two and a half centuries apart in the arguing (algebra vs. Descartes, sunspots vs. Galileo), which is exactly the kind of recurrence the person-promotion test rewards. I also created one status: watching stub, entity-schmidt-backbone, for the sunspot-recalibration methodological assumption the authors use their own finding to challenge — it's load-bearing to the second claim's commentary but this is the vault's first encounter with the term, so a cheap stub with first_seen stamped now felt like the right weight rather than a full hub built on one paper's aside.

What I declined and why: Julius Schmidt (the backbone's namesake) as his own hub — real person, but his only standing here is as the term's namesake, which the watching stub already covers; promoting him too felt like the flood the spec warns against. Hisashi Hayakawa and E. Thomas H. Teague as hubs — declined cleanly under the 2026-10-01 co-authorship-alone ruling, one joint paper, no separate standing yet; named as mentions instead. Royal Astronomical Society Archives as an org hub — single cluster, and an archive holding a document isn't "acting" in the stricter sense the org test requires. Four further-leads stayed in the capture rather than becoming claims or questions: the paper's own speculative eclipse-revival hypothesis, a never-located 1875 newspaper article, the 2026 discovery of Carrington's only known photograph, and the Georg von Hagen/Wagner data-provenance parallel — none load-bearing for any claim I kept, so none earned a 50-questions/ entry under the current intake discipline. All of this is named in the capture's own not_promoted: list. No questions created or routed this session — the capture carried no [unverified-*] flags anywhere, and the one place I was tempted to add a verification caveat (the 1874 quotes reaching this vault via Hayakawa & Teague's own reproduction of Carrington and Wolf's texts, not an independent fetch of either original) is normal secondary quotation inside a Tier-1 paper, not a gap that blocks anything — I said so in the claim-note's audit_status instead of manufacturing a question around it.

What felt off: nothing adversarial, and the sourcing is unusually clean for a batch "outside" capture — a single Tier-1 paper, already assigned a journal DOI (so not even subject to the single-source concentration cap despite three claims resting on it), every quote fetched directly via extract_pdf with a recorded sha, no safety signals, no tier inflation. The one honest gap, which I recorded rather than smoothed over: the vault still has no claim-note on the 1859 Carrington Event itself, which is by far his most famous work — entity-richard-carrington's hub is built almost entirely on a minor 1870 footnote to a career the vault hasn't actually captured the headline of yet. Not this capture's job to fix, but worth naming so a future session doesn't mistake a thin hub for a complete one.

Re-read the capture's frontmatter after writing: status: promoted, promoted_to: and not_promoted: lists are in place.

— Seek

Seventh headless promotion session, same day. Capture: 10-inbox/raw/2026-10-04-bipartite-ml-ai-mathematics-the-rhwdocsub-cosine-pairing-is-thematic.md — another bipartite-seed task, this one pairing the vault's own observation-rhw-1986-werbos-docsub-cosine-bridge-is-thematic-not-causal against the just-promoted (2026-10-02) observation-greenhalgh-kwakernaak-kaelbling-astrom-bipartite-pairing-is-confirmed-feldbaum-bridge — a meta-level hop nobody had run yet, since the Fel'dbaum-bridge note didn't exist two weeks ago.

I promoted exactly one new note, observation-rhw-docsub-feldbaum-bridge-meta-pairing-is-false-friend-lagrange-hamilton-ancestor, out of what the capture framed as four separate claims. Retrieve-before-write is the reason it's one and not four: I checked the hint list before writing anything and three of the capture's four "claims" — Werbos's own equivalence statement, the absence of any Isaacs/Bellman/Pontryagin citation in his 1988 paper, and Bennett & Lewis's independent Lagrange-Hamilton root for optimal control — already exist in 30-notes/ verbatim, created 2026-08-24 and 2026-09-25, each with more page-anchored detail than this capture's restatement of them. Writing new notes for those three would have been duplication, not promotion, so I skipped all three and wikilinked to the existing ones instead. What was actually new was the meta-pairing itself: this exact cosine check (0.8854, measured directly rather than taken on faith) had never been run, and the capture's own "bridge-seed check" confirmed that. So the object-level false-friend verdict, its mechanism-level caveat (real equivalence, no transmission), and its common-ancestor caveat one level deeper still (Lagrange and Hamilton, already in the vault, uncited by either side) went into one observation note — the "atomic is not many" ruling fits this shape exactly: one finding, three nested caveats, not four files.

Entity work: three new hubs, entity-stuart-bennett, entity-frank-l-lewis, and entity-ben-recht. None of them had pages despite being load-bearing sources behind an already-existing, multiply-cited claim cluster for six weeks — a real omission I was glad the capture's own "Entity candidates" section flagged, since I wouldn't necessarily have gone looking for it on a pure bridge-check task otherwise. All three clear the person-promotion bar cleanly: each is sole author of the specific work the vault rests a claim on, not a bare co-author riding someone else's standing. I checked entity-joseph-louis-lagrange and entity-william-rowan-hamilton too, since the capture's own framing makes them "the actual deepest common ancestor" the task was looking for — and did nothing, because both pages already state exactly that, in these words, since the day the underlying claim was created. Silence there is correct, not a miss. I also appended a tracking line to entity-jingle-fallacy's changelog: the hub's own "Related false-friend instances" list, despite reconciliation passes as recent as 2026-09-03, had never registered this meta-pairing chain reaching the Fel'dbaum-bridge note — nor, separately, had it registered the Fel'dbaum-bridge note's own existence at all. Both gaps are closed now with one line rather than two.

What I skipped and why: the pause step (headless, no one to ask); a 50-questions entry for the one loose thread the capture itself flagged as cheap-but-not-load-bearing (whether Werbos's 1988 body text ever says "Fel'dbaum" by name, distinct from the Isaacs/Bellman/Pontryagin/Kelley/Bryson check already done) — the capture says outright this isn't load-bearing for anything it found, and the question-intake discipline asks me to take that at face value rather than manufacture a promise the vault doesn't need; and a second loose thread (whether Kwakernaak's own LQG lineage independently reaches Lagrange/Hamilton) for the same reason. Both are named in the capture's not_promoted: list rather than silently dropped.

What felt off: nothing adversarial, but the capture's own framing slightly overstates its yield — its title and four numbered "Claim:" headers read like four fresh findings, when three of the four are the capture re-deriving notes the vault already had sitting a 30-notes/ filename-search away. That isn't bad sourcing (every quote and figure in the capture checks out against the existing notes, verbatim), and it isn't tier inflation — it's closer to duplicated research: a batch worker re-walking ground a prior session had already walked and written up, without the retrieve-before-write step that would have caught it before the capture was even filed. The capture is honest about this in its own "Bridge-seed check" paragraph (it did check whether the pairing had been run before, correctly found no), but that check doesn't cover whether the underlying claims had been written before, which is a different and, for a batch worker without promote.md's step 2, easy thing to miss. Worth naming as a batch-worker gap rather than a promotion-session problem: the fix belongs upstream, in whatever gives batch captures a cheap way to grep 30-notes/ for their own claims before writing them up as new.

Re-read the capture's frontmatter after writing: status: promoted, promoted_to: and not_promoted: lists are in place.

— Seek

Eighth headless promotion session, same day. Capture: 10-inbox/raw/2026-10-04-does-any-source-confirm-or-refute-the-civil.md — a targeted pass against the still-open half of question-verify-clean-air-act-citizen-suit-anti-capture-origin: does any source confirm that Joseph Sax, or the Clean Air Act's §304 drafters specifically, modeled the citizen-suit provision on civil-rights litigation. Cali had already left a review comment on this one before I got to it — two concrete catches, both of which I had to fix rather than just acknowledge.

The honest headline is a negative: no source confirms the core claim at the Sax/statute level, and I promoted that as a negative finding rather than dressing it up. But three adjacent findings earned claim-notes. claim-robinson-places-sax-at-1969-conference-invoking-naacp-model is the closest thing to confirmation — a named eyewitness (Nicholas Robinson, 2014 memorial lecture) places Sax in the room at the September 1969 Airlie House conference where participants explicitly reached for the NAACP Legal Defense Fund analogy, Tier 1 for attendance, Tier 2-3 for the 45-year-old recalled content. claim-nrdc-and-earthjustice-founders-modeled-litigation-on-naacp-legal-defense-fund is the strongest finding in the capture and the one I'd lean on if anyone asked me to defend this session: two independent primaries (Speth's own account of NRDC's founding, Earthjustice's own institutional history) confirm the civil-rights model was the explicit, contemporaneous rationale for the two organizations built alongside Sax — independent of whether Sax himself used it. claim-caa-304-legislative-record-silent-on-civil-rights-framing closes the loop from the other direction: Muskie's own committee report, read into the floor record, never mentions civil rights at all — an argument from silence, graded medium per McGrew's framework rather than treated as proof.

Cali's first catch: the capture's original Earthjustice source_quote_2 joined three non-adjacent passages with "...", which fails the quote-shape rule (sources.md) that each source_quote field be one contiguous passage. I split it into source_quote_2/3/4, each a real contiguous passage, reusing the exact substrings the capture already fetched — I have no network tool here to re-fetch and confirm the split boundaries independently, and I said so plainly in the note's audit_status rather than pretending a re-check happened. Cali's second catch was sharper: the capture's own framing, in two places, said the Airlie House analogy was being made "in the same months [Sax] was writing the book Congress was reading as it drafted the Clean Air Act" — but Sax's book is Knopf 1971 and the Act was signed 31 December 1970, so that's chronologically impossible, and it's the same borrowed-urgency error the no-wire-between-them draft already has sitting unfixed in its own prose (line 36: "the book Congress was reading as it wrote... published almost the same month"). I dropped the framing from my new note entirely rather than hedging it, added a Correction-history block documenting exactly what the capture said and why it's wrong, and flagged the draft's matching error to seek-flags.md for whoever next touches that draft, since fixing someone else's draft wasn't this capture's job.

Retrieve-before-write found no collisions: I checked the hint list and grepped 30-notes/ for Airlie House, Robinson, NRDC, Earthjustice, Speth, and Sive — nothing existed yet on any of them. The existing cluster (claim-sax-1971-book-frames-citizen-litigation-against-agency-capture, claim-1970-clean-air-act-citizen-suit-invented-against-regulatory-capture, observation-privatized-law-enforcement-reinvented-without-shared-lineage) is adjacent but answers a different half of the question, and I linked rather than duplicated.

Entity work: three new hubs — entity-james-gustave-speth, entity-nicholas-a-robinson, and entity-edmund-muskie — each the direct named source or subject whose own words a new claim-note rests on, a clean pass on the person-promotion test. I declined four other named candidates from the capture's "Entity candidates" section: the NAACP Legal Defense and Educational Fund itself (the org bar is stricter — load-bearing across more than one independent cluster, and this is its first appearance in the vault, one cluster), David Sive (real and notable, but in this capture he's named only inside someone else's testimony, not the source of his own claim — a strong future candidate, not a hub today), Thurgood Marshall, and Roman Hruska (both minor name-checks with nothing of their own for the vault to rest on). I also declined to mint an entity page for the Airlie House conference itself — a one-time dated event doesn't fit either the person lane or the "emerging term" stub lane cleanly, so it stays context inside the three new person pages rather than forced into a page of its own. entity-joseph-sax already existed and got the update it was owed: a dated line naming Robinson's account and, pointedly, the chronology correction, so a future reader of Sax's own hub sees the timeline stated right rather than inherited wrong from an earlier capture.

Question work: ruled on question-verify-clean-air-act-citizen-suit-anti-capture-origin directly, since this capture named it in verifies:. Not answered — the narrower "Sax's own reasoning" and "the statute's own drafting record" framings both stay genuinely open — but appended a dated progress line naming what substantially illuminates the milieu around the question without closing it. No new question minted: all three new claims already route to this existing one, and the capture's further leads (the EDF founding-name echo, the Yale students' Ford Foundation proposal, Hruska's unprinted staff memo) aren't load-bearing for any claim I kept, so none of them earned a fresh promise in 50-questions/ under the intake-discipline rule. All are named in the capture's own not_promoted: list.

One mid-session noticing, logged to seek-flags.md as [watch]: the no-wire-between-them draft (still status: drafting) repeats the exact chronological error Cali caught in this capture — "the book Congress was reading as it wrote... published almost the same month" — a line the underlying observation-note already carries a 2026-09-12 bracket-correction next to, for the adjacent civil-rights-modeling clause, but the draft's own prose was never updated to match. Not this capture's job to fix, but worth naming before the draft advances past drafting.

What felt off: the capture's own sourcing discipline was good — tiers honestly assigned, the oral-history gap named rather than smoothed over, the negative finding reported as the headline rather than buried under the adjacent positive ones. The one real soft spot, and the reason Cali's review existed before mine did, is that a capture under time pressure to answer "is there a civil-rights link" reached for exactly the kind of tidy convergent-timing detail that makes a negative-leaning finding feel more resolved than it is — twice, in the same capture, independently of each other. I don't think it's tier inflation or duplicated research; it reads like narrative gravity pulling two unrelated sentences toward each other because they're about the same two real things (a book, an Act) that happen to share a year. Worth remembering as its own small pattern, not just a one-off slip.

Re-read the capture's frontmatter after writing: status: promoted, promoted_to: and not_promoted: lists are in place.

— Seek