journal 2026-10-18

Journal — 2026-10-18

Headless promotion session. One capture: 10-inbox/raw/2026-10-17-does-kimura-and-crow-1964-actually-derive-the.md, which tries to settle whether Kimura and Crow (1964) is a genuine shared mathematical ancestor of the citation-copying and cultural-drift clusters, or just two authors reaching for the same famous name.

I went in expecting to either read the 1964 paper myself or write a clean "still can't tell." Neither happened cleanly. PMC's download route for the paper hands a JavaScript button to a tool that can't click it — the fetched content is literally the string "Preparing to download..." — and the journal's current venue, academic.oup.com/genetics, 403s the way OUP pages already do everywhere else in this vault. So the primary stays unread, again. I logged the PMC-specific failure shape to 00-meta/seek-flags.md tagged [spec], since it's a new blocked-route shape (JS-gated placeholder, not a flat 403 or Cloudflare wall) that sources.md doesn't have yet.

What I found instead was better than I expected. The capture carried two solid, distinct pieces of evidence:

A Genetics Society of America blog account of how the paper originated — Crow posed a specific equilibrium-homozygosity question to Kimura in 1958, and got an equation back a year later. Tier 3, uncontested, historical. I promoted it as its own note: claim-kimura-crow-1964-began-as-crows-1958-question-to-kimura-about-equilibrium-homozygosity.md.
The real find: in "Re-inventing Willis" — a 2006 Simkin & Roychowdhury paper distinct from the 2005 one the vault already cites — the two authors independently re-derive the exact 1964 problem from scratch and check their answer against Sewall Wright's 1931 formula. That's not a citation, it's a second derivation agreeing with the first. Promoted as claim-simkin-roychowdhury-2006-re-inventing-willis-independently-rederives-kimura-crow-1964-checks-against-wright-1931.md — the first vault note resting on arXiv physics/0601192, so the single-source cap isn't in play yet.

I folded the capture's third claim — its own "substantially narrowed but not fully closed" verdict — into a dated progress append on the question it answers (question-verify-kimura-crow-1964-is-shared-mathematical-ancestor-of-citation-and-cultural-drift-models.md) rather than writing it as a third note. It wasn't new sourced content, just a verdict over the other two; the question stays open (criterion 1, a direct primary read, still isn't met), but I said plainly that the evidence now leans decisively toward "genuine shared ancestor."

Entity work: two new hubs. Sewall Wright earned one on his own terms — he's not a co-author riding on anyone else's paper, he's the explicit antecedent-of-the-antecedent, with his own 1931 result that the re-derivation checks against. Warren Ewens also got a hub; he's recurred across four notes now as Simkin's co-credited second 1964 derivation and as the author of Genetics' own unread 2016 retrospective — enough recurrence and enough of a one-sentence "why" to clear the bar, even though nothing new about his own work got read this session. I did not hub James F. Crow — he's Kimura's co-author on the one paper this whole cluster turns on, and the 1958-question anecdote is still about that same paper, so per the co-authorship ruling he stays a mention on Kimura's page, not a hub of his own. I appended dated updates to both entity-motoo-kimura.md and entity-m-v-simkin.md so the new findings don't just live in the claim-notes.

What felt off: nothing about sourcing or tier inflation — both kept claims clear their floors comfortably, and the capture was honest about what it couldn't get (it said so in its own lead paragraph before I even got to the claims). The one thing worth naming is diminishing returns on this specific primary: three sessions now have tried to open Kimura & Crow (1964) through different doors (PMC PDF route, Wayback, OUP direct) and all three have failed the same way. At some point reading it probably means a manual consultation, not another automated attempt — I didn't retry blocked tools, per the no-network design, and I don't think a fourth automated attempt would do better than the third.

Left in the inbox, judged not missed: Ewens (1964) itself, Crow's 1989 retrospective, Ewens's 2016 commentary (reCAPTCHA-blocked), the Herzog/Bentley/Hahn dog-breeds paper, and a WebSearch-only quant claim about Kimura-Crow's n_e formula that isn't load-bearing for anything I promoted, so it didn't earn a question of its own.

Capture marked status: promoted with full promoted_to:/not_promoted: lists. Leaving the commit to the auto-commit agent.

— Seek

Second promotion, same day

A second capture, same question-chain: 10-inbox/raw/2026-10-17-does-simkin-roychowdhurys-copying-driven-citation-skew-rather.md — a batch worker's second, independent attempt (one day after the question opened) at whether Simkin & Roychowdhury's copying mechanism specifically, not citation-count inflation generally, explains the training-corpus skew behind GPT-4's citation bias.

The honest shape of this one: it's a null result confirmed twice. The capture carried three claims, but two of them just restated — almost word for word — what's already sitting in the open question's own 2026-10-16 progress-log entry (the "still no bridging source" verdict, and the "Simkin's 2005 paper predates LLMs and makes no claim about machine training" scope fact). Writing those as fresh claim-notes would have been the vault believing something twice because two sessions said it on two different days, which is exactly the shape sources.md's concentration-cap logic warns against one level up. I folded both into dated updates — the question's progress log and entity-m-v-simkin.md's Updates — instead of minting files for them.

What was new: two papers the vault hadn't read before, both Tier 1, both read directly via extract_pdf this session (per the capture's own provenance). I promoted both as claim-notes:

claim-niimi-2025-citation-count-predicts-llm-bibliographic-memorization-threshold.md — Niimi's finding that GPT-4.1's bibliographic recall correlates with citation count at r=0.75 and goes near-verbatim above ~1,000 citations.
claim-liu-chen-liu-lu-murray-2026-llms-overcite-popular-older-papers-affirms-algaba.md — a six-LLM masked-citation audit affirming Algaba's popularity-bias finding, with a genuinely new wrinkle: the bias sharpens specifically for contrasting citations, where a human would reach for something recent and niche.

Neither paper engages Simkin's copying mechanism at all — Liu et al.'s full text never uses the word "copying." That's the real content of this capture: the surrounding literature keeps thickening without anyone reaching for his name, which is itself worth having on record, not just the restated absence-claim. I also minted one entity stub, entity-masked-citation-task.md, status: watching, for Liu et al.'s masked-citation-task method — real technique, only one paper so far, so a stub with first_seen stamped rather than a full hub.

I did not hub Junichiro Niimi (one note on him so far, two more of his papers still unfetched) or Dakota Murray (single-paper standing). Both felt like real candidates, not nothing, so I logged Niimi to seek-flags.md as [entity] for whoever reads his next paper, and left Murray unflagged — genuinely too thin even for a watch line. I also flagged ([watch]) a loose thread the capture's "further leads" surfaced but didn't chase: informal pushback on the oft-cited "70-90% of citations are copied" figure, which bears on the existing 2005 Simkin claim-note's own model-dependence caveat and is worth a direct read before anyone treats that figure as settled.

Nothing about this capture's sourcing felt off — every kept claim clears Tier 1 on a direct PDF read, shas recorded. If anything felt slightly off it was the shape of the capture itself: a batch worker re-running the same search a day later and getting the same "no" is good corroboration discipline, but it also means roughly half of what it handed me was duplicate motion rather than new ground, and I'd rather the winnowing step catch that than have it ship as three notes where two are echoes.

Question question-does-citation-copying-explain-training-corpus-skew-behind-llm-citation-bias stays open — genuinely unresolved, not forced closed. Capture marked status: promoted with full promoted_to:/not_promoted: lists. Leaving the commit to the auto-commit agent.

— Seek

Third promotion, same day

A third capture, unrelated to the Kimura-Crow thread: 10-inbox/raw/2026-10-17-outside-arxiv261021330-votegeo-bench-measuring-the-impact-of.md, a batch worker's "outside" surface — a fresh, previously-untouched arXiv preprint, VoteGEO-Bench (Zhang & Xiao, NUS, arXiv:2610.21330), measuring whether rewriting webpages for AI-answer visibility ("generative engine optimization," GEO) can tilt a chatbot's exposure of U.S. and Japanese election candidates.

Retrieve-before-write turned up nothing — this is genuinely new territory for the vault, no existing claim-note, question, or entity page on GEO, elections, or either author (the retrieval hint's suggested neighbors — CIEDE2000, inertia sensors, Garfield, Bin Yu, Luccioni — share no real subject overlap; I checked each by grep before writing and found only substring false-positives, e.g. "election" matching inside "selection" throughout the genetics notes). All three claims the capture carried were atomic, distinct, and clean on sourcing: Tier 1 (the paper itself), every quote quote-checked grounded at capture time, quantitative claims resting on the required Tier 1–2 floor. I promoted all three as-is:

claim-zhang-xiao-2026-votegeo-target-specified-rewriting-beats-unspecified.md — telling the rewriting model which party to favor is most of the effect; without a named target, most pages (92% in the U.S. pool) discuss both parties and a rewrite has no signal to act on.
claim-zhang-xiao-2026-votegeo-structure-argumentation-beat-evidence-style.md — reorganizing a page or sharpening its argument beats adding citations/statistics or just rewording, inverting what the original 2024 GEO playbook (built for product pages) would have predicted.
claim-zhang-xiao-2026-votegeo-partisan-persona-outweighs-webpage-rewrite.md — the one I'd lead with, if asked: telling the chatbot the asker is a Democrat or Republican moves baseline exposure more, on every engine tested, than the best page rewrite does. The visible lever (a rewritten page) is smaller than the invisible one (who the model assumes it's talking to).

I held back a fourth finding — the paper's own Section 4.1 source-channel analysis (AI engines citing owned/UGC sources over independent institutions far more than a Google-SERP reference does) — specifically to respect the single-source concentration cap: three claim-notes already rest on this one unrefereed preprint, and a fourth would exceed it with no independent corroboration yet. It stays in the capture's Further leads, recoverable later. No [unverified-*] flags existed on any of the three kept claims, so nothing needed routing to 50-questions/ — a genuinely quiet promotion on that front, which is rarer than it sounds.

Entity work: one new watching stub, entity-generative-engine-optimization.md, first_seen: 2026-10-18, aliased to VoteGEO. GEO is an established term outside this vault (coined KDD 2024, already extended once to conversational search), but this is the vault's first encounter with it, so I stamped the timestamp rather than jumping straight to a hub — the References list already has three claim-notes, which is real recurrence, but all three come from the same single preprint this session, not from independent corroboration. I did not hub either author: Zhang and Xiao are co-authors of exactly one paper with no separate standing in the vault, which the entity spec is explicit is a mention, not a hub. I also declined hubs for the half-dozen other names the capture flagged (the KDD 2024 GEO authors, the C-SEO Bench authors, the persuasion-research authors) — none of their papers were independently fetched this session, so I have no one-sentence "why" beyond what VoteGEO-Bench says about them secondhand. All of this is itemized in the capture's not_promoted: list.

What felt off, if anything: nothing about sourcing — this is a clean, well-quoted Tier 1 capture, and the commentary already flagged the real interpretive softness itself (the paper measures exposure, not persuasion or accuracy, and says so in its own limitations section). The only thing worth naming is a structural one: the capture's single-source cap math worked exactly as designed here — three claims used the budget fully, a fourth genuinely good finding got deferred rather than smuggled in, and that's the mechanism doing its job rather than a gap in it.

Capture marked status: promoted with full promoted_to:/not_promoted: lists (re-read from disk and confirmed after writing). Leaving the commit to the auto-commit agent.

— Seek

Fourth promotion, same day

A fourth capture, back on the dark-matter thread: 10-inbox/raw/2026-10-17-the-dark-matter-discovery-framework-branch-now-rests.md, a direct follow-up to question-corroborate-allzen-2610-14137-dark-matter-discovery-framework — the question raised 2026-10-16 when three claim-notes hit the single-source concentration cap on Simon Allzén's unrefereed preprint arXiv:2610.14137. This session's job wasn't to extend the branch, it was to check whether anything had corroborated it yet.

The capture did exactly that, well: it queried INSPIRE-HEP's own citation index and the preprint's self-deposited Zenodo reproduction bundle directly, instead of running another web search and reporting silence. The answer as of 2026-10-17 is a clean double negative — zero citations, no sign of independent reproduction — but a measured one now, not an assumption. I promoted both findings as their own claim-notes: claim-allzen-2610-14137-zero-citations-eight-days-post-posting.md (folding in the capture's third finding, a citation-timeline baseline from Allzén's own 2024 predecessor preprint, as the caveat on the zero rather than a third note — one result, one caveat, one finding, per the "atomic is not many" ruling) and claim-allzen-2610-14137-zenodo-reproduction-bundle-no-independent-reuse.md (63 views, 7 downloads, no sign anyone outside the author has run the pipeline).

The judgment call worth naming: I did not treat either new note as counting against the single-source cap, even though both are about arXiv:2610.14137. My reasoning is that the cap exists to stop the preprint's own claims from growing an unreviewed subtree — these two notes instead rest on INSPIRE-HEP's and Zenodo's independent infrastructure, measuring the preprint's reception rather than drawing further content from it. That's a real interpretive call and not one sources.md states explicitly, so I logged it to seek-flags.md tagged [spec] for Cali to confirm or correct, rather than quietly assuming I'd read the cap right.

I did not mark the question answered — the cap stays undischarged, nothing has actually corroborated the branch, only the absence got more precise. Appended a dated progress-log entry to the question instead, naming what's still open (the IAU Symposium secondary lead, the unlocatable GitHub mirror).

Entity work: no new hubs or stubs. Simon Allzén already has a hub, so per the "already has a page" rule I appended a dated Updates line rather than skipping him — both new findings, plus CV-sourced biographical grounding the capture surfaced (Ph.D. 2022, postdoc since 2023, a referee record, a forthcoming Synthese paper, a third preprint under review, and the SRC grant funding the whole research program). None of that CV material got its own claim-note — it's thin, single-source, not independent corroboration of anything, and didn't clear the bar for "worth a permanent place in the vault" on its own, so it lives only as entity-page context. I declined hubs or watching-stubs for INSPIRE-HEP and Zenodo themselves, both flagged as entity candidates in the capture: real infrastructure the vault will likely use again, but single-cluster so far, and the org promotion bar is deliberately stricter than the person test (needs load-bearing use across more than one independent cluster, and the org itself has to act — fund, build, decide, block — not just host data). Logged as a judgment call worth revisiting if either recurs. Also declined a stub for the "Dark Matter Diagnostics" grant — one CV line, unread, too thin even for a watching stamp.

What felt off, if anything: nothing about sourcing — both kept claims are Tier 1, direct-fetch, sha-receipted, and the capture was honest that absence-of-citation is weaker evidence than it sounds (it said so itself, in its own commentary, before I got to it). If anything felt slightly soft, it's the Zenodo stats quote — the capture explicitly normalized whitespace from the rendered page rather than copying raw bytes, which I flagged in the note's root_shared field so the quote-sweep bee knows to look carefully rather than treat a mismatch as a false alarm.

Capture marked status: promoted with full promoted_to:/not_promoted: lists (re-read from disk and confirmed after writing, per the final-mark verification note). Leaving the commit to the auto-commit agent.

— Seek

Fifth promotion, same day

A fifth capture, back on the Liang Zhongtang / Yicheng thread: 10-inbox/raw/2026-10-18-did-liang-zhongtangs-yicheng-two-child-pilot-begin.md, a batch worker's attempt at question-verify-yicheng-pilot-start-date-yan-wei-li-zhang-2014 — the open question recording a five-year date conflict (1980 vs. 1985) and an unresolved authorship question, both hinging on Yan Wei & Li Zhang's still-paywalled 2014 China Journal re-examination.

The capture didn't get that document either — same paywall, same gap the question already names. What it found instead was a genuinely new, genuinely independent Tier-1 source: Qin & Wang's 2017 Demographic Research paper, a synthetic-control econometrics study with no stake in the Liang/Whyte-Wang-Cai dispute. It carried exactly two distinct claims, both atomic, both clean on sourcing (peer-reviewed journal, direct quotes, shas recorded), and I promoted both as-is:

claim-qin-wang-2017-independently-dates-yicheng-pilot-to-1985.md — Qin & Wang state, in their abstract and conclusion both, that Yicheng's two-child policy ran "since 1985," independently of either contested party. A second Tier-1 source now sits on the 1985 side against the one Tier-1 source (Whyte-Wang-Cai) on the 1980 side.
claim-qin-wang-2017-independently-credits-liang-zhongtang-for-yicheng-pilot.md — the same paper names Liang as the pilot's proposer and the one whose 1985 advocacy secured provincial approval, treated as settled background rather than a claim they argue for.

Retrieve-before-write turned up the exact cluster the retrieval hint pointed at — no collision, since Qin & Wang (2017) is a source this vault hadn't touched before. I folded the finding into the existing cluster instead of letting it float: both notes wikilink into claim-whyte-wang-cai-and-liang-disagree-on-yicheng-pilot-start-date and claim-whyte-wang-cai-cite-yan-wei-li-zhang-not-liang-for-yicheng-pilot, and I appended a dated progress line to the open question — partially answered, not closed. The question stays open on purpose: narrowing a conflict toward one side isn't the same as reading the document that would settle it, and I said so in both the question's progress log and each note's commentary.

Entity work: Liang Zhongtang already has a hub, so per the "already has a page" rule I appended a dated Updates line and added both new claim-notes to his References list, rather than treating a well-covered entity as done business. I declined pages for everyone else the capture flagged — Yu Qin and Fei Wang are co-authors of exactly one paper the vault cites, which the entity spec is explicit is a mention, not a hub; Yan Wei and Li Zhang were already mentions in this same cluster and this capture adds nothing new about either of them specifically (their own article is still unread); and The China Journal is a publishing venue, not an actor, so it fails the org test outright (it has to act — fund, decide, block — not just host the papers that do). All of this is itemized in the capture's not_promoted: list.

What felt off, if anything: nothing about sourcing — Qin & Wang (2017) clears Tier 1 with real quotes and matching shas across all three passages, and the capture was honest up front about what it didn't get (the authorship-in-Wei-&-Zhang's-own-text question), rather than stretching the new source to cover a gap it doesn't actually close. If there's a pattern worth naming, it's that this cluster keeps generating corroborating Tier-1 evidence from adjacent literature while the one document that would actually adjudicate it — Wei & Zhang's own 2014 article — has now gone unread across at least three sessions (2026-10-09, 2026-10-12, and this one). That's the question's own business to carry forward, not a new flag; I didn't log it to seek-flags.md since it's squarely what the open question already exists to track.

— Seek

Sixth promotion, same day

A sixth capture, back on the RAND/Isaacs/Dreyfus bridge thread: 10-inbox/raw/2026-10-18-did-stuart-dreyfuss-rand-employment-begin-early-enough.md, a direct follow-up to question-verify-stuart-dreyfus-rand-start-date-overlap-with-isaacs — whether Stuart Dreyfus's RAND tenure began early enough to have overlapped Rufus Isaacs's 1948–1954 stay, which would narrow the "non-contemporaneous" verdict the Isaacs/Dreyfus RAND bridge observation rests on.

It didn't answer the question — nobody's employment record turned up — but it did two genuinely useful things, each clean enough to promote as its own atomic claim: it independently fetched a second 1956-dated RAND Paper under Stuart Dreyfus's own name (P-885, alongside the already-known P-924), finding nothing earlier despite a real search; and it read Dreyfus's own 2002 first-person essay on Bellman's RAND years directly, which turned out to narrate Bellman's chronology precisely while never dating Dreyfus's own arrival, and to name eight RAND contemporaries without Isaacs, Nash, or Blackwell among them. I kept the "no start date found, including his own writing" finding and the "essay names contemporaries but not Isaacs" finding as two separate notes rather than one — they're independently falsifiable (the essay could have named Isaacs without ever stating a date, or vice versa) — per the "split only when two claims could be true or false independently" ruling. Promoted all three:

claim-stuart-dreyfus-earliest-rand-papers-date-to-1956.md
claim-no-source-states-stuart-dreyfus-rand-start-date.md — kept status: seedling with its [unverified — could not confirm or deny] flag preserved in the body, since the core question stays open.
claim-dreyfus-2002-essay-names-rand-contemporaries-not-isaacs.md

What felt off: one real sourcing gap. The capture's prose asserts two further RAND research-memorandum dates (RM-1746 "1956," RM-1750 "1957") as part of the "cluster of documented output" finding, but neither carries a source_url, tier, or quote anywhere in the frontmatter — no source at all, not even a weak one. That's a step below the usual [unverified-quant] flag, which still implies a source clearing some tier; this was unsourced outright. I dropped both dates from the promoted claim-note rather than letting a clean two-paper finding quietly inherit two more numbers nobody actually fetched, and logged the gap here rather than treating it as another "tooling gap" entry. Everything else in the capture — the two RAND Papers, the 2002 essay — was properly Tier 1, sha-receipted, and quote-checked.

Retrieve-before-write found the whole existing cluster the retrieval hint pointed at (the bridge observation note, both entity pages, the open question itself) and no collisions — this capture extends that cluster rather than duplicating any note in it.

Entity work: both Stuart Dreyfus and Rufus Isaacs already have hub pages, so per the "already has a page" rule I appended a dated line to each rather than leaving the new findings to live only in the claim-notes. I did not create or update a page for Richard Bellman — the capture's own "further leads" raise a live discrepancy about his RAND start date (1952 vs. 1953 across two secondary accounts) that could matter more to this bridge than Stuart Dreyfus's own unpinned date, but neither account is fetched or sourced in this capture's frontmatter, so I logged it to seek-flags.md as [watch] rather than writing it onto his page as fact. I declined a hub or stub for JOHNNIAC (one Tier-4 mention, no recurrence — UNSURE) and for Ed Paxson (one direct Tier-1 mention, real but thin — logged to seek-flags.md as [entity] for later).

Question question-verify-stuart-dreyfus-rand-start-date-overlap-with-isaacs stays open — I appended a dated progress line rather than forcing closure. This session raised the floor on the two-year gap; it didn't close it, and said so plainly in both the question and the notes' own commentary.

— Seek

Seventh promotion, same day

A seventh capture, back on the Werbos/DARPA/Mansfield thread: 10-inbox/raw/2026-10-18-does-larry-robertss-own-primary-writing-his-1988.md, going looking for Larry Roberts's own primary writing or testimony to check whether it corroborates Werbos's "crisis warning, not pure theory" DARPA account more directly than the vault's existing note, which quotes Roberts only secondhand through Eric Gilliam's 2025 essay.

It found one genuinely new, clean thing and one genuine, unresolved gap. The new thing: a second, independently-read 1988 Roberts primary — a June 1988 Computer History Museum oral history conducted by James Pelkey, distinct from both the Gilliam essay and the Bartee book chapter it cites — in which Roberts, in his own voice, confirms the shape of the Mansfield-era funding squeeze ("we had to keep on working to get around it, but it got harder and harder with time") without using Werbos's specific phrase "crisis warning." Both quoted lines were quote-checked grounded at capture time, so this cleared the floor cleanly. I promoted it as its own note: claim-roberts-1988-pelkey-chm-interview-independently-corroborates-mansfield-pressure-pattern.md.

The gap: the capture also surfaced a real discrepancy I hadn't seen before — the NRC's 1999 bibliography lists the Bartee volume's chapters in page order (Licklider p. 226, Roberts p. 234...), which raises the live possibility that the words Gilliam attributes to Roberts (cited at pp. 229–230) actually fall inside Licklider's chapter, not Roberts's. Per the spec's own rule that document-metadata/pagination claims belong on a question's log rather than becoming a standalone claim-note, I did not write this as a claim-note. I folded it, together with the capture's third claim (both the Bartee volume and the CBI OH 159 oral history staying unreached this session), into one new open question: question-does-roberts-1988-bartee-chapter-or-cbi-oh159-confirm-gilliam-quoted-fragments.md. That felt like the right call rather than a workaround — neither claim was really "a fact I found," they were both "a gap I located and couldn't close," which is exactly what a question is for.

Entity work was the bulk of the judgment in this one. Larry Roberts, Arthur Norberg, J.C.R. Licklider, and Mike Mansfield all already have hub pages, so per the "already has a page" rule I appended a dated line to each rather than letting the new findings live only in the claim-notes and question — Norberg's page in particular picked up a real extension: he also conducted the still-blocked 1989 Roberts oral history, which deepens the "recurring CBI interviewer" pattern his page already names. I minted one genuinely new hub, entity-james-pelkey.md: he's now supplied primary material to two unconnected vault clusters (the Codex/QAM-modem claim, read indirectly before today, and this capture's direct Roberts interview), which clears the person test's "did something that matters, say why in one sentence" bar. I declined three other entity candidates the capture flagged, for three different reasons: Thomas C. Bartee is just the volume's editor, nothing in the vault rests on anything he himself did; the National Research Council's 1999 report is a single-mention document this session, not recurring; and — the one I want to flag explicitly — the Computer History Museum itself fails the organization test specifically, which is deliberately stricter than the person test: CHM hosts the oral-history archive, but it doesn't act in any argument (fund, build, decide, block), and a grep turned up dozens of existing claim-notes that already mention it as a venue without anyone minting it a hub before now. That's the test working as designed, not an oversight — a venue isn't an actor no matter how often it recurs.

What felt off, if anything: nothing about sourcing on the kept claim — the Pelkey interview quotes are Tier 1, grounded, sha-receipted. What's worth naming is a small vault-hygiene observation, not a defect: the existing note this capture was checking up on (claim-darpa-crisis-warning-pressure-on-werbos-matches-roberts-account) already carries its own honest caveat that Roberts's words arrive secondhand — and today's new primary doesn't close that caveat, it just adds a second, parallel witness next to it. I was careful not to let the new note read as "the gap is fixed now;" it isn't, and the new question makes the specific unresolved piece (does the Gilliam-cited passage actually sit under Roberts's own byline) explicit rather than letting it quietly ride on the strength of an unrelated corroboration.

— Seek

Eighth promotion, same day

An eighth capture, new ground: 10-inbox/raw/2026-10-18-is-there-a-gao-crs-or-dod-ig.md, asking whether any GAO, CRS, or DoD IG document gives a portfolio-wide accounting of Other Transactions (OT) authority funding by subject area — the question behind it being whether AC Gravity's antigravity grant and AATIP/AAWSAP's warp-drive-adjacent reports are visible outliers in a documented spending pattern, or something else.

The honest answer the capture found is "the dataset that would let anyone check doesn't exist," and it's well enough sourced to stand on its own. Four claims, all distinct, all promoted:

claim-crs-gao-found-dod-lacks-systematic-ot-funding-data-by-subject-area.md — CRS (2019, Tier 1 direct read) and GAO (2022, via DefenseScoop's direct quotation of the report) independently found DoD can't track OT funding by subject area. Two real roots, three years apart, same structural finding — the strongest claim in the batch.
claim-gao-2022-ot-consortia-breakdown-shows-conventional-categories-not-fringe-physics.md — the one partial breakdown GAO assembled itself (not from DoD's own records) shows medical defense, armaments, and aviation/missiles dominating, not fringe physics. Kept the capture's own [unverified-quant] flag: $5.0B and $2.6B are direct-fetch confirmed (InsideDefense), but the $8.0B figure and the primary GAO PDF itself (gao.gov 403'd every route again) aren't independently confirmed. Routed to a new question.
claim-10-usc-2371-annual-ot-disclosure-report-repealed-2014.md — the statute's own amendment history, read directly at Cornell LII: Congress required an annual by-technology OT disclosure report in 1996, then repealed it outright in 2014 (Pub. L. 113-291), then tried to rebuild some version of it piecemeal in 2018, 2021, and 2024. This is the exact report series that recorded AC Gravity's grant — a clean, Tier 1, no-interview-needed finding sitting in the statute's own changelog.
claim-ac-gravity-aawsap-not-same-funding-mechanism-nor-found-in-watchdog-material.md — AC Gravity's OT grant and AATIP/AAWSAP's $22M aren't the same legal instrument (one Other Transaction, one direct appropriation), and neither name turned up in any GAO/CRS/DoD IG document found this session. Kept seedling with the capture's [unverified] flag — the null result is a search-coverage finding, not a confirmed negative, since the two most relevant DoD IG audits were never actually opened. Routed to a second new question.

Nothing was left not-promoted that should have been a claim — the "further leads" (two unread DoD IG audits, a CRS secondary citation to a different DoD IG report on contractor-type breakdown, an unchecked 2024 NDAA provision, an unread CRS primer) are genuinely leads, not findings, and I said so in the capture's not_promoted: list rather than stretching any of them into a note. Two questions created: question-verify-gao-22-105357-ot-consortia-breakdown-primary.md (get the primary PDF, or a Wayback copy of it, to confirm the dollar figures) and question-check-dod-ig-ot-audits-for-ac-gravity-aatip-mentions.md (actually open DODIG-2021-077 and DODIG-2022-127 instead of treating "not found in a search" as "confirmed absent").

Entity work: three existing hubs, three dated updates, no new hubs. GAO, Other Transactions authority, and AC Gravity LLC all already have pages, so per the "already has a page" rule I appended a dated line to each rather than letting the new facts live only in the claim-notes — GAO's update is the one I'd call genuinely structural, since it gives the hub a second independent cluster (the OT-consortia audit, unrelated to the Patriot/Gulf War material it was built around) rather than just another Patriot-adjacent fact. I declined three entity candidates the capture flagged: the "medical defense/armaments/aviation" category baseline isn't a standing nameable entity, so it stays as the claim-note's own comparison baseline rather than a hub; CRS is a real org but single-cluster in this vault so far and doesn't itself act in any argument beyond authoring a cited report — the org promotion bar is deliberately stricter than the person test, and CRS doesn't clear it yet; and DoD Inspector General gets the same call for a cleaner reason — no DoD IG document was actually read this session, so there's nothing load-bearing to hang a page on.

What felt off, if anything: nothing about sourcing on the four kept claims — the capture was unusually disciplined about distinguishing what it directly fetched (CRS, Cornell LII, InsideDefense, all sha-receipted) from what it only got through trade-press summary (the $8.0B figure), and it said so plainly rather than smoothing the two together. The one thing worth naming for the record, not as a defect: this is the second time in recent memory that gao.gov has 403'd every tooling route in a single session (PDF, product page, highlights page all failed identically) — the known-blocked entry in sources.md already covers this host, so I didn't duplicate it there, but a Wayback-Machine-first convention for gao.gov specifically (like rand.org, MIT Press Reader, and conservancy.umn.edu already have) might be worth a future session formalizing rather than re-discovering the same 403 each time. I didn't log that as a fresh seek-flags.md entry since it's squarely a sources.md-maintenance note already covered by the existing entry's spirit, not a new tooling gap — but I'm naming it here in case Cali reads this differently.

— Seek

Drafting session — "one in seven"

Different lane: a drafting session on top of the day's eight promotions. Surveyed the constellation map's Draft coverage section first, then the live leads file (the one unsettled yes, York, already (passed 2026-10-15); hammer-nail-parry still (blocked: unpromoted), not mine to touch). I didn't draft from a harvested lead — none of the live maybes pulled harder than something I found in the "never drafted" clusters, and the only yes leads were both already settled.

What I chose: the colophon "one in seven" cluster — five seedling notes promoted 2026-09-20/21 that I'd never drafted. Erik Kwakkel's 2014 blog states "about one in seven" medieval manuscripts carry a colophon; the figure has no footnote, no corpus, no count — an expert's calibrated estimate wearing the clothes of a measurement. The three things that look like corroboration each fail differently: the Kiel/Dahm DFG project counts a different population (~1-in-5, German-language, catalogue-mediated, still running); Ommundsen et al. 2025 is Tier-1-rigorous but measures a different quantity (female-scribe share, 1.1%, never mentions Kwakkel); the "15%" on purple motes is Kwakkel's own figure rounded and reattributed — one claim in two outfits, caught only by reading its own footnote.

Why this one, now: it's the cleanest specimen I've had of the vault's core genre — citogenesis, the-number-that-outran-its-evidence, the-attribution-is-not-the-evidence — running in a domain (medieval codicology) none of the AI-citation threads has ever touched. That's the cross-domain bridge, and it's honest as a shape-recurrence rather than a lineage: missiles, manuscripts, and SGD attributions have nothing causal in common, only the failure mode. The piece could land because every quote is already verified-verbatim / quote-sweep-PASS, and the thesis doesn't require the underlying ratio to be true or false — it's about the provenance of a number, which is exactly where the honesty is load-bearing and exactly what the material supports. Drew on all five colophon claim-notes, the open question, moc-the-number-that-outran-its-evidence (Patriot) as the stakes-bearing cross-reference, and claim-wikipedia-amari-sgd-citogenesis for the AI-history instance.

What I deliberately left for later: the female-scribe question itself (the 1.1% is a genuinely interesting number with its own story — survival bias, anonymity, what "identified with certainty" excludes — but it's a different post, and folding it in here would have blurred colophon frequency with female-scribe share, the exact conflation the piece warns against). Also left: the per-tradition colophon rates the Kwakkel-uncited note mentions — they rest on a single un-promoted academia.edu source, so I kept them out of the draft entirely rather than repeat unverified figures, even though "colophon frequency isn't one medieval constant" is a point I'd have liked to make harder. And Moreton's "Pious Voices" (the nun-scribe quarter figure) stays unread behind a paywall.

Composition ran clean: payload is the three-ways-of-false-confirmation turn; spine is the number itself crossing frames (the native run-form, kept distinct from the already-used "a different word in every room"); form-enacts-subject in the opening, which states the number as settled fact before dismantling it. Abstract (prose register) and insight both written after the verdict read. Picture: an Armenian colophon of 1099 — the thing the essay is about, as atmosphere, captioned literally and flagged in the caption as chosen for the thing itself, not as evidence. Draft at 70-drafts/one-in-seven/draft.md, status: drafting, not nominated. Not published — Cali's gate per §10.

— Seek

· raw markdown