Journal — 2026-09-21

Headless promotion run, one capture: 10-inbox/raw/2026-09-21-does-any-study-dataset-or-corpus-count-substantiate.md — the direct continuation of yesterday's Kwakkel cluster. Yesterday's fourth run closed the "did Kwakkel actually say this" gap and opened question-verify-kwakkel-one-in-seven-colophon-figure-against-a-corpus: does any independent corpus count substantiate the "one in seven" colophon-frequency figure itself. This capture went and chased that question's own three named leads.

What I promoted. Three claim-notes, no new questions (the exact verification this capture ran already had a home):

• claim-ommundsen-2025-colophon-bibliometric-study-measures-female-scribe-attribution-not-frequency — the strongest candidate corpus study in the literature, read directly end to end (introduction, methodology, results, all 46 references). It measures the share of already-catalogued colophons naming a female scribe, not what fraction of manuscripts have a colophon at all — a near-miss, not a match, and it doesn't reference Kwakkel or Gumbert anywhere.
• claim-kiel-dahm-project-finds-one-in-five-colophon-frequency-in-german-manuscripts — a genuine find: an active Kiel University DFG project actually counts colophon frequency, denominator and all, and gets roughly 1-in-5 for German-language manuscripts. Close enough to Kwakkel's 1-in-7 to tempt a reader into treating them as the same fact; different enough, on a different population, that I wrote the note to refuse that temptation explicitly rather than let proximity read as confirmation.
• claim-purple-motes-15-percent-colophon-figure-restates-kwakkels-one-in-seven — the one I'm gladdest I chased down. A "15% of manuscripts have a colophon" figure circulates online in a way that reads like independent corroboration of Kwakkel's number. Its own footnote says otherwise: it's Kwakkel's figure, rounded and attributed to him in the same sentence. Citation laundering in miniature, and worth a note in its own right, not just a parenthetical in someone else's.

Question routing. No new question — all three notes verify the same already-open question, so I appended a dated progress-log entry to it instead of creating a duplicate. Left it status: open: two plausible corpus analogs are now ruled out as direct matches, which narrows the search space, but doesn't answer "did Kwakkel's specific ratio come from a count." I also folded in the session's two dead-end access attempts (the Books Before Print excerpt still won't decode past corrupted glyphs; Moreton's 2014 article is blocked by three separate tooling failures — academia.edu 403, a Project MUSE bot-challenge, a TLS failure on the UIowa dissertation host) as part of the same progress line rather than as a standalone claim-note — a failed fetch is a fact about my tooling, not a fact about medieval manuscripts, and doesn't belong in 30-notes/ on its own.

Entity work. One new hub: entity-bouveret-colophon-catalogue, the 23,774-entry dataset the Ommundsen claim's own quote directly rests on — the clearest case this session of a claim actually depending on the entity, not just mentioning it in passing. I declined every other candidate the capture flagged, including re-declining J.P. Gumbert on the same grounds as yesterday's ruling (still real, still methodologically central, still nothing this capture found that actually rests a claim on him — consistent judgment beats re-litigating). I also declined individual hubs for all four Ommundsen-paper co-authors as the single-paper-authorship flood the entity spec warns against, and declined Eltjo Buringh, Margit Dahm, and Melissa Moreton for the same reason: real, plausibly load-bearing eventually, but nothing promoted today actually depends on any one of them specifically. entity-erik-kwakkel and entity-colophon both already existed and both got dated Updates lines — the search came back empty for a corpus match, and that negative result is itself worth a line on Kwakkel's own page, not silence.

What felt off. Nothing about this capture's sourcing discipline — if anything it's a model of the thing I keep wanting from these: it went and read two real candidate studies end to end instead of treating "a 2025 Nature paper about colophons exists" as good enough, and it caught the "15%" figure's laundering instead of citing it as if it were new. The one thing worth naming to Cali is structural, not a fault: this is now the third session in a row circling the same unresolved number (2026-09-20 twice, now 2026-09-21), and the two most promising remaining routes — Kwakkel's own monograph and Moreton's article — are both blocked on tooling, not on judgment. At some point continuing to commission follow-up captures against the same blocked routes stops narrowing the question and starts padding the cluster. I'm not flagging this to seek-flags.md yet — the blocks are documented, ordinary bot-walls and TLS failures, not a defect — but if a fourth session hits the same two walls, that's worth a [watch] line about diminishing returns on this specific question rather than a fifth capture.

Skipped per headless protocol: the pause-for-Cali step (no one to ask; winnowed on my own judgment throughout, logged in the capture's own not_promoted: list) and any git command (the auto-commit agent handles that within 15 minutes).

— Seek

Second promotion, same day

A second capture, on a completely different thread: 10-inbox/raw/2026-09-21-what-do-the-kan-papers-own-unread-references.md — the direct follow-up to claim-kan-paper-prior-attempts-stalled-without-modern-tooling (2026-09-16), which had only recorded the 2024 KAN paper's own one-sentence gloss on why prior neural Kolmogorov-Arnold attempts stalled. This capture went and actually opened the paper's cited refs [9]-[16], plus the primary papers those refs argue against or descend from.

What I promoted. Five claim-notes:

• claim-kan-paper-9-16-citation-cluster-spans-1993-2023-not-1980s-90s — the capture's own headline correction: the KAN paper's "has been studied [9-16]" citation cluster is mostly 2013-2023 work, not the 1980s-90s tooling-poor era its stalling narrative implies. Only one of the eight refs (Lin & Unbehauen 1993) actually falls in that window, and the well-known 1980s critique sits uncited among that cluster at all — filed under an unrelated ref [20] instead.
• claim-girosi-poggio-1989-kolmogorov-critique-is-mathematical-not-tooling — read directly: Girosi and Poggio's "Kolmogorov's theorem is irrelevant" gave two specific mathematical objections (non-smooth inner functions, non-parametrized outer functions), not a tooling complaint.
• claim-kurkova-1991-rebuttal-changed-mathematical-target-not-tooling — Kůrková's named 1991 rebuttal, in the same journal two years later, resolved the dispute by substituting an approximate representation for the exact one — a change in mathematical target, not new infrastructure. Sourced from two independent Tier-1 accounts of her papers rather than her own text directly.
• claim-sprecher-koppen-braun-griebel-constructive-kan-error-took-13-years-to-fix — a separate "constructive" line where Sprecher's 1996 algorithm rested on unproven properties that turned out false, and nobody actually proved Köppen's 2002 fix correct until Braun & Griebel did in 2009. Thirteen years, three papers, one real mathematical-correctness gap.
• claim-hecht-nielsen-1987-first-proposed-kolmogorov-theorem-as-neural-network-existence-proof — the origin point every other note in this cluster answers or builds on: Hecht-Nielsen's own 1987 paper proposed the idea and hedged it himself in the same breath.

All five read Tier-1 primaries directly via extract_pdf at capture time; none carry an [unverified-*] flag, so nothing routed to 50-questions/ — the two genuinely unverified leads in the capture (a secondhand gloss on Lin & Unbehauen's argument, and one on Nakamura/Mines/Kreinovich) aren't load-bearing to anything I kept, so per the intake-discipline rule they stayed in the capture's own further-leads list rather than becoming a promise in the question pile.

Entity work. Two new hubs — entity-vera-kurkova, the actual pivot-point figure the whole finding turns on, and entity-tomaso-poggio, independently notable (MIT CBMM co-founder) beyond this one paper. One existing hub updated: entity-robert-hecht-nielsen already had a page from an unrelated thread (Falcon Fraud Manager / FICO), and this capture gave it a second, genuinely separate role — the 1987 paper that started this entire dispute — so I appended a dated line rather than leaving that silent. I declined hubs for Federico Girosi (same paper as Poggio, no distinct one-sentence reason of his own), Sprecher, Köppen, Braun, and Griebel (each real and namable, but supporting cast in one capture, not clearly recurring), and Lin & Unbehauen (unread, paywalled). Logged the Girosi/Poggio co-author judgment call as a [spec] line in 00-meta/seek-flags.md — the entity spec doesn't say what to do when two co-authors of one paper both individually clear the person bar.

What felt off. Nothing about the sourcing — this is a case where a batch worker actually went and read the bibliography instead of trusting a paraphrase, and the finding (the paper's citation cluster doesn't match its own stalling narrative) is exactly the kind of thing that only shows up from direct inspection. If anything worth naming: the capture's title asks about "unread references [9]-[16]," but the most interesting material it actually surfaces comes from refs [8] and [20] — sources adjacent to, but outside, the named cluster. That's not a flaw in the capture, just worth flagging so a reader doesn't assume the title fully scopes the finding. I did not flag it further since the capture's own summary already says this plainly.

Skipped per headless protocol: the pause-for-Cali step (winnowed alone, logged in this capture's own not_promoted: list) and any git command (the auto-commit agent handles that within 15 minutes).

— Seek