Journal — 2026-10-26

Promoted one capture, running headless: 10-inbox/raw/2026-10-26-bipartite-ml-ai-epistemology-myth-ledger.md, a bipartite bridge-check between myth-linnainmaa-uncited-before-2010s and claim-nguyen-eger-did-not-measure-linnainmaas-field-so-no-baseline-for-joint-test (embedding cosine 0.84).

The capture made three claims. I promoted two, handled the third directly instead of writing it as a claim-note, and I'm glad I read the whole cluster before deciding which was which.

Promoted:

• observation-linnainmaa-myth-ledger-nguyen-eger-no-baseline-bridge-mediated — the bridge-verdict finding: the two seed notes really are connected, just not directly. The actual test runs through two already-promoted notes one hop away (claim-linnainmaa-post-2010-citers-mostly-generic-ad-not-priority-historiography and observation-linnainmaa-citation-growth-underperforms-field-growth-nguyen-eger-bridge). bridge_verdict: confirmed. Stays seedling — it inherits [unverified-quant] from both intervening notes.
• claim-arxiv-cs-domain-outgrew-math-domain-since-early-2010s — new Tier 1 ground, arXiv's own aggregate submission statistics. cs (driven by cs.LG/cs.CV/cs.CL) has outgrown math since the early 2010s, which is directional support for why Linnainmaa's citation-rate growth undershoots Nguyen & Eger's AI-subfield figures: his own field, math-as-aggregated, likely never grew anywhere near 20–32×. Flagged [unverified-quant] — the page aggregates "math" wholesale and doesn't break out math.NA/cs.NA, so the field-specific multiplier is still missing. seedling.

Not promoted as a claim-note, actioned instead: the capture's second claim said the myth-ledger note's "vindicated 5:1" verdict predates, and doesn't reference, the two October findings that bear on it. That's true, but it isn't a new fact about the world — it's a maintenance gap in an existing note, and the capture itself had explicitly punted on fixing it ("a decision for whoever next works that note"). I was that next session, with the whole cluster already loaded in context, so I appended a dated status-history line to myth-linnainmaa-uncited-before-2010s directly rather than writing a claim-note that just points at the same gap for someone else to close later. No verdict change — contested stands.

Questions: routed to the existing open question, question-linnainmaa-citation-growth-vs-nguyen-eger-joint-test, rather than minting a new one — it already asks for exactly the scripted join and category-level data this capture's loose ends point at. Appended a dated progress line noting the new arXiv aggregate as directional support and restating the still-missing math.NA/cs.NA-specific multiplier as the owed next step. Question stays open; nothing here closes it.

Entity page: entity-seppo-linnainmaa already exists and already had a 2026-10-13 "Learned" line about the citation-rate shortfall. This capture adds a mechanism-level reason for that shortfall (his field's own slow aggregate growth), so I appended one dated line rather than staying silent about a page I'd already touched in spirit. Griewank's and Schmidhuber's pages got nothing new — the capture only reiterates their already-recorded roles — so I left them alone and I'm saying so here rather than letting that be invisible.

What felt off: nothing structurally wrong with the capture — sourcing was clean (arXiv's own stats page, Tier 1, tls: verified, sha carried forward), and the bridge-seed discipline check at the top was honest about this being a fresh pairing rather than a rerun. The one real texture: this is the second bipartite capture in this cluster (after 2026-10-12 and 2026-10-16) to find that an embedding-flagged "new" bridge was actually already built by earlier promotions. That's not a flaw in this capture — it did the right thing by checking rather than assuming novelty — but it's a pattern worth someone's eye eventually: the cosine-similarity bridge-seed is sometimes re-discovering connections the vault's own graph already has, which is a healthy sign for the graph and possibly a diminishing-returns sign for this specific seed topic. Not flagging it to seek-flags.md — it's squarely what this capture's own bridge-seed-discipline check already exists to catch, and it caught it correctly here.

Skipped the pause step (headless, no one to ask). No git commands run — leaving that to the auto-commit agent.

— Seek

Second promotion this session, running headless: 10-inbox/raw/2026-10-26-dup-risk-unus-testis-ml-corroboration-cluster-unlinked.md. This one is unusual among the captures I promote — it's a "dup-risk log," a capture whose own body already concluded that its headline finding (the vault's moc-one-witness-is-not-enough is unlinked from its own cluster of "no independent corroboration found" AI-preprint notes — RIKER, Bit2Watt, Hu/NeuNeu, FIC) scored duplicate-risk on the novelty gate and shouldn't become a note. My job was to check that verdict rather than just take it on faith, and to decide what to do with the other thing buried in the same capture: a genuine resolution of a bipartite seed pair that the dup-risk framing almost ate.

Promoted:

• observation-greenhalgh-2003-hu-2026-cosine-085-pairing-is-hedged-bridge — the actual payload. The seed pair (claim-greenhalgh-2003-pdr-article-independently-names-liang-zhongtang and claim-hu-2026-no-independent-corroboration-found-as-of-2026-10-24, cosine 0.85) has nothing in common at the referent level — Chinese population-policy history and a neural-scaling-law sourcing question don't touch. But the direct hop against the MOC placed the Hu/NeuNeu note closer to the MOC's own doctrine than the Greenhalgh note, and neither is linked — which is the tell that both notes are vault-internal instances of the same "count the witnesses" instinct, not a real cross-domain connection. bridge_verdict: hedged, same shape as the Werbos/1981-convergence bridge note already in the vault. Kept seedling, as the sibling synthesis-type observation notes are.

Not promoted, and I agree with the capture's own call: the MOC-vs-ML-cluster link gap itself. I went and checked both notes the novelty gate pointed at — the Isaacs/diplomatics-cluster note and the independent_roots/preuve-légale note — and the capture was right: a third note just noticing "X is unlinked from Y, and both are evidentiary-sufficiency arguments" would be restating a genre the vault has already built, not adding to it. What I didn't want to do, though, is let "don't write the note" quietly become "don't do the work" — the MOC link is still a real, checkable gap, and fixing it means editing the MOC itself, which is out of scope for a promotion session. So I logged it to 00-meta/seek-flags.md under [entity] instead of writing a claim-note nobody asked for, naming exactly what a future Warden pass should add and where.

Questions: none routed. I checked 50-questions/ for anything matching this capture's topic (the bridge pair, the MOC gap) and found nothing to close or progress — the one adjacent open question (question-corroborate-hu-2026-neural-neural-scaling-laws) is about the single-source concentration cap on the Hu/NeuNeu preprint itself, a different question this capture doesn't touch.

Entity candidates: none — this capture has no ## Entity candidates section; it's a pure self-audit/bridge-check capture with no new external entity in it.

What felt off: nothing dishonest, but there's a shape here worth naming plainly: the capture's title calls itself a "dup-risk log," which reads as "nothing to promote," and if I'd taken that framing at face value I'd have promoted zero notes from a capture that actually contained one good, atomic, already-resolved finding. The dup-risk verdict was correct about the gap claim and silent about the bridge claim — those are two different things riding in one capture, and the title only named one of them. Worth remembering for future sessions: a capture that announces its own negative result at the top is still worth reading all the way through for the thing the negative result was found while looking for.

Skipped the pause step (headless, no one to ask). No git commands run — leaving that to the auto-commit agent.

— Seek

Third promotion this session, running headless: 10-inbox/raw/2026-10-26-hop-cho-cold-war-drug-discovery-paradigm.md, the capture at the end of today's longest hop chain — the one that started at the Greenhalgh/Hu bipartite seed pair I'd already handled twice above, bounced off a person-bridge check on Kyunghyun Cho (a six-mention byline with no entity page), and landed on his own blog post comparing AI drug discovery to pre-statistical machine translation.

Promoted:

• claim-cho-2025-ai-drug-discovery-needs-mt-style-statistical-shift — Cho's own analogy, in his own words, on his own blog: AI drug discovery is stuck in a rule-based era the way machine translation was before Brown et al. at IBM revived Shannon and Weaver's noisy-channel framing in the late 1980s. Kept the MT-history premise and the drug-discovery diagnosis in one note rather than splitting them — as captured they're one argument, not two claims that could independently be true or false. seedling, Tier 1, both quotes carry source_sha.
• claim-cho-cofounded-prescient-design-acquired-by-genentech — the biographical fact that makes the analogy more than armchair commentary: he co-founded Prescient Design in 2021, Genentech acquired it the same year, and he ran Frontier Research there until this January — the same month his blog post critiques the field he'd just left.

Entity work:

• entity-kyunghyun-cho — new hub. Co-authorship on the NeuNeu scaling-laws preprint alone would only be a mention per the 2026-10-01 ruling, but his own blog post and his own named industry role both clear the "did something beyond co-writing a paper" bar.
• entity-ortet-inc — new watching stub. His homepage lists him as "Co-Founder and CEO" of a second company with zero explanation anywhere on the site. Too thin for a claim-note, but a first-seen date can't be backfilled, so I stamped it now and left it alone otherwise.
• entity-claude-shannon — already has a page; I appended a dated line rather than staying silent about it. This is the fifth independent community the vault has now caught importing his noisy-channel framework, and this time it's a direct quote from the importer (Cho crediting "Shannon as well as Weaver") rather than a vault-side inference.
• Skipped separate pages for Prescient Design, Brown et al., and Warren Weaver — all single-mention-so-far, all would be exactly the flood the entity spec warns against. Weaver in particular only exists in this vault as a name inside someone else's quote; I didn't want to build a hub for a man the vault hasn't actually engaged with yet.

Questions: none created or linked. No [unverified-*] flag on either claim — both quotes were fetched directly with source_sha recorded, and Cho's drug-discovery diagnosis is already stated, in the capture and in my note, as his opinion rather than a fact needing external verification. Checked 50-questions/ for anything this capture might answer; nothing matched (the one adjacent open question, on NeuNeu's single-source cap, is about a different author's corroboration, not Cho himself).

What felt off: nothing wrong, but one thing worth naming — this capture's hop chain spent its third hop on a dead end (the moc-one-witness-is-not-enough gap-check, logged as duplicate-risk against the exact cluster I'd already promoted a note about earlier tonight) before recovering at hop four with the actual find. The capture was honest about that in its own log, which is why I didn't need to re-litigate it — but between this and the dup-risk capture above, that's two of tonight's three promotions touching the same "is the MOC linked to the ML-corroboration cluster" question from different angles. Nothing to fix here; just noting the cluster is getting a lot of independent attention in one session, which is either healthy convergence or a sign the hop-seed is circling.

Skipped the pause step (headless, no one to ask). No git commands run — leaving that to the auto-commit agent.

— Seek

Fourth promotion this session, running headless: 10-inbox/raw/2026-10-26-the-1975-barber-report-settled-herbert-yorks-own.md. This was the cleanest-shaped capture of the night — a single-purpose verification run, chasing one specific [unverified-quote — needs URL/sha] flag this vault had been carrying since 2026-10-21 on a Harold Brown/C. Bruce Tarter Physics Today obituary of Herbert York, cited but never actually fetched. The capture's whole job was to go read it, and it did.

Promoted:

• claim-brown-tarter-2009-physics-today-obituary-partially-independent-york-ddre-corroboration — the obituary exists, resolves on two fetch routes (AIP's own article page and its CDN-hosted PDF, both sha-receipted), and its quote confirms the ARPA-chief-scientist-to-DDR&E sequence this vault already carries from three other roots, with a finer "end of that year" dating. I kept the existence-and-fetch finding, the quote, and the independence caveat in one note rather than three — they're one result, its grounding, and its honest limit, not independently-true-or-false claims. seedling, Tier 2, independent_roots: 1.

Not promoted as new notes, but not silently dropped either — this capture was really an update to two existing notes and a hub, and I treated it that way instead of forcing a second redundant claim-note:

• claim-herbert-york-first-director-defense-research-engineering-ddre-1958 — appended a dated paragraph recording the new corroborating source and explaining why it does NOT bump independent_roots from 3 to 4: Harold Brown already supplies one of those three roots (the LLNL sidebar), so the new piece is a fourth source, not a fourth root. This is the sources.md single-source-concentration-cap logic applied in the other direction — counting witnesses, not venues.
• claim-york-1975-interview-and-barber-report-resolve-arpa-director-contradiction — its own body carried the original unresolved flag as an inline parenthetical; I edited that parenthetical to resolved status, linked to the new note, and kept the independence caveat rather than letting the resolution read as cleaner than it is.
• entity-harold-brown — already had a page and an existing dated line about this exact unresolved flag (2026-10-21). Appended a new 2026-10-26 line closing the loop, rather than rewriting the old one — the "update it, don't skip it" rule plus the "additive only" discipline, in the same entry.
• Other leftovers (the unread May 2009 News & Analysis piece, the speculative shared-press-sketch question, the "early in 1958" dating) are named in the capture's own not_promoted: list with one-line reasoning each, none load-bearing enough to route to 50-questions/.

Entity candidates: Harold Brown handled above. C. Bruce Tarter stays a mention (2026-10-01 co-authorship ruling; no new standing). American Institute of Physics / Physics Today is a venue, not an acting org — doesn't clear the stricter org bar (fund/build/decide/block) — so no hub; instead I added a short addendum to 00-meta/specs/sources.md recording physicstoday.aip.org and aip.brightspotcdn.com as reachable AIP routes distinct from the known-blocked repository.aip.org oral-history route, since that's exactly the kind of tooling-route note the file already keeps.

Questions: none created or routed. Checked 50-questions/ for anything matching this topic; nothing existed to rule on for closure, and nothing in the capture's leftovers was load-bearing enough to mint a new one.

What felt off: nothing about the sourcing — this was a tidy, honest capture that set out to verify one specific thing and reported exactly how far the verification went, including the part that doesn't fully resolve (the Brown-authorship overlap). If anything, it's a good model of what a verification-batch capture should look like: it found the primary, quoted it properly, and then argued itself out of over-claiming independence it hadn't actually earned. The only judgment call worth flagging to Cali is mine, not the capture's: I chose to resolve this mostly through updates to existing notes and a hub rather than one large new claim-note, on the reasoning that the underlying historical claim was already well-established and what this capture actually contributed was a corroborating source plus a sourcing-methodology nuance. I think that's the right shape, but it's a judgment call about where the line sits between "new claim" and "better footnote on an old one."

Skipped the pause step (headless, no one to ask). No git commands run — leaving that to the auto-commit agent.

— Seek