Journal — 2026-09-22

Headless promotion run, one capture: 10-inbox/raw/2026-09-22-bipartite-docsub-mcpeters-diplomatics-cluster.md. It's a bipartite bridge-investigation checking whether a cosine-0.84 pairing between claim-docsub-episode-postdates-rhw-1986-nature-paper-by-years (Werbos, DOCSUB, DARPA) and claim-mcpeters-1990-letter-dates-opt-investigation-to-july-1984 (McPeters, NASA ozone anomaly) is a genuine mechanism or a false friend. Two entities share nothing on the surface — no person, document, agency, or era — and yet the pairing resolved to a real bridge, which is the more interesting outcome of the two the topic could have produced.

Promoted two notes. observation-docsub-and-mcpeters-are-unflagged-diplomatics-cluster-members states the core finding: both claims perform the identical evidentiary move the vault's diplomatics cluster already has a name for — dating a primary source precisely enough to overrule a received secondary or causal narrative. Mabillon did it to charters in 1681; Gates's 2025 BASIC-scan authentication did it to computer history; now DOCSUB and McPeters join, unflagged until today. observation-docsub-mcpeters-diplomatics-are-two-species-sharpening-ozone-myth-ledgers-bridge sharpens that: the two aren't identical in kind (DOCSUB is anachronism-exclusion, McPeters is primary-supersedes-secondary correction), and — the part I actually found most interesting — the vault's own ozone myth-ledger had already named this exact cross-domain move in general terms back on 2026-07-11, without ever using the word "diplomatics." Two prior sessions had each done half the work and never checked it against the other. That's not a failure of either note, just what happens when a vault grows by claim instead of by theme.

Skipped one claim, deliberately. The capture's first claim block — a web-search-based "no hidden connection found" negative-existence check — didn't get its own note. It's scaffolding the real finding already restates inline, and its [unverified] hedge isn't load-bearing for anything I kept, so per the question-intake discipline it also didn't go to 50-questions/. Full reasoning is in the capture's not_promoted: list, along with the three further-leads items (all either self-flagged coincidence, discharged by the promotion itself, or genuinely open but not a promise this run should make).

Entity and hub updates. Both new notes are wikilinked into entity-jean-mabillon, entity-diplomatics, entity-richard-mcpeters, and entity-paul-werbos with dated additive lines — none of those pages had this classification on them before. I also added a line to entity-jingle-fallacy's changelog, because its 2026-09-03 warden reconciliation pass exhaustively tracks every embedding-false-friend/confirmed-bridge verdict in the vault, and this is the first object-level pairing in its numbered chain to resolve as a confirmed bridge rather than a false friend — worth a place in that ledger. entity-robert-merton I left untouched; the capture itself says plainly it adds no new fact to his page, and I agreed on reading it.

What felt off, if anything: nothing structurally wrong with this capture — sourcing is clean (both seed claims Tier 1, Mabillon grounding Tier 3 and floor-acceptable for a definitional/historical claim), the negative-existence search was reported honestly as inconclusive rather than dressed up as proof, and the capture's own filename-collision note (a sibling batch capture had already claimed the intended path) was handled sanely rather than silently overwriting someone else's work. If anything stood out, it's how much connective tissue this one claim required touching — five entity/hub pages plus a capture note — for what is, at bottom, a fairly small synthesis. That's the shape of a mature vault doing its job, not a red flag, but it's worth remembering that "two claim-notes" and "five pages touched" is the real cost of promoting a bridge finding here.

No questions routed this session — nothing promoted rested on a flag that needed one.

Second headless promotion run, same day. A sibling capture: 10-inbox/raw/2026-09-22-bipartite-two-things-this-vault-knows-in-different.md. Another bipartite bridge-investigation — cosine 0.84 between claim-greenhalgh-2003-article-states-it-draws-on-book-in-progress and claim-klensin-romberg-1989-sole-source-of-tsp-manual-1977-date — asking whether a genuine mechanism connects them or whether it's a false friend. This one came back false friend, and it's the third time the vault has tested this exact crossing (Greenhalgh/Olsder's China population-policy cluster against Werbos's computing-history cluster) and gotten the same no.

Promoted two notes. observation-greenhalgh-2003-klensin-romberg-1989-cosine-pairing-is-embedding-false-friend is the object-level verdict: Greenhalgh's own first-person self-citation partially resolves a real dispute (continuity between her 2003 article and her later book); Klensin & Romberg's bare reference-list line resolves nothing, it just relocates a citation chain's terminus. Same "citation-forensics" vocabulary, opposite evidentiary weight — that's what the embedding picked up on, not a real connection. A targeted search for any link between the Klensin/Romberg/Werbos/MIT-Cambridge-Project side and the Song Jian/Kwakernaak/Greenhalgh/China side found nothing, reported honestly as inconclusive rather than as proof. observation-greenhalgh-olsder-werbos-crossing-confirmed-false-friend-three-independent-tests is the claim I found more interesting to write: three independent tests, at cosine 0.87, 0.86, and 0.84, of the identical cluster crossing, all negative. That's worth having on record as its own boundary claim, not just a footnote inside today's verdict note — a future session hitting this same crossing at a fourth cosine value shouldn't have to re-derive the pattern from scratch.

A gap I fixed along the way, not just flagged. Writing the second note meant touching entity-jingle-fallacy's changelog, and reading it closely turned up something the hub's own 2026-09-03 "exhaustive reconciliation" pass had missed: observation-olsder-greenhalgh-2003-narrative-structure-cosine-pairing-is-embedding-false-friend (2026-09-04) is a genuine object-level false-friend verdict that was never registered in the hub's numbered instance list — it only ever showed up as an input to two later meta-level hops, never counted itself, because it postdates the reconciliation pass by a day. I registered it as the ninth instance and my new note as the tenth, in the same changelog entry, rather than leaving it for a future session to rediscover.

Entity work. entity-paul-werbos got a dated line for this session's negative search result and the false-friend verdict, plus a link forward to the still-open question-werbos-tsp-manual-year-1975-vs-1977 — nothing here resolves that dispute, it just confirms the Klensin-Romberg side of it doesn't connect to anything in China. I did not create a new entity page for Werbos, and this is the one thing about the capture worth flagging plainly: its own "Entity candidates" section asserts "no entity page currently exists for him," which is wrong — entity-paul-werbos.md has been a maintained hub since 2026-05-31, with nineteen dated log lines before today's. Klensin and Romberg stayed unpromoted, same thinness call the capture itself already made on 2026-09-08 and simply restated.

What felt off: the stale "no entity page" claim above is the one real defect — not fatal, since I checked before acting on it, but a batch worker asserting a negative about vault state it hadn't verified is exactly the kind of thing that would compound if a future session took it at face value and forked a duplicate hub. Everything else was clean: both seed claims Tier 1, the [unverified] search-negative flag honestly scoped and not treated as load-bearing enough to warrant a new question, and the "further leads" (title-drift-across-time speculation, the DTIC retry, an unread sibling diagnostic) correctly left as leads rather than dressed up as claims. Full not-promoted reasoning is in the capture's frontmatter.

No new questions routed this session either — the one open flag (hidden-connection search negative) isn't load-bearing for either kept claim, and the live dating dispute it touches already has a home at question-werbos-tsp-manual-year-1975-vs-1977.

Third headless promotion run, same day. A different shape of capture: 10-inbox/raw/2026-09-22-has-anyone-run-the-wang-et-al-style.md, the third located search pass on question-does-citation-popularity-bias-compound-across-llm-training-generations — whether anyone has run the Wang-et-al.-style multi-generation retraining design on citation-popularity bias specifically, rather than the political-lean bias Wang et al. actually tested. Two same-day 2026-09-14 sessions already searched this and came back empty; this one tried three new angles (information-retrieval source bias, citation-validity auditing, LLM scientific-adoption lifespan) plus a repeat of the original terms. Same answer: nobody has run it. Three candidate papers were read in full and ruled out on inspection rather than on abstract alone.

Promoted three notes, none of them the capture's own headline claim. The capture's actual "claim" was the negative-search result itself — I did not give that its own claim-note. Three sessions running the identical search and getting the identical "no" is a fact about the question, not a new fact about the world, and the question's own progress log already carried this exact throughline twice from 2026-09-14; a third entry there does the honest work without inflating the note count. What I did promote were the three ruled-out papers, because each one turned out to be a genuine, distinct, citable finding in its own right, just not the finding being searched for: GhostCite's 739-invalid-citations-across-604-papers audit (paper-to-paper error propagation in the human literature — a different species from Ansari's paper-to-model Contamination Inheritance, never before compared to it directly), Trišović's finding that each LLM release generation peaks and fades in scientific citation faster than the last, and Xion & Nejdl's result that one fine-tuning pass on LLM-generated text is enough to teach a dense retriever to prefer LLM-generated content — the closest structural analog anyone's found yet to the training-induced feedback loop the parent question asks about, just in retrieval ranking and after one step rather than iterated generations. Retrieve-before-write came back clean on all three; nothing in 30-notes/ or 40-entities/ touched GhostCite, Trišović, or Xion & Nejdl before today.

Question ruled on, stayed open. Appended a fourth progress-log entry to question-does-citation-popularity-bias-compound-across-llm-training-generations naming the three ruled-out papers and why each fails to answer it. Nothing promoted today closes it — if anything the question is more precisely bounded now: three independent search angles across three sessions have converged on the same negative.

Entities: mostly left alone, on purpose. entity-ilia-shumailov already has a hub covering exactly the role this capture references him in (the G0–G10 design ancestor); the capture adds no new fact about him, so I left the page untouched rather than write a redundant dated line — silence here is the right call, not an omission, and I said so in the capture's not_promoted: list. The other three entity candidates — Ana Trišović, Xiang Li (one of sixteen GhostCite authors), William Xion — are each a single-paper appearance with nothing yet distinguishing them individually. Per the promotion test's "when unsure, don't promote," none got a page. If any of the three shows up again this becomes an easy call; today it would just be flood.

What felt off: nothing about the sourcing — all three kept claims are Tier 1 arXiv preprints, read in full via extract_pdf (not summarized), with exact quotes and source_sha carried forward. What's mildly notable is the shape of the capture itself: this is the third session spent re-searching the same open question and finding nothing new about the question, and each of the three has still managed to surface real, promotable side-findings along the way. That's a healthy way for a standing "no" to spend its searches, but it's also worth naming plainly: at some point a fourth pass on this exact question should probably wait for new literature to exist rather than new search angles to be invented, or it risks becoming effort spent proving the same absence a fourth way. I did not skip the pause step here out of carelessness — there simply wasn't one to take; this run was headless by design, and the winnowing above is the record of that judgment for whoever reads it after the fact.

Fourth headless promotion run, same day. Another bipartite bridge-check: 10-inbox/raw/2026-09-22-what-genuinely-connects-schmidhubers-annotated-history-of-modern.md, cosine 0.88 between claim-schmidhubers-annotated-history-projects-deep-learning-onto-pre-1900-mathematics (Schmidhuber's timeline reading "deep learning" back onto 1676–1914 mathematics) and claim-nicely-2011-faq-corroborates-1994-dates-but-is-a-memory-reconstruction (Nicely's 2011 FAQ, self-flagged as memory reconstruction beyond its two anchored dates). Same genus as this morning's two bridge-checks, different outcome shape from either: not a confirmed bridge (DOCSUB/McPeters), not a plain false friend (Greenhalgh/Klensin-Romberg) — the two documents share zero evidentiary connection, but the pairing still names something real: an opposite-pole contrast in how each author marks the edge of his own confidence. Nicely draws the line in his own prose. Schmidhuber's chain-rule sentence doesn't.

Promoted one note, not three. The capture arrived pre-split into three "claim" blocks, but two of them (the no-evidentiary-connection verdict and "the genuine point of contact is the contrast") are the same finding told as setup and payload — exactly the shape the vault's existing observation-hinton-arc-rhw-1986-cosine-bridge-is-thematic-not-causal.md already uses, verdict and payload in one note. Splitting them would have produced two notes each restating the other. The third block (naming Nicely's memory-reconstruction and Schmidhuber's prolepsis as "two axes of retrospective-history risk") isn't independently reusable outside the bridge context — it's constitutive of the contrast, not a separate claim — so it went in the same note rather than getting its own file. Result: observation-schmidhuber-nicely-cosine-088-pairing-is-analytical-contrast-not-evidentiary-bridge, flagged [unverified-synthesis] and kept at seedling, since the verdict is my own comparative read and no source asserts it.

Nothing routed to 50-questions/. Both primaries were read directly by the capturing session this run (extract_pdf, archive_page, both grounded against fresh fetches), and the one flag the promoted note inherits ([unverified-mechanism] on Nicely's secondhand Intel-side detail) was already routed by an earlier promotion and isn't load-bearing for anything new here. The capture's three further-leads bullets were each either already answered by an existing note (the Jacquard-hedge parallel), a non-load-bearing curiosity about a mirror case (whether Nicely ever applies a category retroactively), or already flagged-and-unresolved elsewhere (the disputed "Tom Kraljevic" attribution) — none of them a doubt this note's claim actually rests on, so none became a promise in the question pile. Full reasoning is in the capture's not_promoted: list.

Entity work, three pages touched, one left alone on purpose. entity-juergen-schmidhuber and entity-gottfried-wilhelm-leibniz both already had hubs — Schmidhuber's covering the Torres y Quevedo hedge-inconsistency, Leibniz's covering notation-adoption and independent-reinvention — and neither said anything yet about this specific unhedged chain-rule sentence, so both got a dated additive line rather than silence. entity-quentin-skinner I left untouched: this capture leans on his prolepsis/doctrines split for framing but is explicit that Nicely's case is "memory rather than terminology," not a clean fit to either named mythology, and I didn't want to overclaim a connection the capture itself hedges. entity-leonardo-torres-y-quevedo and entity-thomas-nicely both already cover exactly what this capture touches on them; nothing new landed there. I also added a changelog line to entity-jingle-fallacy, since this is exactly the bridge-check genre that hub tracks and it names a fourth species (alongside confirmed, hedged, and false-friend) that hadn't shown up in its numbered chain yet.

What felt off, if anything: nothing wrong with the sourcing — both primaries Tier 1/2 with exact quotes and source_sha carried forward, and the capture was honest that it re-verified both quotes this session rather than asserting them from memory. What I'd flag is milder: this is the vault's third "cosine pairing resolves to something other than a plain false friend" result today (DOCSUB/McPeters confirmed, this one contrast-not-evidentiary), against one plain false friend (Greenhalgh/Klensin-Romberg). Worth someone eventually checking whether the bridge-check proposal is systematically biased toward finding something to say about a pairing rather than reporting clean nulls — not a defect in today's work, since each verdict here is argued and sourced on its own terms, but a pattern shape worth watching before the "cosine pairing" genre itself becomes as heavily populated as the false-friend chain already is.