journal 2026-11-02

Journal — 2026-11-02

Headless promotion session. I promoted 10-inbox/raw/2026-11-01-does-an-independent-source-corroborate-tieman-markou-2026s.md — the capture checking whether anyone outside Tieman & Markou has corroborated their "generative-process diversity" finding.

The capture held three claims (arXiv/NeurIPS record check, the Artificial Hivemind contrast, Bradley's "Madness of Crowds" ruled out), and I winnowed them into one claim-note, claim-tieman-markou-2026-no-independent-corroboration-as-of-2026-11-01. All three were angles on the same finding — no independent corroboration found — not three independently-falsifiable claims, so splitting them would have manufactured atomicity rather than found it. That note keeps the [unverified-mechanism] flag on the Bradley sub-point (abstract-only read this session, no verbatim quote) and stays status: seedling.

I did not route a new question. The capture's own "further leads" already named the one thing that would actually move this forward — the NeurIPS listing's unopened OpenReview page, which could carry outside reviewer critique — so I appended a progress line to the existing question-corroborate-tieman-markou-2026-generative-process-diversity instead of either closing it or duplicating a question that already asks for exactly this. Forcing it to answered felt wrong: the capture itself flagged the gap, and a question that already names its own next move doesn't need a sibling.

Entities: I gave entity-artificial-hivemind a full hub — it's real, peer-reviewed (NeurIPS 2025 D&B Oral), and it's the paper Tieman & Markou themselves measure their own contribution against, so it's foundational to this whole branch, not just adjacent to it. I updated the existing entity-generative-process-diversity stub with a dated line rather than touching its body — it already existed (the capture was wrong to say no page existed yet; a stale assumption from whoever wrote the capture, not a researched claim, so I didn't flag it as a sourcing problem). I did not promote Ross Tieman, Evan Markou, Liwei Jiang, William F. Bradley, or Yejin Choi to entity pages — every one of them is standing on co-authorship of a single paper this vault reads, and the 2026-10-01 ruling is specific that co-authorship alone is a mention, not a hub. Four names is a flood I'd rather not start. I also skipped Normalised Compression Distance — a real technique, but it shows up in the vault exactly once, inside this one preprint's method section, with nothing yet asking it to recur.

What felt a little off: the capture's "further leads" section carries emergentmind.com and academy.dair.ai summaries that it correctly identifies as Tier 4/5 noise inflating search-hit counts without adding a source — a good catch by whoever wrote the capture, and exactly the kind of thing I'd rather see named and discarded than silently absorbed into a "found several results" impression. Otherwise the sourcing here was clean: two direct Tier-1 quotes off the arXiv and NeurIPS pages, a third off Artificial Hivemind's own abstract, and an honest, correctly-flagged weak spot on the one source (Bradley) that was only read at abstract depth. No tier inflation I could find.

Skipped per the headless-operation rules: no pause for sign-off (none available), no git commands (the auto-commit agent handles that), and the report above replaces what would otherwise have gone to Cali directly.

No mid-session noticings worth a 00-meta/seek-flags.md line this time — the one "could be read as structural" candidate (the capture's wrong claim that no entity page existed yet) turned out to be a one-off stale assumption inside this single capture, not a repeating pattern, so I left it in the journal rather than the ledger.

— Seek

Second headless promotion, same day. I promoted 10-inbox/raw/2026-11-01-hop-tches-hexels-peer-review-false-friend.md — a hop-batch capture adjudicating a vault-index pairing between claim-tches-peer-review-process-clears-bit2watt-of-unrefereed-cap and claim-hexels-2dgs-triangle-splatting-baselines-peer-reviewed-hexels-is-not at cosine 0.8857 (the Hexels note sits at the 99.9th percentile of the TCHES note's own neighbour list — about as strong an unlinked pairing as this vault's index produces).

I winnowed the capture down to one claim-note, observation-tches-hexels-peer-review-role-inversion-is-embedding-false-friend. The core finding — the two notes share "peer review clears an unrefereed cap" vocabulary but assign the reviewed/capped roles to opposite parties (Bit2Watt is the reviewed, capped subject; Hexels' baselines are reviewed while Hexels itself stays capped) — is one atomic claim, and the cosine measurement plus the negative author-link search fold into it as supporting detail rather than spinning off into their own notes, per the "atomic is not many" ruling. I left out two things on purpose: the capture's own "why this was hop-worthy" meta-line (that's a reflection on the one finding, so it went in the note's commentary callout, not a second note), and the Alan Yuille career hook the capture flagged as a saved lead — that one already has its own sibling capture (2026-11-01-hop-yuille-hawking-marr-bridge.md), so it's a different session's promotion, not mine to duplicate.

Retrieve-before-write turned up no collision — nothing in 30-notes/ already names this exact TCHES/Bit2Watt × Hexels/2DGS pairing, despite the vault having ~70 siblings in the observation-*-cosine-pairing-is-embedding-false-friend genre. I gave the new note the same audit_status: "synthesis — [unverified-synthesis]" treatment the genre already uses when the verdict is Seek's own connective read rather than an externally-asserted claim (closest precedent: observation-perceptrons-scan-merton-1957-cosine-pairing-is-false-friend-sharper-instance-elsewhere). No new question — there's no unresolved external verification here, both underlying notes are already capture-verified/verified-verbatim, and this genre doesn't typically spawn questions since its subject is the vault's own retrieval geometry, not a fact pending confirmation. I did check seek-flags.md for the long-standing "moc-embedding-false-friends is unbuilt infrastructure" thread before adding to it myself — it's already logged and was explicitly declined by a 2026-09-25 Warden pass ("a naive duplicate restates the hub"), so I left it alone rather than re-flagging settled territory.

Entities: no ## Entity candidates section in this capture, so the promotion test didn't apply. entity-hexels already carries a dated line from 2026-10-29 covering both notes this capture's finding touches; this capture's own contribution is about the vault's embedding behavior, not a new fact about Hexels itself, so I left the entity page alone rather than padding it with a line that doesn't teach it anything new.

What felt a little off, worth naming: nothing in the sourcing — both underlying notes are clean, Tier 1, already audited, and the capture's own due-diligence hop (checking for a real author-team connection) is good discipline, matching the method the vault used on the earlier RIKER/Babylonian-Diaries false friend. The only soft spot is structural, not this capture's fault: I noticed (but didn't act on, since it's outside this capture's scope) that the observation-*-embedding-false-friend genre still sits in a documented, unresolved tension with promote.md's 5a rule about vault-search claims belonging in questions, not notes — flagged already on 2026-10-10, still open, still unruled-on. I followed the same precedent every prior promotion session in this genre has followed rather than re-litigate it here.

Skipped per the headless-operation rules: no pause for sign-off (none available), no git commands (the auto-commit agent handles that), and this entry replaces what would otherwise have gone to Cali directly.

— Seek

Third headless promotion, same day. I promoted 10-inbox/raw/2026-11-02-hop-citation-as-legitimation-bridges-song-jian-chatgpt.md — a hop-batch capture testing whether the vault's seed pairing of claim-greenhalgh-2003-article-documents-kwakernaak-1977-citation-in-song-1981 and claim-petiska-2023-chatgpt-cites-by-google-scholar-count-perpetuates-matthew-effect (flagged unlinked near-neighbors) was real or just shared vocabulary.

I winnowed it into two notes, not one, because they're genuinely separable: claim-greenhalgh-2003-frames-kwakernaak-citation-as-legitimacy-transfer-independent-of-content carries the primary-text finding on its own — Greenhalgh's 2003 article says, in her own words, that citing foreign science gave Song Jian's group "unassailable legitimacy" while the foreign work itself was only a "heuristic device," never a policy recommendation. That claim stands whether or not the ChatGPT comparison holds. The bridge itself — observation-greenhalgh-kwakernaak-citation-petiska-chatgpt-citation-confirmed-legitimation-bridge — is a second, separately falsifiable claim: that this mechanism also describes what Petiška's 2023 study found in ChatGPT's citation behavior. Retrieve-before-write turned up no collision: I checked both quotes ("unassailable legitimacy," "heuristic devices") against the full vault and neither string exists anywhere yet, despite this Song-Jian/Kwakernaak cluster already running to ~90 notes. I also read observation-petiska-greenhalgh-footnote-meta-pairing-is-false-friend-with-shared-olsder-entity, the nearest-sounding existing note, closely enough to confirm it adjudicates a different pairing entirely (two meta-level vault notes about Olsder, not these two seed claims), and read claim-greenhalgh-2003-poses-legitimation-versus-origination-question-explicitly and moc-legitimation-not-origination to confirm this capture's "citation as legitimation ritual, independent of content" is a different axis from that cluster's "did science originate or merely legitimate the policy" — the capture's own hop 2 novelty check (frontier, 0.735, P44.4) already flagged this distinction, and a direct read confirmed it rather than just trusting the number.

I left the Latour & Fabbri / Bourdieu theoretical framing out of both notes and did not route a question for it. It's genuinely unread — a WebSearch summary, never fetched, correctly flagged [unverified-quote — needs direct read] by the capture — but it isn't load-bearing for either kept claim: the bridge stands on Greenhalgh's and Petiška's own words without needing a named theoretical school to back it. Per the question-intake discipline, that makes it a nice-to-verify lead, not a promise-worthy question, so I recorded it as an unread lead inside the observation note's body and named it again in the capture's not_promoted: list instead of opening a 50-questions/ file for it. I checked 50-questions/ for anything already covering Latour, Bourdieu, or "citation as legitimation" first — nothing did.

Entities: "scientism" was a genuine first encounter (vault_mentions 0), so it got a status: watching stub immediately, entity-scientism, stamped first_seen: 2026-11-02 — the etymology (OED ~1870, G.B. Shaw, popularized by Hayek) is recorded honestly as an unfetched search lead, not a verified claim. Huibert Kwakernaak already has a hub; this capture taught it something new — that his own relative obscurity as a cited figure is itself evidence against reading this as a literal Matthew-effect case — so I appended one dated line rather than leaving the page silent about what this session learned. I did not touch Qian Xuesen's entity page: the capture's own saved-hooks section says the vault already documents both his recruitment of Song Jian and his later qigong advocacy as separate notes, and a fresh read of that hub confirmed nothing here adds to it.

What felt off, worth naming: nothing in the sourcing — both anchor quotes are Tier 1, read directly off a PDF already archived elsewhere in the vault with an unchanged sha, and the capture was honest about the one claim (Latour/Bourdieu) it hadn't actually verified rather than dressing it up as settled. If anything, this capture is a good example of the hop method working as designed: it went looking for a false friend and found a real, previously-unwritten primary-text finding instead, five pages from a citation the vault had already mined once. The only soft spot is density, not quality — this is now the fourth or fifth "confirmed bridge" observation note in a cluster that is well past the size where a navigation MOC was flagged as wanted (seek-flags.md), and I didn't build one this session; noted below rather than taken on scope-creep.

— Seek

· raw markdown