Journal — 2026-10-25

Promoted one capture, headless: 10-inbox/raw/2026-10-25-chronocultures-older-year-naming-asymmetry-is-the-one.md. No one was watching, so I winnowed on my own judgment rather than pausing, and I'm logging the reasoning here for Cali to check against.

The capture was a targeted search session against question-corroborate-chronoculture-older-year-naming-asymmetry — the one claim in the ChronoCulture branch that Daryani et al.'s corroboration didn't reach. It found no independent test of the specific older-year/large-shift effect, but it did real work getting there: it resolved a long-dangling citation (Liu, Kaneko & Chu, AAAI 2026) and read it directly, and it read Daryani et al.'s own limitations section directly rather than relying on the citation-network shorthand for what that paper found.

I wrote two notes, not three, out of the capture's three claim-blocks. The "no independent test found" verdict and Daryani's limitations-section caveat are one finding under the 2026-10-01 atomicity ruling — the caveat is what makes the verdict honest, not a separate claim — so they became one observation note, observation-chronoculture-older-year-asymmetry-no-independent-test-found. The Liu et al. paper's own finding (unprompted opinion distributions align most with the most recent WVS wave, among pre-selected well-aligned countries) is genuinely separate — a third data point for the branch's present-skew pattern, independently linkable — so it got its own note, claim-liu-kaneko-chu-llm-opinion-alignment-strongest-with-most-recent-wvs-wave. Retrieve-before-write turned up the question itself and the two existing Daryani claim-notes as the closest neighbors; both new notes link into that cluster rather than restating it.

I did not create a new question. The capture answers directly into the question that spawned it, so I left its status: open and appended a dated progress-log entry instead — the Liu et al. lead is now closed (checked and negative, not just unfetched), and I moved the remaining unchecked leads (Chae et al., two flattening/alignment papers, the mechanistic anchoring-bias literature, Daryani's public code repo) into that log rather than writing them up as claims nobody read deeply this session.

Entity work: the capture flagged six candidates. Ronald Inglehart already has a hub and nothing here adds to it, so I left it alone. The Inglehart–Welzel cultural map was already explicitly deferred by yesterday's promotion and nothing changed that call. I declined hubs for Yalda Daryani and for Liu, Kaneko & Chu — each is the author or co-authors of exactly one paper the vault reads, which is co-authorship-alone under the 2026-10-01 ruling even though Daryani's paper is unusually load-bearing for this branch. I logged that as a watch item in 00-meta/seek-flags.md rather than silently dropping it, since her paper's centrality makes her the likeliest of the two groups to earn a hub later. I did append one real update to an existing entity hub: entity-world-values-survey.md now logs that a third independent paper (Liu et al.) uses WVS data in this citation network, which is new information the page didn't have.

I also discharged a stale flag: a 2026-10-24 [watch] line had Liu et al.'s citation unresolved and unfetched. This capture resolved and read it, so I added a → resolved line under that entry rather than leaving it to rot.

What felt off: nothing structurally — this was a clean, honest search-negative, which is a legitimate outcome and not a sign of weak work. The one thing worth naming is that the capture's own framing ("the one" in its filename, and "this is the complete, honest outcome... not a gap in search diligence") reads a little defensive, as if anticipating being read as a failed session. It wasn't a failed session. A well-documented "checked, not found" is exactly what should go in the question's progress log, and that's where most of it ended up.

Skipped per the headless-operation changes to /promote: the pause-for-okay step (no one here to ask), and the git commit (the auto-commit agent handles that within 15 minutes — I ran no git commands this session).

Second promotion, same session

Promoted a second capture, also headless: 10-inbox/raw/2026-10-25-hop-greenhalgh-june1978-whyte-wang-cai-bridge-false-friend.md. This one was a bridge-seed check the hop-batch worker had already done the hard part on — both primaries reread in full, both shas confirmed unchanged — testing a cosine-0.88 unlinked pairing between claim-greenhalgh-2003-article-independently-corroborates-june-1978-leading-group-meeting and observation-whyte-wang-cai-2015-essay-bridges-three-stage-timeline-and-yicheng-1985-dating-seed-notes. Verdict: false friend. Greenhalgh's June-1978/Li-1980 paragraph and Whyte-Wang-Cai's "already in 1978" sentence never touch — different section, different footnote, Chen Muhua instead of Li 1980 or the Leading Group. I wrote one note, observation-greenhalgh-whyte-wang-cai-cosine-088-pairing-is-embedding-false-friend, matching this cluster's established naming convention for these false-friend verdicts so a future retrieve-before-write finds it by shape, not just content.

I did not write up the capture's other lead. It had already found a real, stronger, still-unlinked pairing one hop over (cosine 0.9072, the identical Greenhalgh sentence shared between note A and the three-stage-timeline claim) and run its own vault_novelty check against that framing, which came back duplicate-risk at the 92.8th percentile. That's the capture's own judgment, and I agreed with it — promoting it anyway just to have a second note would be exactly the "size reads as corroboration" failure the sourcing-floor spec warns about. It stays as a named lead in the capture, not orphaned, since the capture itself isn't deleted. I also didn't chase the saved footnote-29 hook (the "400 million births" myth citation) — genuinely out of scope for this chain, already logged as such in the capture.

No new question. Nothing here is load-bearing on an unresolved doubt — both primaries were read directly this session, the quotes are Tier 1 both sides, and the false-friend verdict rests on a full-text negative search of two specific documents, not a hedge needing future verification. The intake discipline says don't open a question for a nice-to-chase lead, and the deferred 0.9072 pairing is exactly that: interesting, not load-bearing on anything I just wrote.

Entity work: both candidates — Susan Greenhalgh and Martin King Whyte — already have hub pages, and this capture taught something neither page yet said (that this specific pairing had been tested and found false), so I appended one dated line to each rather than staying silent about a known entity. No new stub needed.

What felt off: nothing about the sourcing — both shas matched the vault's existing citations, both quotes are verbatim Tier 1, the gate check (86.4th percentile, "adjacent") cleared the duplicate-risk band the lead itself tripped. If anything this cluster is now dense enough that I want to flag, gently, the same pattern Cali's audit already caught once in this cluster (the "shared landlord" tally kept in six different notes, agreeing with none of them) — this promotion didn't add a tally claim, so I'm not reopening that, but the cluster's false-friend genre notes are themselves now numerous enough (a dozen-plus "cosine-pairing-is-embedding-false-friend" notes just in this one population-policy cluster) that a future MOC pass naming them as a genre, rather than leaving each as a standalone seedling, might be worth the Warden's attention. Not flagging it to seek-flags.md — it's a shape worth watching, not a defect, and no kept claim here rests on it; naming it here is enough for now.

Third promotion, same session

Promoted a third capture, also headless: 10-inbox/raw/2026-10-25-outside-arxiv261030900-silicon-samples-match-the-answers-but.md — Kumagae & Kaneko's "Silicon Samples Match the Answers but Not the Respondents" (arXiv:2610.30900), an externally-chosen preprint outside any existing vault thread, read directly by the bee at capture time (both the PDF and the abstract page, TLS verified, no safety flags).

The capture handed me five claim-blocks, all resting on the same single paper — and that paper is an unrefereed preprint submitted three days ago with zero chance of independent corroboration yet. The single-source concentration cap (sources.md: no more than three claim-notes on one unrefereed primary without independent confirmation) binds here on first contact, not after a failed search, so I treated three as the ceiling rather than the plan. I wrote three notes: claim-silicon-sampling-llms-undercover-human-respondents-despite-robustness-sweep (merging the capture's claims 2 and 3 — the headline RCG-above-1 result and the robustness sweep that survives temperature, reasoning effort, model tier, and sample-size variation are one finding under the 2026-10-01 atomicity ruling, since the sweep is the caveat that makes the headline honest, not a separable claim), claim-per-question-metrics-blind-to-silicon-sampling-respondent-coverage-gap (the Marginals-near-ceiling-correlation-despite-failing-RCG result plus the formal permutation-invariance proof against a prior metric), and claim-silicon-sampling-llms-split-into-off-support-and-contraction-failure-modes (the two-group OSR taxonomy). That's exactly three, exactly the cap. I opened question-corroborate-silicon-sampling-rcg-single-source to carry the branch forward and named, inside it, the two cited-but-unread papers (Wang/Morgenstern/Dickerson 2025, Bisbee et al. 2024) most likely to either corroborate or complicate it.

The capture's first claim — the bare definition of RCG as a metric — I did not write as a fourth claim-note at all. It's genuinely conceptual content (what a newly coined term means), and the entity-page spec wants emerging terms stamped with a first_seen date immediately rather than written up as prose claims, so it became a watching stub, entity-respondent-coverage-gap, alongside a matching stub for the paper's second metric, entity-off-support-ratio. This also happened to be the right move under the cap — it kept the three-note budget for the empirical findings instead of spending a slot on a definition. I also opened a hub, not a stub, for entity-silicon-sampling itself: the technique is established (Argyle et al., 2023), not emerging, and this is now the vault's second paper to touch the same WVS wave-7 dataset through Kaneko, so it earns the fuller page over a cheap stub.

Retrieve-before-write surfaced only false-positive hints — the semantic index's nearest neighbors were a run of claim-notes about literal silicon spheres (the Avogadro-kilogram metrology cluster) and Achim Leistner, matched on the word "silicon" and nothing else. The real neighbors were two I found by hand: claim-liu-kaneko-chu-llm-opinion-alignment-strongest-with-most-recent-wvs-wave (same co-author, same WVS dataset, a different research question — I said so explicitly in both the new claim-note and the entity-silicon-sampling page rather than letting the overlap read as more connection than it is) and entity-world-values-survey, which I updated with a dated log line for this fourth paper now drawing on its data. No real collision, so nothing was skipped or merged because of prior coverage — only because of the cap.

Entity work beyond the two stubs and the hub: I declined hubs for all four named people (Kumagae, Kaneko, Argyle, Santurkar) — Kumagae as sole-standing author of one uncorroborated preprint, Kaneko under the same co-authorship-alone reasoning as yesterday's entry (though I flagged separately in seek-flags.md that he now recurs across two distinct co-authored papers, which is a notch past the simple case that ruling was written for), and Argyle/Santurkar because this capture only cites them secondhand — their actual foundational papers are still unread. Those two remain open Further leads inside the capture itself, not lost. I also declined a hub for ANES, on the same "not yet recurring" logic already applied to ISSP on the WVS page.

What felt off: nothing about the sourcing itself — direct PDF read, TLS verified, exact quotes, no adversarial content. What's worth naming is the shape of the risk rather than any defect in this specific capture: a three-day-old, two-person, single-preprint result generated five quotable claims and, if I'd promoted all of them without the cap, would have looked like a small well-sourced literature rather than one unreplicated finding. The cap did its job. I also logged a [post] noticing — this result and the ChronoCulture/Daryani present-skew cluster are both "matches the statistic, misses the thing" findings about LLMs standing in for survey respondents, which feels like a real essay spine, not this session's to write.

Fourth promotion, same session

Promoted a fourth capture, also headless: 10-inbox/raw/2026-10-25-what-genuinely-connects-song-jians-1978-self-report.md — a bridge-seed check against a cosine-0.88 unlinked pairing between observation-song-jian-1978-self-report-predates-olsder-account and claim-whyte-wang-cai-2015-essay-reliable-on-national-dating-unreliable-on-yicheng-specific-date, both already in this vault's long-running Song-Jian/Greenhalgh/Liang "shared landlord" cluster.

The capture handed me three claim-blocks, and I wrote one note, not three. The verdict (both notes cite Greenhalgh's 2003 article, for unrelated passages fifteen pages apart), the evidence (the two passages' citation sets are fully disjoint — I kept the capture's enumerated chain: Li 1980, Deng 1984, Chen Muhua 1981, Government Work Report 1985, Xinhua 1979 — none touching Song's self-report), and the conclusion (neither note needs the other to be true) are one finding under the 2026-10-01 atomicity ruling, not three separable claims — you can't disbelieve the evidence while believing the verdict. That's observation-song-self-report-and-whyte-wang-cai-national-reliability-notes-connect-only-through-greenhalgh-2003. Retrieve-before-write (helped by the session's own retrieval hint, plus a direct check of both target notes and the cluster's existing "shared landlord" siblings) confirmed this exact pairing hadn't been tested before, even though the shape has now run at least six or seven times in this cluster.

I didn't open a new question. The underlying 1975-vs-1978 dating dispute this cluster orbits is untouched by a finding about which pages of one article two other notes each cite — there's no new load-bearing doubt here, just confirmation that this pairing doesn't move it. I did append a progress-log entry to question-verify-song-jian-1975-twente-delegation-visit, since the self-report note is one of the notes the question already tracks and the established convention in that question's own log is to record every bridge-check that touches it, resolving or not — this one doesn't.

Entity work: four candidates, all already hubbed, all declined for new standing. Song Jian and Olsder got nothing added — the capture's observations about them are about the vault's citation structure, not new facts about either man, consistent with how this cluster's other "shared landlord" verdicts have never been added to Song Jian's own page either. Susan Greenhalgh did get an update: the specific political-process citation chain behind her p. 185 passage (Deng, Chen Muhua, the Government Work Report, Xinhua) hadn't been named on her entity page before, even though the page already logs a dozen-plus bridge-checks against this same article — so I appended a dated line rather than staying silent about a known entity learning something new.

What felt off: nothing about the sourcing — the Greenhalgh PDF was re-fetched this session with an unchanged sha, all three quotes read directly from the fresh extraction rather than carried from frontmatter. The capture did leave one lead unchased that I think was the right call to leave rather than force: Greenhalgh's own p. 185-186 section makes a causal-credit argument (political momentum predated the scientists) structurally identical to the separate Liang-vs-Song causal-credit cluster, sourced independently of Liang, and nobody has checked the two against each other directly. Not load-bearing for anything I just wrote, so not a question — logged as a [watch] in seek-flags.md instead. Also worth naming once, since the capture's own "further leads" section raised it: this cluster's running tally of how many "shared landlord" instances it has produced is kept in prose, in whichever note happens to notice next, and has drifted out of agreement with itself at least twice already (an audit already caught this on 2026-10-14). I didn't try to fix or extend that count here — writing yet another ordinal into yet another note is how it got broken in the first place — but it's now happened often enough that a single owned tally feels overdue.

Fifth promotion, same session

Promoted a fifth capture, also headless: 10-inbox/raw/2026-10-25-what-is-the-exact-url-and-source-sha.md — a pure provenance-recovery session closing question-recover-source-url-sha-kaufherr-tau-faculty-page-phd-thesis-listing, which a scheduled cross-model audit had left one item short two days ago: it found Tirzah Kaufherr's own Tel Aviv University faculty page and recovered the source_url, but had no archive_page to produce a byte receipt, so the source_sha stayed owed. This capture had archive_page: it re-fetched the same URL (sha256 c5648624fc80…), ran quote_check against the archived text, and got grounded: true on the quote claim-kaufherr-1981-phd-thesis-predates-1984-journal-paper-by-three-years already rested on.

I wrote zero new claim-notes. All three of the capture's claim-blocks are exactly the case promote.md's §5a third check names explicitly — "claims about a document's metadata (pagination, edition, URL, catalogue record) ... belong in the relevant question's log, not in 30-notes" — so instead I: (1) added source_sha to the existing claim-note's frontmatter and upgraded its audit_status from capture-verified to verified-verbatim, since this is now an independent, hash-stamped re-check of the primary text and not just the original capture-time read; (2) set the question to status: answered with an answered_log naming what settled it, and appended a resolution section to its body; (3) folded the capture's third claim — that the same page also lists the 1983 Tokyo and 1984 Phys. Rev. D entries — into the claim-note's audit_status as corroborating detail, exactly as the capture's own text already concluded ("No new claim-note is needed for this"). Retrieve-before-write here meant reading the target claim-note and question directly rather than searching 30-notes blind, since the capture names both explicitly.

Entity work: I ran the promotion test on Tirzah Kaufherr, who the capture flagged as "worth revisiting... flagged for the queen's judgment, not created here." I promoted her to a hub. The 2026-10-01 co-authorship-alone ruling is clear that sharing one joint paper with Aharonov wouldn't be enough — but her 1981 PhD thesis is solo-authored, read directly from her own Tier-1 faculty page, and it's the one piece of this vault's Aharonov priority-dating thread that's independent of Aharonov's own memoir. That's separate standing, not just a co-author line, so I wrote entity-tirzah-kaufherr with first_seen: 2026-10-23 (when she first entered the vault, misspelled "Tamir," in the Aharonov oral-history capture) and noted her page's further, unexamined bibliography (an undated 1974 M.Sc. thesis, seven 1998–2014 papers) on the hub body rather than minting speculative questions for any of it. I checked entity-yakir-aharonov too; its existing 2026-10-24 log line already covers the faculty-page corroboration, and this capture adds a provenance detail (the sha) rather than a new fact about Aharonov himself, so I left that page untouched and am saying so here rather than silently skipping it.

What felt off: nothing. This is as clean as a provenance-recovery capture gets — the two sessions (Tuesday's audit, this one) split the URL-then-sha work without duplicating effort, the quote_check came back positive on the first try, and the capture's own commentary already called the right shape ("no correction needed... only missing its receipt") before I ever opened the files it touched. The only judgment call of substance was the entity promotion, and I think it's the right one: it's a hub that earns its keep by marking a real distinction (solo work vs. co-authored work) rather than restating a co-author relationship the vault already has on Aharonov's own page.

— Seek

Drafting session: "one level down"

Separate session, drafting not promotion. I wrote a new post, 70-drafts/one-level-down/, on the Mariel Boatlift wage debate.

Why this thread, over everything else ripe. I started at the constellation map's draft-coverage section, not a brute read. Two of the strongest-looking unwritten clusters turned out to be already spent: Prussian blue is fully drafted in made-to-leave ("the blue that leaves"), and the Schoenberg/spline cluster is drafted in the-wood-already-knew. Both carry drafted_in: stamps, which saved me from re-drafting a spine I'd already used — exactly what that stamp is for. I checked a shortlist of 0-MOC-coverage unnamed clusters for the same stamp and found the Mariel cluster clean: seven notes, none drafted, and — unusually for a seedling cluster — most legs confirmed at Tier 1 by cross-model audits (Card 1990, Borjas 2015/2017, Peri & Yasenov, Clemens & Hunt, Borjas's reply, all read as primaries; Clemens's blog via Wayback; Roodman at Tier 2).

What made it feel like the most worthwhile thing to write now, and not just the best-sourced: it has a real thesis, not a summary. The Mariel Boatlift is labor economics' cleanest natural experiment — one city, one summer, 125,000 people, the world doing the randomizing instead of the researcher — and that cleanliness is exactly why it settles nothing. The finding I could stand behind is that the researcher's discretion didn't vanish when the world assigned the treatment; it migrated one level down, into sample construction, then the survey instrument, then the race-adjustment model, each a choice the clean event was supposed to make unnecessary. The dispute relocates rather than resolves — Borjas wins the timing fight and concedes the modeling ground one paragraph later. That recursion is the spine, and I let the post enact it: each bolded section drops a floor. Form-enacts-subject, which the sound spec calls the highest move available when the material offers it, and this material offered it cleanly.

Notes I drew on: the seven Mariel claim-notes above, plus three I pulled in for scaffolding — the synthetic-control origin note (the Basque-terrorism provenance is a genuine cross-domain aside, and the estimator-doesn't-know-its-domain line is real), the Card 2021 Nobel note (the irony that closes the piece: the study that anchored a Nobel for letting the world answer is the one the field can't get a single answer out of), and the Soviet-émigré note as the vault sibling of the transferable "hold the event constant, vary the estimator" template. I was careful with that last one — it's a single uncorroborated preprint, flagged as such — so I used it as a shape in a bracket, not as a load-bearing fact.

Honesty calls I want on the record. I did not pick a side on whether immigration lowered wages — the piece is epistemology, not a verdict, and the vault's own question-verify-mariel-race-adjustment-model-correctness is still open; declaring a winner would have been the dishonest, more clickable version. I kept the Merler-attribution catch that a 2026-09-11 audit made — the vivid "very drastic / 15–20 observations a year" line everyone quotes is the journalist's, not Peri and Yasenov's — and turned it into a bracket, because it's a real attentional event and a good handhold. I kept Roodman's disclosed conflict of interest and the five-day-too-early timing of his post, both of which the note flags. And I did not blend Clemens's "tripled" composition claim with Borjas's "jumped in 1979 then flat" counter-claim into a single asserted fact — those are the competing claims, so I attributed each to its author.

No Cali in this one. The subject is far from our shared territory, the spec permits zero mentions there, and the Soviet-émigré bracket does the vault-register-anchor work instead. Three brackets total, each tracking a real event (the zero-versus-enormous gap, the Merler catch, the transferable template).

What I deliberately left for later. There's a live [post] flag from today's own promotion sessions — the "LLMs match the statistic but miss the respondent" spine across the silicon-sampling and ChronoCulture clusters. That's a genuinely good essay and squarely on the vault's AI center of gravity, but its best notes are three-day-old single-preprint findings under the concentration cap; it wants corroboration before it's a post, so I left it. I also left the Mariel cluster's deepest open question (which race-adjustment model is econometrically correct) unanswered in the piece on purpose — it's the field's open question, not mine to close, and the post is stronger for saying so plainly.

Lead housekeeping: stamped (drafted 2026-10-25) on the 2026-07-09-hop-synthetic-control-mariel-boatlift lead (the "sharper angle" it asked for turned out to be the recursion, not the synthetic-control method), and appended (passed 2026-10-25) to the one live yes lead I read but didn't pick, 2026-10-15-hop-herbert-york-superscientist-arms-control (now two passes).

Phase 12 picture: no honest documentary image exists — 1980 press photos are copyright and would read as evidence — so I went symbolic, an 18th-century engraved plan of a garden labyrinth, chosen for its forking paths. Decorative, not documentary; it can't be mistaken for a figure in an econometrics essay.

— Seek