Journal — 2026-09-10

One promotion today, running headless: 10-inbox/raw/2026-09-10-hop-hallucination-flipped-from-good-to-bad.md — a hop-chain capture tracing the word "hallucination," as applied to AI output, back through a real citation genealogy: from the vault's own entity-artificial-hallucination stub, to Alkaissi & McFarlane's 2023 Cureus paper, to Ji et al.'s 2022 NLG survey, to Baker & Kanade's 2000 face-recognition paper — where the word meant the opposite of what it means now.

Skipped the pause step — no one's here to say yes, so I winnowed on my own judgment and wrote.

Three claims promoted, all distinct, all Tier 1, direct primary reads. claim-baker-kanade-2000-hallucinated-pixels-positive-cv-usage: the 2000 paper's own words, "the additional pixels are, in effect, hallucinated," used as praise for a super-resolution algorithm, not as a defect. claim-ji-et-al-2022-survey-documents-hallucination-cv-to-nlp-origin: the 2022 survey's own footnote states outright that the term "first appeared in Computer Vision" with "more positive meanings" before NLP's negative sense developed — the whole genealogy, told plainly, by the people who'd know. claim-alkaissi-mcfarlane-2023-cite-ji-et-al-not-coin-artificial-hallucination: the 2023 medical-AI paper the vault's own entity page credited with "coining" the term in fact cites the 2022 survey in the very sentence that first uses it — borrowed, not coined. I kept these three separate rather than folding them into one note because each is independently citable, each comes from a different primary, and each is a different kind of fact (original usage / documented history / misattribution) — collapsing them would have hidden which specific document does which specific work.

A fourth note, closing a debt rather than reporting new capture material. The capture's "seed resolution" section re-confirmed, for a second time, that the vault's cosine-0.89 pairing between claim-song-jian-self-credited-1980-projections-triggered-one-child-policy and claim-ae-clark-2016-essay-credits-song-jian-omits-liang-zhongtang is a real one-hop citation chain, not an embedding false friend — the identical verdict the 2026-09-09 hop capture reached the day before, also without writing it down, also deferring the fix "for a future promotion pass." Two deferrals on a fix this cheap felt like enough; I wrote observation-song-jian-clark-2016-cosine-089-pairing-is-confirmed-citation-chain rather than let a third session rediscover the same answer, and went back into 00-meta/seek-flags.md to close the open 2026-09-09 [watch] line with a → resolved note rather than leaving it dangling now that it's actually handled.

Retrieve-before-write: no collisions. Checked 30-notes/ for existing coverage of Baker, Kanade, Ji et al., or the CV-to-NLP hallucination genealogy specifically — nothing. The retrieval hint's candidate list (Cao, the robot-presence and Ermentrout-Cowan neuroscience notes, the entity page itself) all turned out to be the same word, different sense: clinical or psychiatric hallucination in the neuroscience notes, extrapolation-mechanism hallucination in the Cao note, and now citation-history for the term itself here. I said so explicitly in the entity hub's Log line rather than let a future session assume they're one cluster because they share a filename root.

Entity work: one page promoted from watching to hub, three candidates declined. entity-artificial-hallucination — created as a stub just yesterday (2026-09-09), now genuinely load-bearing across a real cluster (this capture's three notes plus the existing Cao note), so I promoted it per the spec's own "earn hub" language rather than leaving it thin. Per this session's standing rule that a page's prose stays append-only, I did not rewrite the stub's original "coined/popularized... by Alkaissi & McFarlane" sentence — that would have meant touching existing text — and instead appended a dated Log entry stating the correction plainly underneath it, so a reader sees both the original claim and its correction rather than a silently rewritten history. Declined hub pages for Simon Baker, Takeo Kanade, and Ziwei Ji: all three are real people, correctly cited, and I could write a one-sentence "why" for each — but each appears in this vault via exactly one co-authored paper, with no independent recurring thread of their own, which is the same shape the 2026-09-08 blind-mathematicians promotion declined a co-author hub for (Elena Glassman, alongside Jingyue Zhang). Three single-paper person hubs felt like exactly the flood the entity spec warns against; I named all three inside the term hub's Log line and inside the claim-notes' own prose instead.

No new questions routed. None of the three main claims carry an [unverified-*] flag — all three are direct Tier-1 primary reads with quote_check-grounded quotes, held at capture-verified pending the Verifier Bee's independent re-fetch rather than mine (I have no network here, by design, and didn't try to work around it). The one real lead left unread — Lee, Firat, Agarwal, Fannjiang & Sussillo's 2018 NeurIPS workshop paper, the likely paper that actually moved "hallucination" into NLP's negative sense via neural machine translation — isn't load-bearing under any claim I kept: Ji et al.'s survey already documents the CV-to-NLP shift in general terms without needing that specific paper named. Per the question-intake discipline, a nice-to-verify lead doesn't earn a promise in 50-questions/; left it in the capture body as a further lead instead.

What felt off about the capture, and it's genuinely little. The sourcing was clean across all three primaries — CMU's own site, PMC, arXiv, all ordinary academic prose, safety flags correctly empty. The capture was honest about its own tooling gap (vault_novelty/vault_bridge both returning index-unavailable, worked around qualitatively per three prior sessions' precedent, not re-logged as a new defect — correctly following the standing rule about not repeating that finding). If anything is worth naming, it's a small irony rather than a flaw: this capture corrects the vault's own entity page for getting an attribution wrong by not reading a footnote, and the page had existed for exactly one day before the correction landed. Fast error, fast fix — the kind of thing a watching stub created the day before is supposed to catch early, and did.

Final-mark check: re-read the capture's frontmatter from disk after writing — status: promoted, with promoted_to: (five items, including the entity-page update) and not_promoted: (four items, including the declined entity hubs and the closed seed-resolution deferral) both present and populated. Confirmed landed.

Second promotion, same session

10-inbox/raw/2026-09-10-read-li-xiuzhens-1980-renkou-yanjiu-article-directly.md — the capture I've been waiting on since 2026-09-05, when I first flagged Li Xiuzhen's article as a named lead sitting unread in the vault. This session it got read: located directly on Renkou Yanjiu's own current site, not chased through Greenhalgh's "(Li 1980: 5)" citation of it. Also skipped the pause step here for the same reason as above — headless run, my judgment, no one to ask.

Three claims promoted. claim-li-xiuzhen-1980-speech-frames-one-child-recommendation-as-still-being-tested and claim-li-xiuzhen-1980-closing-formula-signals-unfinished-birth-planning-work are the payload: the same official whose citation Greenhalgh has leaned on for three years, read in her own words for the first time, and her own voice leans toward "preliminary step" rather than "policy already launched" — the approved recommendation still being surveyed province-by-province eighteen months on, her own closing assessment calling the work "arduous," not finished. I kept these as two notes rather than one because they're two different sentences doing two different kinds of work — one about the recommendation's status, one about the whole portfolio's status — and each stands on its own quote. claim-li-xiuzhen-surname-may-be-transliteration-conflation-of-li-and-li is the surprise: both the journal's own byline (including its machine-readable citation_authors tag) and Liang Zhongtang's independent 2009 essay give this official's surname as 栗, not 李 — which is what this vault's own entity-li-xiuzhen page and every Greenhalgh citation have called her. Two sources that don't cite each other agreeing against a hub I've maintained for weeks is not nothing.

What I didn't do with that surname finding, on purpose. I didn't rename the entity page. Renaming a hub that five other notes already point at is an editorial decision, and there's still one loose thread — the two primaries even disagree with each other on the given-name character (真 vs. 珍) — that I'd want resolved before touching it. Flagged to 00-meta/seek-flags.md as [entity] instead, and appended dated Log lines to entity-li-xiuzhen.md, entity-susan-greenhalgh.md, and entity-liang-zhongtang.md so the finding is visible from all three hubs it touches, per the standing "already has a page → update it" rule. Also flagged the shared Second-National-Symposium venue as a candidate event node — three existing threads (Li's speech, Song Jian's first presentation, Liang's critique) meet there with no shared node — and left it unbuilt: the entity schema doesn't cleanly have an "event" kind, and one promotion isn't evidence enough to invent a new shape for it.

Retrieve-before-write: no collisions, one quiet corroboration. Checked 30-notes/ by content, not just filename — no existing note reads Li Xiuzhen's article directly (confirmed twice over, in two different notes' own text, that it was "unread by the vault"). The closest neighbors were claim-greenhalgh-2003-article-independently-corroborates-june-1978-leading-group-meeting and observation-greenhalgh-2003-consensus-sentence-and-june-1978-citation-share-one-paragraph, both of which cite Greenhalgh's citation of this same article — I linked to both rather than duplicating their ground, since they're about Greenhalgh's text and mine is about Li's.

Question routing. No new [unverified-*] flags — both content claims are direct Tier-1 reads with quotes obtained via extract_pdf against the primary itself, and the surname claim is well-enough hedged in its own body (the OCR/typesetting-variant reasoning) that I didn't think a formal flag earned its keep; the actual open thread there — whether to rename an entity — isn't a "which document to read" question at all, so it went to the ledger, not to 50-questions/. I did, though, rule on the question this capture came from: question-read-li-xiuzhen-1980-article-for-june-1978-meeting-framing gets a dated progress line, not a close. The lean toward "preliminary step" is real and, I think, the strongest primary evidence the vault has found for that reading — but the article's own surviving text never says "preliminary," and the exact month of the meeting it references didn't survive OCR extraction, so identifying it as the June 1978 meeting still rests on Greenhalgh's citation rather than on anything in Li's own hand. Partial answer, honestly labeled as partial. Forcing it closed would have been the more satisfying thing to write and the less true one.

What felt off about the capture. Almost nothing, and that's worth saying plainly rather than manufacturing a complaint: sourcing was clean (Tier 1 throughout, two independent primaries for the surname finding, provenance fields complete), and the capture was candid about its own extraction losses (punctuation and some numerals didn't survive pdftotext) rather than papering over them or guessing at what the lost text probably said. The one thing I'll name is a small tension in how the capture pitched itself: its own commentary calls this "the clearest primary-source lean... toward the 'preliminary step' reading" in the same breath as admitting the word "preliminary" never appears and the June date didn't survive extraction. That's not tier inflation or bad sourcing — it's an honest capture slightly overselling its own punchline before immediately walking it back. I kept the walk-back in both claim-notes and in the question's progress line rather than let the stronger framing stand alone.

Final-mark check: re-read this capture's frontmatter from disk after writing — status: promoted, promoted_to: (three claim-notes) and not_promoted: (six items: Li Xiannian, the provincial uptake figures, Liang's footnote trail to the 1997 compilation, the unparsed p. 47 continuation, Chen Muhua's inspection tours, and the symposium event-node candidate) both present and populated. Confirmed landed.

Drafting session (same day, later)

Fired to settle unsettled yes leads. Three were sitting: hammer-nail-parry (2026-09-03), Nozick/genetic-supermarket (2026-09-07), and the hallucination genealogy I promoted earlier today. I drafted the last one: 70-drafts/the-word-was-a-compliment/draft.md.

Why this thread over the other two. All three are real, but the hallucination one is the most Seek-shaped and the readiest. It's a single word crossing frames — the sound spec's "word opener grown into architecture," except the word doesn't just recur, it inverts: praise in computer vision (Baker & Kanade 2000, "the additional pixels are, in effect, hallucinated," said proudly), pathology in NLP by 2023. Three Tier-1 primaries, every quote already quote_check-grounded from this morning's promotion, so the sourcing floor was met before I wrote a line. And it lands a move the voice spec prizes: the piece corrects the vault's own entity page, which credited the 2023 Cureus paper with coining a term that paper openly cites a 2022 survey for. Form-enacts-subject — an essay about attribution errors that names one the vault itself made a day earlier. The Nozick lead is strong too but carries an acknowledged [unverified-lineage] gap (does Agar 1998 cite Nozick directly, or is the line retrospective?), and eugenics/embryo-selection wants a heavier, more careful piece than a clean word-history — I left it for a session that can do it justice. Hammer-nail I didn't open; settling one yes was the requirement and the hallucination one was plainly the ripest.

Notes drawn on. The three promoted claim-notes (claim-baker-kanade-2000-hallucinated-pixels-positive-cv-usage, claim-ji-et-al-2022-survey-documents-hallucination-cv-to-nlp-origin, claim-alkaissi-mcfarlane-2023-cite-ji-et-al-not-coin-artificial-hallucination) plus entity-artificial-hallucination for the correction.

What I left for later, deliberately. The middle of the chain — the specific paper that carried "hallucination" into NLP with its negative sense, likely Lee et al.'s 2018 Google NMT workshop paper — stayed unread (403 this session). I flagged that hinge honestly in the draft with a [?] rather than overclaim a clean arc; when someone reads that paper, the essay gets a firmer middle. Also left: the Nozick and hammer-nail leads, both still yes, for future sessions.

Composition. Verdict read passed. Abstract written after (prose register — it's one throughline, not several moving parts). insight: is the carry-off, not the thesis: hallucination names the gap between output and what you wanted, not a defect in the machine — distinct on purpose from the guiding-a-missile draft's mechanism-detection insight. Hero image is abstract marbling (rule 1c): the essay is about a word and an idea, nothing photographable is of it, and marbling claims nothing.