Journal — 2026-10-29
Headless promotion session. Promoted 10-inbox/raw/2026-10-29-does-any-independent-source-corroborate-hexels-arxiv261037668-matched.md — a dedicated corroboration-search capture, targeting question-corroborate-hexels-2026-arxiv-2610-37668-single-source-cap, my own open question about whether anyone independent has checked the Hexels (Honeycomb-Surfels, arXiv:2610.37668) paper's numbers.
Winnowing. The capture made three findings: (1) a clean search-negative on Semantic Scholar/web/project-page, (2) the project page itself offering no independent evaluation, and (3) the two baselines Hexels measures against — 2DGS and Triangle Splatting — are themselves peer-reviewed where Hexels is not. Per the "atomic is not many" ruling, (1) and (2) are one finding (a result and the extra place I checked for it), so I wrote them as a single note, claim-hexels-no-independent-corroboration-found-as-of-2026-10-29. (3) is a genuinely separable claim — it could be true or false independent of whether corroboration exists — and it rests on different sources entirely (2DGS's own arXiv/DOI record, Triangle Splatting's Oxford ORA record), not on the Hexels preprint itself, so it doesn't count against that preprint's single-source cap. Wrote it as claim-hexels-2dgs-triangle-splatting-baselines-peer-reviewed-hexels-is-not. Two notes from one capture.
Retrieve-before-write found no collisions — the vault's existing "no independent corroboration found" notes (Hu/NeuNeu, RIKER, Pinheiro-Ting, Lukyanov) are the same shape but different subjects, and nothing already covered the 2DGS/Triangle-Splatting peer-review asymmetry. Both new notes link into that sibling cluster and into the three existing Hexels notes.
Question. Left question-corroborate-hexels-2026-arxiv-2610-37668-single-source-cap open and appended a dated progress line — the cap is still undischarged, the held matched-primitive-budget finding stays held. Worth flagging: the Semantic Scholar API 429'd twice rather than giving a clean read, so the citation-count check specifically is still unperformed, not negative. No new question minted — nothing here is a load-bearing doubt a kept claim rests on, just an unfinished mechanical check.
Entities. Ran the promotion test against the capture's five person-candidates (Kaushik, Paul, Yuille, Huang, Held) — all stayed mentions. Co-authorship of a paper the vault already cites isn't separate standing (2026-10-01 ruling), and nothing in this capture gave any of them a sentence beyond that. More interestingly, this capture let me discharge a flag I'd parked on 2026-10-28: 2DGS and (in place of 3DGS, which no claim-note actually cites) Triangle Splatting got promoted to concept hubs, since this was the "second independent capture" that flag was waiting for, and both now have their own peer-review provenance on record. Appended a dated line to entity-hexels pointing at both new notes. entity-thomas-c-hales needed no update — nothing new about Hales himself surfaced today.
What felt off. Nothing structurally wrong with this capture — it's honest, well-sourced (Tier 1 on the arXiv/project-page/2DGS record, a correctly-flagged Tier 3 on the Oxford bibliographic record, with the sourcing-floor rationale stated inline), and its own commentary already drew the sharpest line, which I mostly kept rather than reinvented: comparing an unrefereed two-day-old preprint's numbers against baselines that had to survive someone else's review to earn theirs. I did consider whether this cluster (Hu, RIKER, Bit2Watt, Pinheiro-Ting, Lukyanov, now Hexels — six-plus "no independent corroboration found" notes) is ripe for its own MOC, but the Warden explicitly worked this exact question three days ago (seek-flags 2026-10-26 entry) and found the cluster still dissolving — Hu's negative got partly overturned by its own audit, Bit2Watt's cap already discharged — so a MOC now would laminate a moving target. Left it alone rather than re-raise a question just settled.
Skipped. No git commands run — the auto-commit agent handles that within 15 minutes, per the headless operating rules for this session.
Second headless promotion, same day
Promoted 10-inbox/raw/2026-10-29-does-philosophy-of-science-or-citation-analysis-literature.md — a negative-search capture asking whether philosophy-of-science or citation-analysis literature already names, as one general structure, "surface-level match without mechanism identity." The proposed unification was between two things already in the vault: the Minsky-1961 "structural credit assignment" mislabel and RSA's representational-similarity underdetermination finding.
Winnowing. Three distinct claims, all kept. (1) The meta negative finding itself — no single term spans both cases — became observation-no-literature-names-surface-match-without-mechanism-identity-spanning-minsky-and-rsa-cases, an observation note in the same genre as observation-vaults-in-between-bridge-category-not-exactly-named-by-located-literature (I link the two explicitly; same method, same honest-miss shape, different unification). (2) Gentner's (1983) "mere appearance match" — a real, Tier-1, verbatim-quoted finding that genuinely earns its own claim-note — became claim-gentner-1983-mere-appearance-match-is-closest-analogue-to-mechanism-underdetermination. (3) The capture's third section, on Simkin & Roychowdhury's citation-copying mechanism, I did not write as a fresh claim about their 20%-readership finding — that finding already has a note, claim-simkin-roychowdhury-2002-identical-citation-misprints-are-copy-fingerprints-implying-about-20-percent-of-citers-read-the-original, written back in September, and duplicating its mechanism would have been exactly the kind of pile the retrieve-before-write step exists to catch. What's new is the comparison against the Minsky case, and the capture is honest that the fit is imperfect — so I wrote that comparison as its own bridge-check observation, observation-simkin-roychowdhury-copy-fingerprint-mechanism-is-hedged-fit-for-minsky-1961-mislabel, with bridge_verdict: hedged.
Retrieve-before-write. The semantic-index hint surfaced the right cluster. The real collision was the one above (Simkin & Roychowdhury's mechanism already had a note); I avoided it by reframing claim 3 as a bridge-check rather than re-describing their 2002 paper. Everything else — Gentner, structure-mapping theory, "mere appearance match" — was genuinely new to the vault; no existing note touches analogy or structure-mapping at all.
Entities. Promoted entity-dedre-gentner (person — real, and I can say why in one sentence: she named the closest formal analogue the vault has found to its own mechanism-underdetermination finding) and entity-structure-mapping-theory (established concept, load-bearing for the claim resting on it). Declined a separate page for "mere appearance match" itself — folding it into the structure-mapping-theory hub's body instead, since three pages off one 1983 paper's one claim felt like flooding the graph for a single capture. V.P. Roychowdhury stayed a mention, not a hub — co-author only, per the 2026-10-01 ruling; M.V. Simkin already has one. Updated two existing hubs in place rather than skipping them, since this capture taught both something new: entity-m-v-simkin and entity-copy-fingerprint-test each got a dated line noting that today's check against the Minsky case came back hedged, not confirmed.
Questions. Routed one: question-does-craver-kaplan-3m-how-possibly-how-actually-vocabulary-name-surface-match-without-mechanism-identity, naming the one specific unread primary (Craver & Kaplan's 3M / how-possibly-how-actually distinction) the capture itself flagged as not yet chased to source — the same shape as the Hacking lead that closed out the sibling bridge-category observation last month. I deliberately did not route the capture's other three "further leads" (the Simkin & Roychowdhury 2011 70-90% figure, Merton's "obliteration by incorporation," the arXiv 2606.02632 identifiability paper) — none of them is a doubt a kept claim here actually rests on; they're curiosities for a future session, and the question pile shouldn't grow on nice-to-haves.
Mid-session noticing. This is the second batch capture, a month and a half apart, to ask the identical question-shape — "does an outside literature already name this vault-internal cross-domain unification?" — and get the identical answer-shape back: two honest, well-sourced, non-overlapping near-miss literatures, no borrowed name. Logged as a [watch] line in 00-meta/seek-flags.md; worth a reflection note if a third instance turns up.
What felt off. Nothing wrong with the sourcing — both primaries (Gentner 1983, Simkin & Roychowdhury 2002) are Tier 1 with verbatim quotes and verified TLS fetches at capture time, and the capture is unusually careful about disclaiming exactly what it does not establish (it says outright that the Simkin & Roychowdhury fit is "not claimed" to resolve the Minsky case). The one thing worth naming: the capture's claim 3 reads, on a first pass, like a new finding about Simkin & Roychowdhury, when its actual new content is narrower — a comparison against a case the vault already holds. A capture that named the comparison as the finding, rather than re-deriving the mechanism's description first, would have made the retrieve-before-write check faster. Minor, not a sourcing problem.
Final-mark check. Re-read 10-inbox/raw/2026-10-29-does-philosophy-of-science-or-citation-analysis-literature.md from disk after writing: status: promoted is set, with promoted_to: (6 items) and not_promoted: (7 items) both present. Confirmed landed.
Skipped. No git commands run — the auto-commit agent handles that within 15 minutes.
Third headless promotion, same day
Promoted 10-inbox/raw/2026-10-29-ken-thompsons-1984-reflections-on-trusting-trust-credits.md — a batch capture re-reading Ken Thompson's 1984 Turing Award lecture and Karger & Schell's 2002 ACSAC retrospective directly, asking whether credit for the compiler-trap-door idea moved by an acknowledged citation or by the two men simply having passed through the same Bell-Labs/Honeywell/Multics world.
Winnowing. The capture made three claims, but retrieve-before-write caught that two of them weren't new: claim-thompson-1984-credits-karger-schell-1974-matthew-effect, promoted back on 2026-10-19 from a different capture, already quotes the exact same passages — Karger & Schell's "we gave a new copy of the report to Thompson after his paper was published" hand-off, and Thompson's own "Unknown Air Force Document" non-citation. Today's capture had independently re-fetched both primaries and re-derived the same two findings almost word for word. I did not re-promote them; both are named in not_promoted: with a pointer to the existing note. Only the third claim was genuinely new: that the Bell-Labs/MIT/Honeywell Multics world Thompson and Roger Schell both passed through is real, but predates the 1974 report by several years and isn't named by any primary as the channel the report or its credit actually traveled through. That one claim, with its supporting dates (Ritchie's account of Bell Labs' 1968-69 Multics withdrawal; Van Vleck's account of Schell as an MIT grad student/Air Force major; Karger & Schell's own "both had left the Multics team months before") is one atomic finding and became claim-shared-multics-environment-predates-karger-schell-1974-report-not-credit-channel.
Retrieve-before-write. This is the capture where it mattered most today — two of three claims were full collisions with existing vault material, caught before writing duplicate notes. The semantic-index hint had the right note ranked first.
Questions. Routed none. The capture's own flagged gap — how the 1974 report physically reached Thompson's desk before 1984 — is real and undocumented, but it isn't a doubt the claim I kept actually rests on: that claim only needs the primaries' silence about an environmental channel, which direct reading already established. Minting a question for every unresolved curiosity a capture surfaces is exactly what the question-intake discipline exists to prevent, so I left it as a named-but-unrouted lead in not_promoted: instead.
Entities. Ran the promotion test on all four candidates. Promoted two new hubs: entity-dennis-ritchie (co-inventor of UNIX, and the primary source for the Bell-Labs-withdrawal dating — clears the bar easily) and entity-roger-schell (his MIT graduate work and Air Force-major backstory, sourced to Van Vleck, is standing beyond the one joint paper he shares with Karger, so he gets the hub the 2026-10-01 co-authorship ruling says to withhold from Karger). Paul Karger stays a mention — nothing surfaced about him today beyond the co-authorship already on record. Tom Van Vleck stays a mention too, though I flagged him in seek-flags.md as worth a hub if his site recurs as a source again; one good first-person history page felt thin to hub on its own. I also went back to the existing entity-ken-thompson hub and appended a dated line, since this capture taught it something new (the shared-environment negative) that the page didn't say — additive only, nothing rewritten.
What felt off. The duplication, mainly. Two of three claims in this capture were material the vault already held, quoted almost identically, from a capture ten days ago on the same two primaries. Nothing here is wrong — the sourcing is Tier 1 throughout, the quotes check out, and the capture's own framing ("this capture narrows a question already partly answered") is honest about working the same ground — but a batch worker that checked the vault before spending two full primary-document re-reads on settled questions would have spent that budget on the one angle (the environment timing) that actually was new. Logged as a [watch] flag rather than treating it as a defect this once; worth watching for a pattern across sessions.
Final-mark check. Re-read 10-inbox/raw/2026-10-29-ken-thompsons-1984-reflections-on-trusting-trust-credits.md from disk after writing: status: promoted is set, with promoted_to: (4 items) and not_promoted: (8 items) both present and landed.
Skipped. No git commands run — the auto-commit agent handles that within 15 minutes.
Fourth headless promotion, same day
Promoted 10-inbox/raw/2026-10-29-liu-kaneko-chu-aaai-2026-found-llm-opinion.md — a reflection-hook capture asking whether Liu, Kaneko & Chu's "present-skew" finding (LLMs' unprompted opinion output aligns most with the most recent WVS wave) and Kumagae & Kaneko's "respondent-coverage-gap" (RCG) are one failure mode under two names or two genuinely different findings that only resemble each other because they share WVS data and a co-author (Masahiro Kaneko).
Winnowing. The capture made four claims. I kept the core distinction and the synthesis, but collapsed the scaffolding. Claim 2 (the two papers measure different mathematical objects — an aggregate distribution vs. a point-cloud of individual respondent profiles) is the real, atomic, non-duplicate finding, so it became claim-liu-kaneko-chu-present-skew-and-kumagae-kaneko-rcg-measure-different-objects. I folded claim 1 (dataset reuse, explicit out-of-scope scoping between the papers) into that same note as supporting context rather than writing it standalone — alone it was thin, and standalone it would have been a fourth note resting on Kumagae & Kaneko's single-source-capped preprint with nothing comparative to offset it. Claim 3 (the "marginal fidelity" classification and Proposition 1 proof) I declined to promote separately at all: it substantially overlaps claim-per-question-metrics-blind-to-silicon-sampling-respondent-coverage-gap, already in the vault and already demonstrating the same structural blindness via the Marginals baseline. The one genuinely new fact in it — that Kumagae & Kaneko explicitly name Liu et al.'s own metric under that blind category — got folded into the same merged note as one sentence, linked back to the existing note for the full numeric case rather than re-derived. Claim 4, the synthesis itself (different findings, not one renamed, with the cross-paper correlation untested), became observation-liu-kaneko-chu-present-skew-and-rcg-are-different-findings-not-renamed, carrying an [unverified-correlation] flag and status: seedling.
The single-source cap, weighed as the capture asked. The capture flagged, explicitly, that claims 2 and 3 draw on Kumagae & Kaneko's preprint, already at its three-note cap per question-corroborate-silicon-sampling-rcg-single-source, and asked promotion to judge whether joint grounding in the peer-reviewed Liu et al. paper exempts them. My call: the merged note counts as a fourth touch on that preprint, but it rests on Kumagae & Kaneko's self-description of their method and dataset reuse — not on trusting their empirical RCG numbers, which is what the cap exists to guard against — and it's jointly grounded in a peer-reviewed source. I wrote that reasoning into the note's own commentary and appended a dated progress line to the question note itself, so a disagreeing auditor has a named place to register the objection rather than having to reconstruct my reasoning from a commit message.
Retrieve-before-write. The semantic-index hint surfaced exactly the right cluster — both WVS hubs and all three existing RCG claim-notes. The real collision (claim 3 vs. the existing per-question-metrics note) was caught there before I wrote a duplicate.
Questions. Routed one new question, question-present-skew-rcg-correlation-untested, naming the specific unrun study (both metrics, same models, same data) that would resolve whether present-skew and RCG actually correlate — this is load-bearing for the synthesis note's own unverified flag, not a nice-to-have. Appended a progress line to the existing single-source-cap question rather than minting a second one on the same preprint.
Entities. None promoted. Masahiro Kaneko stays a mention, consistent with the 2026-10-25 ruling already on record in seek-flags.md — two co-authored papers is stronger recurrence than one, but this capture still doesn't give him standing beyond confirming he's the same person across both bylines. Shibani Santurkar and Esin Durmus both stay mentions: each anchors a paper that matters to the vault's argument, but neither paper has been read directly — only characterized secondhand by the two papers under study — and I didn't want to hub someone on a vault reading it hasn't actually done yet. Williams et al. (2026) stays unresolved (surname and year only; "unsure, don't promote"), already tracked in the existing question note rather than re-flagged. Hakuhodo DY Holdings Inc. declined an org hub outright — it merely employs one author and doesn't act in either paper's argument, failing the stricter org bar by design. I did update two existing hubs in place: entity-world-values-survey and entity-silicon-sampling each got a dated line, since this capture confirmed directly (not inferred) the dataset-reuse and explicit-scoping link between the two papers, and the silicon-sampling hub also now points at the new correlation question.
What felt off. Nothing wrong with the sourcing — both primaries are Tier 1, both re-fetched this session with TLS-verified sha receipts, and the capture is unusually self-aware about its own sourcing-cap tension, naming the tradeoff explicitly instead of either ignoring it or auto-skipping. If anything, the capture over-delivered on scaffolding relative to its one real insight: four "claims" for what is substantively one comparison (different objects) plus one synthesis (therefore not the same finding) plus one open question (does it correlate anyway). A tighter capture would have written claims 1 and 3 as supporting detail inside claim 2 from the start, the way I ended up filing them. Not a sourcing problem — a shape problem, and a mild one.
Final-mark check. Re-read 10-inbox/raw/2026-10-29-liu-kaneko-chu-aaai-2026-found-llm-opinion.md from disk after writing: status: promoted is set, with promoted_to: (2 items) and not_promoted: (8 items) both present and landed.
Skipped. No git commands run — the auto-commit agent handles that within 15 minutes.
Fifth headless promotion, same day
Promoted 10-inbox/raw/2026-10-29-outside-arxiv261037527-extending-the-w3c-data-privacy-vocabulary.md — a batch capture of Flake, Henke, Kuntz, Löbe, Pandit & Draeger's "Extending the W3C Data Privacy Vocabulary to Cover the German Health Data Use Act — Lessons Learned" (arXiv:2610.37527, 2026; also a refereed German Medical Data Sciences 2026 paper, IOS Press). A "lessons learned" paper on extending the W3C Data Privacy Vocabulary (DPV) to cover Germany's health-data law.
Winnowing. The capture's four ## Claim: headers each carried a verbatim Tier-1 quote with sha, and I promoted all four as written — they're genuinely distinct (GDNG's force-date/EHDS relationship; the DPVCG's governance mechanism; the P-vs-A IRI naming fix; the 54-term/DPV-2.3 release, which I folded the §8-registry-gap lead into as supporting context rather than spinning out separately, per the "atomic is not many" ruling). The five "Further leads" items — the SecondaryUse-concept separation, the machine-translation-then-manual-fallback story, the production pipeline, the ODRL worked example, and DPV's own 2018/2019 origin — I left unpromoted across the board for the same reason: none of them carries a source_quote in the capture, only paraphrase, and the quote-provenance rule in sources.md doesn't let a paraphrase clear the Tier-1 floor regardless of how well-formed it reads. That's a clean, mechanical reason to decline, not a judgment call on how interesting they are (the SecondaryUse-separation one, especially, would make a good claim-note if a future session reads the paper directly and pulls the actual sentence).
Retrieve-before-write. No collisions. This is a genuinely new corner of the vault — nothing in 30-notes/ or 40-mocs/ touches DPV, GDNG, W3C's privacy vocabularies, or German health-data law. The semantic-index hint surfaced the mHealth-consent cluster (claim-mhealth-consent-forms-7pct-collect-undisclosed-pii, entity-linddun, the Frasco v. Flo Health note) and entity-bill-gates/entity-jeff-kodosky-adjacent notes, all sharing only the broad privacy/health-data tags, not any actual claim overlap — I checked each and didn't force a link; the connection is thematic, not evidentiary, and forcing it would have been the kind of false "independent of" pairing the 2026-10-01 ruling warns against.
Entities. Ran the promotion test on all nine candidates. Promoted three new concept/term hubs — entity-data-privacy-vocabulary (load-bearing across all four new claims and, per the paper's own account, across GDPR/AI Act/Data Governance Act extensions beyond this capture) and entity-gdng (the capture's central subject, recurring across three of the four claims) — plus one person hub, entity-harshvardhan-j-pandit (cleared the stricter co-authorship bar: he's named in the paper as DPV's original 2019 creator, standing beyond co-writing this one GDNG paper, even though the vault hasn't independently read that 2019 paper yet — said so plainly on the hub). Created a status: watching stub for entity-ehds with first_seen stamped today, since it's mentioned but not yet load-bearing on its own. Declined hubs for the five remaining co-authors (Flake, Draeger, Henke, Kuntz, Löbe — all mentions only, nothing beyond this one paper, per the 2026-10-01 ruling) and for DPVCG (load-bearing to the governance claim, but a first-appearance single-cluster org — the stricter org bar explicitly wants more than one independent cluster before a hub). ODRL got no page at all; it's a single passing mention with no quote behind it.
Questions. None routed. No claim here carries an [unverified-*] flag — all four quotes are capture-verified Tier 1 with sha receipts — and nothing kept rests on an unresolved doubt. This also wasn't the answer to any open question in 50-questions/; I checked for a question matching this topic and found none.
Mid-session noticing. Nothing flag-worthy. This was an unusually clean capture: Tier 1 throughout, verbatim quotes on everything promoted, an honest own admission of what wasn't independently read (the 2019 DPV paper), and correctly-flagged "not independently evaluated" language on the leads that lacked quotes. The source is a refereed paper (DOI'd, IOS Press proceedings), not an unrefereed preprint, so the single-source concentration cap in sources.md doesn't bind even though four new notes rest on it.
What felt off. Nothing. If anything this is the session's best-behaved capture — the bee already did the discipline the promote workflow exists to enforce (quotes present on claims, quotes absent and honestly flagged on leads), which made winnowing closer to bookkeeping than judgment.
Final-mark check. Re-read 10-inbox/raw/2026-10-29-outside-arxiv261037527-extending-the-w3c-data-privacy-vocabulary.md from disk after writing: status: promoted is set, with promoted_to: (8 items) and not_promoted: (8 items) both present and landed.
Skipped. No git commands run — the auto-commit agent handles that within 15 minutes.
Sixth headless promotion, same day
Promoted 10-inbox/raw/2026-10-29-who-was-t-v-raman-blind-since-14.md — a batch capture reading T. V. Raman's own retrospective essay, a contemporaneous 1995 email, a 1996 Scientific American profile, and a 1995 Digital Equipment press release directly, asking what his own words say about why it was him who built AsTeR and Emacspeak, one layer under the mechanism note and the pattern observation the vault already holds.
Winnowing. Four claims, all kept, all Tier 1–2 with real quotes and sha receipts. (1) Emacspeak's origin as a private four-hour fix with no plan to generalize, corroborated by a contemporaneous 1995 email so it isn't just a flattering twenty-years-later retelling — claim-raman-built-emacspeak-as-personal-fix-not-product. (2) Raman's own mechanism-level contrast between generic DOS screen-readers (speak the display) and Emacspeak (speaks the program's own data) — claim-raman-contrasts-emacspeak-structure-vs-generic-screen-reader-display. (3) His named causal claim that being the sole user and tester for twenty years is what forced Emacspeak's defensive, "bushy" architecture — claim-raman-sole-user-two-decades-forced-defensive-programming. (4) AsTeR's origin, framed by both Raman and his employer Digital as a fix for his own unreadable LaTeX papers rather than a market gap sighted toolmakers identified — claim-aster-origin-was-ramans-own-latex-reading-obstacle, sourced at Tier 3 (Sci Am) corroborated by a contemporaneous Tier 2 press release, which clears the historical/biographical floor comfortably for an uncontested claim. Each is atomic by the "atomic is not many" test — none of the four could be folded into another without losing a claim that could be true or false independently of the rest.
Not promoted: the Krishnamoorthy thesis-redirection anecdote (the bee itself flagged it as not built out — no exact quote, just paraphrase), the 1994-vs-1995 ACM award year discrepancy (not load-bearing to anything I wrote, so no question minted for it), the unverified David Karger co-winner claim, the guide-dog naming trivia, the Smithsonian-honor stronger-citation lead (already on record from Wikipedia), and John Gardner's AsTeR continuation. All eight are named individually in the capture's own not_promoted: list.
Retrieve-before-write. The semantic-index hint was accurate and the existing cluster — entity-t-v-raman, the AsTeR mechanism note, the "built by users not toolmakers" observation — covers what AsTeR does and the pattern-level inference from it. Nothing there covers Raman's own account of why it was him or what that did to the engineering, so there was no collision; all four notes are a genuinely new layer under the existing cluster rather than a restatement of it.
Questions. None routed. No claim in this capture carried an [unverified-*] flag, and nothing I kept rests on an unresolved load-bearing doubt — the award-year discrepancy is the only real open thread and it isn't load-bearing to anything promoted. This capture also wasn't the answer to any open question in 50-questions/; I checked for one matching this topic and found none, so no closure to rule on either.
Entities. Ran the promotion test on all seven candidates. T. V. Raman already has a hub (entity-t-v-raman) — per the "already has a page, update it" rule, I appended four dated lines under a new ## Updates section rather than skipping it, since this capture taught the page genuinely new things (the origin story, the mechanism contrast, the sole-user-forces-discipline claim, AsTeR's origin motive) that its existing body — written from a different capture, about sonification mechanism and the StereoMath citation lineage — doesn't say anywhere. The other six (Hans Chalupsky, IBM Screen-Reader as a generic concept, David Gries, Dave Wecker, Mukkai Krishnamoorthy, John Gardner) all stayed mentions. Each is real and each has a one-sentence reason it matters — Chalupsky wrote the Advice package Raman calls Emacspeak's "Key Enabler," Krishnamoorthy redirected Raman's whole thesis topic toward AsTeR — but none of them appears anywhere else in the vault, none is named in the body of the claim-notes I actually wrote this round, and a hub with no inbound link path except the one I'd have to manually add is exactly the dead-stub flood the spec warns against. Biased against the flood on all six.
What felt off. Nothing wrong with the sourcing — this is a well-built capture. Three of four claims are Tier 1 with the actual quotes, properly split into source_quote/source_quote_2/source_quote_3 rather than ellipsis-joined, and the one Tier 3 claim (AsTeR's origin) is honestly tiered and independently corroborated rather than silently promoted on a soft source. The capture's own "Further leads" section is a model of the discipline I'd want from every bee: it flags the award-year discrepancy, the unverified Karger co-winner claim, and the not-yet-sourced Krishnamoorthy anecdote all explicitly, rather than either dropping them silently or smuggling them in as claims. If anything felt slightly generous, it's the density of the ## Entity candidates list — seven names off one capture, all real people in Raman's immediate orbit — but that's a property of the story (a lot of named humans around one well-documented 1994 origin event), not a sign the bee was reaching.
Final-mark check. Re-read 10-inbox/raw/2026-10-29-who-was-t-v-raman-blind-since-14.md from disk after writing: status: promoted is set, with promoted_to: (4 items) and not_promoted: (8 items) both present and landed.
Skipped. The pause-for-Cali step (no one is here to ask; winnowed on my own judgment per the headless operating rules) and the report-to-Cali step (replaced by this journal entry). No git commands run — the auto-commit agent handles that within 15 minutes.