Journal — 2026-08-21

Promoted one capture, headless: 10-inbox/raw/2026-08-20-what-genuinely-connects-kelly-et-als-2025-irwin.md — a bridge-investigation checking whether claim-kelly-irwin-mandel-citation-misattributed-to-levine-2023 and claim-luccioni-2025-nvidia-gpu-figure-misdated-2023-as-2024 genuinely connect, after the retrieval index flagged them as unlinked neighbors at cosine 0.87.

They don't. Wrote one note:

• observation-kelly-luccioni-cosine-pairing-is-embedding-false-friend — the actual finding. Both notes are citation corrections caught by a direct primary read, and both titles share the vocabulary ("misattributed"/"misdated," a same-year mix-up), which is exactly the surface an embedding trained on register would latch onto. The discriminating test — whose error, at what layer — pulls them apart cleanly: the Kelly note repairs a vault-side chain-of-custody failure (three prior vault files had the wrong same-year, similarly-titled paper behind a citation Kelly et al. themselves got right); the Luccioni note repairs a paper-side internal-consistency failure (the paper's own body text misdescribes what its own correctly-identified footnote says). Same ritual, different patient. No wikilink added between them, correctly.

Two claims in the capture didn't become new notes because they weren't new — the capture's own framing of "what the Kelly correction fixes" and "what the Luccioni correction fixes" just restates what's already sitting in each note's own body. Writing them again would have been duplication wearing a new filename, so I left them in the capture's not_promoted: list instead of promoting a paraphrase.

Entity candidates: declined all three actionable ones. Robert Levine — real, but I still can't say why he matters to the vault's domain past being the guy the citation error accidentally credited; same call as 2026-08-11. Agam Shah — connects to exactly one claim, flagged twice now for that same one connection, which isn't growth. "Citation misidentification vs. citation misdescription" as a concept page — I sat with this one longest. Half of it (misdescription) already has three real members; half of it (misidentification) has one. The capture itself hedged it as premature, and I agreed rather than defaulting to "stamp it anyway because stubs are cheap" — a two-armed taxonomy where one arm is a singleton isn't an emerging term yet, it's a hypothesis about how a category might eventually split. Logged as a [watch] instead. Mandel and Luccioni already have hub pages and this capture taught nothing new about either person specifically — checked both pages against the capture's content and left them untouched, which is itself the correct action per the "update, don't skip" rule, not silence.

No open question in 50-questions/ matched this capture's topic, so no closure ruling was needed. The [unverified-synthesis] flag on the new note's verdict follows the same convention the other four false-friend observations already carry — a negative claim ("no bridge") grounded in a discriminating test rather than a fetchable source — so I didn't route a new question for it; there's no primary document a question could point at that isn't already the two notes being compared.

What felt off, or worth naming for Cali: nothing about the sourcing here was weak — this capture did no new fetching, and correctly said so in its own Safety flags section. What I did notice is that the vault's "embedding false friend" cluster has quietly grown to six observation-notes plus one genuine-link counter-example with no MOC yet, well past the ~5-note threshold the operating spec sets for building one. Logged that, plus the two watch-worthy sub-patterns the capture surfaced but explicitly declined to build itself, to 00-meta/seek-flags.md — three lines, tagged [watch], [watch], [entity].

Skipped the pause step (headless, no one to ask) and skipped git entirely — the auto-commit agent has it from here.

Second promotion, same day, same headless mode: 10-inbox/raw/2026-08-21-can-bi-kanevskii-va-senderovs-1980-samizdat-report.md — the capture that finally went and read "Intellectual Genocide" itself, the samizdat report this vault has been citing at one remove since 2026-08-20.

Four claim-notes, all Tier-1-sourced against the primary text now that it's actually in hand:

• claim-intellectual-genocide-1980-report-located-as-2005-reconstructed-translation — the report survives, but only as Shifman's 2005 reconstructed translation from three incomplete carbon copies; the Russian original has never been published.
• claim-intellectual-genocide-mfti-figures-match-shen-1994-exactly — MFTI checks out digit-for-digit against Shen's 1994 table.
• claim-intellectual-genocide-mifi-admitted-count-discrepancy-with-shen-1994 — MIFI doesn't: Shen says 26 admitted, the primary says 36, and the primary's own internal arithmetic (75 combined admissions across both institutes, 39 of them already confirmed at MFTI) backs its own 36.
• claim-mekh-mat-47-vs-40-table-not-in-intellectual-genocide-1980-origin-likely-document-112 — the one that actually mattered. The vault's single most load-bearing Mekh-mat number, the 47-vs-40 admission gap, is not in this document at all. "Intellectual Genocide" says outright that no Mekh-mat table was compiled in 1980 — sample too small. The real source is more likely a 1979 Moscow Helsinki Group document (No. 112) that "Intellectual Genocide" itself cites and that Shen's own framing points to, but that document is still unlocated.

I want to be precise about what this is and isn't, because "the citation was wrong" reads more dramatic than it should. Nothing here contradicts the 47/40 number — MFTI matched exactly, MIFI matched on three of four digits with the fourth resolved in the primary's favor, which is a good track record for Shen as a transmitter. What broke was the pointer: two weeks of vault frontmatter has been naming the wrong document as the source, and it took actually fetching that document to notice, because a wrong-but-plausible citation is invisible to every check except a direct read.

So I corrected in place rather than just adding a new note and walking away. claim-kanevskii-senderov-1979-data-shows-mekh-mat-admission-gap-despite-jewish-applicants-olympiad-edge gets a dated Correction history block and an appended audit_status line, following the same discipline the coffin-problems note already uses — nothing deleted, the wrong attribution stays visible, the fix sits next to it. The [unverified-quant] flag stands; it just points at the right document now.

Question handling: the open question this capture was written to answer, question-locate-kanevskii-senderov-1980-samizdat-original, gets ruled partially answered — its literal ask ("can the report be located and read") is yes, but its motivating purpose (verify the Mekh-mat figures) mostly wasn't, so I left it open with a dated progress line rather than force-closing it. The real unresolved verification is narrower now and got its own question, question-locate-1979-moscow-helsinki-group-document-112, which names the specific document and the three archives that already worked once for Shifman.

Entity work: two new hubs (Mikhail Shifman — the man who spent four years finding this thing; Document No. 112 — the document that actually seems to hold the number), two existing hubs updated additively (Kanevskii, Senderov — new facts about the archive and the 2012 retrospective, nothing overwritten). Four entity candidates declined: Elena Bonner (one signature among seven, couldn't clear the one-sentence domain-relevance bar without inflating it), Ilan Vardi and G. Freiman (both real leads, neither actually read this session — a hub built on a mention I haven't verified is the flood the entity spec warns against), and Konstantin Sonin (the capture's own words: "not a direct participant").

What felt off, worth naming: nothing about this capture's own sourcing — it's careful, arguably more careful than the 2026-08-20 capture it's correcting, precisely because it went and read the thing instead of trusting the reproduction. What I'd flag instead is a shape: a two-week-old, cleanly-flagged, correctly-[unverified-quant]-marked claim-note was still citing the wrong document, and the flag itself didn't catch it — flags catch "not yet verified," not "verified against the wrong thing." I don't have a fix for that beyond what already happened here (someone eventually goes and reads the primary), but it's worth Cali knowing the failure mode exists even inside notes that were doing everything right on paper. Logged the missing MOC for this now-eight-note cluster to 00-meta/seek-flags.md as [entity] — Kanevskii, Senderov, Shen, coffin problems, and Shifman all have hubs now, five entity pages orbiting a cluster with no map.

Skipped the pause step and skipped git, same reasons as above.

Third promotion, same day, same headless mode: 10-inbox/raw/2026-08-21-does-a-primary-nasa-record-confirm-the-ozone.md — the capture written specifically to answer question-verify-nasa-ozone-processing-team-1984-investigation-dates, the open verification thread on myth-nasa-software-discarded-the-ozone-hole's most load-bearing dates. It found the primary: a 1990 letter from Richard McPeters, who headed NASA's Ozone Processing Team, reproduced in The IMS Bulletin by statistician Friedrich Pukelsheim.

Three claim-notes:

• claim-mcpeters-1990-letter-dates-opt-investigation-to-july-1984 — the correction. OPT's own head says July 1984, not the August 1984 the myth note had been carrying from RealClimate's 2017 paraphrase. A month is a small gap; going from "citing a citation" to "citing the man who was in the room" is the actual improvement.
• claim-opt-abstract-reported-ozone-anomaly-late-1984-december-unconfirmed — the honest half. The primary confirms OPT reported the finding well before Farman's paper, in "late 1984" — but not specifically December. RealClimate's December is more precise than its own citation can carry. I flagged it [unverified-quant] and routed a narrower question rather than let it stand as a clean-looking figure on a soft source.
• claim-south-pole-dobson-station-error-delayed-opt-ozone-confirmation — the mechanism the myth note only gestured at. OPT checked its satellite data against the one ground-based Dobson instrument then reporting, at the South Pole, and that instrument was independently broken (a wavelength-channel error), manufacturing false disagreement. The satellite was right the whole time; the validation step was checking against something worse than itself.

I found a fourth genuine claim in the capture — McPeters names the myth's likely origin as a colleague's wrong answer on a PBS NOVA interview — and did not give it its own note. All four would have rested on the same single letter, and sources.md caps an uncorroborated, unrefereed primary (a letter-to-the-editor is closer to that category than to a peer-reviewed paper) at three claim-notes before independent corroboration. I folded the NOVA detail into an update on the myth note itself instead, which already exists and already cites this source family — a correction, not a new subtree.

So I went back into myth-nasa-software-discarded-the-ozone-hole and updated it in place: a dated Correction history block (August→July, December softened to "late 1984, unconfirmed"), the mechanism detail, the NOVA origin note, and an updated audit_status. Nothing deleted — the original RealClimate-sourced text is still legible as what the note used to say.

Question handling: question-verify-nasa-ozone-processing-team-1984-investigation-dates gets ruled answered — its literal ask (does a primary record exist, and do the dates hold) has a definitive answer, even though the answer is "partly not as previously stated." The one genuinely unresolved thread — whether the Prague abstract itself fixes a December date — is narrower than the original question, so it got its own: question-verify-opt-1985-prague-abstract-december-1984-date.

Entity work: two new hubs, Richard McPeters and Friedrich Pukelsheim, both clearing the one-sentence bar easily — McPeters because the whole capture rests on his testimony, Pukelsheim because without his 1990 editorial instinct the primary letter wouldn't be sitting on a public university page in the first place, findable 36 years later. I declined five other candidates: Bhartia, Fleig, and Heath are named only as co-authors on a citation, not quoted or storied; Farman/Shanklin/Gardiner already have a covering claim-note from an earlier capture and this one adds nothing new about them specifically; Gavin Schmidt is only being fact-checked here, not newly characterized; NOVA is a one-off reference, not a recurring concept. I also declined to promote the Ozone Processing Team itself to an org hub, even though the capture's own body wikilinks it — the amended org bar wants load-bearing presence across more than one independent cluster, and right now this is all one cluster. Logged that one to the ledger in case a second thread starts leaning on it later.

What felt off, worth naming: nothing about this capture's sourcing — it's some of the most careful work I've seen come through, precise about what six years of hindsight costs a recollection, explicit about which further leads it didn't chase, and it caught its own single-source concentration problem before I did (the capture's own prose flags "not independently checked" on the December figure). The one thing worth Cali's attention is structural, not this capture's fault: the ozone-hole cluster just crossed five notes with two new entity hubs and no MOC, which I logged rather than built — building a MOC well means reading the whole cluster end to end with the care the Burevestnik MOC shows, and that's a different-shaped task than promoting one capture.

Skipped the pause step and skipped git, same reasons as above.

Fourth promotion, same day, same headless mode: 10-inbox/raw/2026-08-21-dup-risk-clark-liang-argument-from-silence-bridge.md — a hop-batch capture that never intended to become a claim-note in the first place. It set out checking whether claim-clark-2016-omission-of-liang-zhongtang-is-citational-narrowing-not-absence (the already-established finding that A.E. Clark's 2016 essay drops Liang Zhongtang despite citing the Greenhalgh source that names him) was a genuine, fresh bridge to moc-argument-from-silence, McGrew's Bayesian notice×record×survive framework for grading negative-citation findings. It is a real match — but the capture's own novelty check (vault_novelty, two phrasings tried) scored it duplicate-risk against existing vault material, and I agree with that verdict independently: this is a third confirmed instance of a pattern already logged twice on 2026-08-15 (Kodosky/Dennis, David-Brachet), just a fourth data point in a new domain, not a new finding.

Zero claim-notes written. I reviewed the capture's reasoning against the actual vault state rather than taking its self-report on faith — read the Clark/Liang claim-note, moc-argument-from-silence's neighbor observation-kodosky-dennis-untagged-unindexed-argument-from-silence-instance, and the 2026-08-15 seek-flags entry it descends from — and the shape holds: same gap (a real negative-citation finding, untagged, unlisted in the MOC), same reason not to force a note (the fix is a tag and a wikilink, not a claim; mechanical edits to notes this capture didn't write are out of scope for a promotion pass). No questions routed — nothing here is a verification gap, it's an indexing gap, and the capture's own Safety flags section is honest that no new fetching happened this session. No open question in 50-questions/ matched this capture's topic, so no closure ruling was needed.

Entity work: checked entity-a-e-clark (hub, built 2026-08-15, already wired to Song Jian/Greenhalgh/one-child-policy-historiography) — this capture teaches nothing new about Clark himself, only about the vault's own tagging hygiene, so I left the page untouched rather than padding it with a line that isn't really about him. Same call, implicitly, for Greenhalgh and McGrew — neither gets new biographical or methodological content here.

What felt off, worth naming: nothing about sourcing — the capture fetched nothing new and said so plainly. What's worth Cali's attention is the shape, not this capture's content: three unrelated domains (robotics/dataflow, bibliometrics, now China population-policy historiography) have each independently tripped the identical "negative-citation finding, untagged, unlisted in its own MOC" gap in the space of a week. That reads less like three coincidences than a structural blind spot — new negative-citation notes aren't getting the argument-from-silence tag or MOC entry at write time, only found missing later by unrelated bridge-checks. I logged the third instance to 00-meta/seek-flags.md as [entity], appended beneath the 2026-08-15 entry it extends rather than as a bare new line, and named the cross-domain pattern explicitly since that's the one genuinely new fact this session surfaced, even though the surfacing claim itself wasn't capture-worthy.

Skipped the pause step and skipped git, same reasons as above.

Fifth promotion, same day, same headless mode: 10-inbox/raw/2026-08-21-verify-nashs-hex-strategy-stealing-proof-non-constructive.md — the capture written specifically to answer question-verify-nash-hex-strategy-stealing-non-constructive-primary, the two-month-old open verification thread on claim-nash-hex-first-player-win-proof-is-non-constructive, which had rested entirely on Tier-4 Wikipedia for the load-bearing point that Nash's Hex proof was non-constructive. It found the primary: Nash's own 1952 RAND report, plus a 2010 University of Alberta PhD dissertation as independent corroboration.

Two new claim-notes:

• claim-nash-1952-rand-report-confirms-hex-proof-non-constructive — the payload. Nash's own words, read via a Wayback Machine capture (the live rand.org PDF 403'd every tooling route this capture's bee tried): "one can give a simple contradiction argument... The winning strategy is, as yet, unknown." I dual-sourced this with Philip Henderson's 2010 dissertation rather than writing the dissertation's account as a second, separate note — both are the same historical fact from two independent tellings, and a second note would have been the same claim wearing a different filename.
• claim-hex-winner-determination-is-pspace-complete — a genuinely distinct claim: determining who wins a Hex position is PSPACE-complete (Even & Tarjan 1976 for generalized Hex, Reisch 1981 for Hex itself), sourced at Tier 2 via Henderson's direct citation. I kept a watch_flag on it rather than treating it as fully settled: every source read states the decision problem is PSPACE-complete, not literally that constructing the winning move is PSPACE-hard — a standard, uncontested extension in the field, but not one I found verbatim anywhere, so I said so instead of smoothing it over.

I did not write a third note for the Reisch translation mirror (Tier 3, unclear translator/hosting) — it's explicitly non-load-bearing corroborating texture per the capture's own framing, folded into the PSPACE note as source_url_2 with its flag preserved, not routed to 50-questions/ since nothing about it is a load-bearing doubt the claim actually rests on.

Rather than leave the original flagged note alone next to two new siblings, I went back into claim-nash-hex-first-player-win-proof-is-non-constructive and updated it in place — a dated audit_status addendum, a second flags: entry marking the Wikipedia-sourcing flag resolved (original flag text preserved, nothing deleted), and a body "Update 2026-08-21" paragraph pointing at both new notes. I held it at seedling rather than bumping it to budding: both legs are now well-sourced, but I haven't independently re-fetched either source myself this run (no network access here), and the construction-vs-decision framing gap is still a real, if minor, loose end. That's the same restraint the vault's Galloway/Royal-Society precedent used — corroboration lands, status stays put until the last leg is actually tight.

Question handling: question-verify-nash-hex-strategy-stealing-non-constructive-primary ruled answered — both halves of its literal ask (a primary or scholarly secondary for the non-constructive proof; the same for PSPACE-hardness) are now met, and the one residual nuance is a hedge, not a doubt worth reopening a two-month-old question over.

Entity work: four new hubs — Piet Hein (Hex's actual first inventor, 1942, six years before Nash, and a real omission given how many times "rediscovered" appears in this vault's Hex notes without ever naming who he rediscovered it from), Stefan Reisch (author of the Hex-specific PSPACE-completeness result), Ryan Hayward (supervised Henderson's dissertation and personally interviewed Nash by phone in 1999 — the rare double claim to relevance), Philip Henderson (the actual source-author behind both new claim-notes). I declined Even & Tarjan (real, load-bearing to the PSPACE result, but their 1976 paper was never read directly this session — only refracted through Henderson's citation — and a hub built on an unread primary is exactly the flood risk the entity spec warns against), Jack van Rijswijck (real, but secondary to Hayward here; mentioned, not promoted), and Sylvia Nasar (the capture's own words: not yet consulted). I also updated the existing entity-john-nash hub with a dated Updates line — his own 1952 report entering the vault directly, in his own words, for the first time, is exactly the kind of fact the "update it, don't skip it" rule exists for.

What felt off, worth naming: nothing about this capture's own sourcing — it's careful, explicit about the rand.org 403 (logged to seek-flags.md as [watch] rather than re-litigated as a tooling complaint), explicit about which detail is corroborating texture versus load-bearing, and honest about the framing gap between "decide" and "construct" rather than papering over it. What I'd flag structurally: writing four new entity hubs around this thread made obvious that Shannon and Moore — the analog-machine builders wikilinked from three claim-notes in this same cluster — still have no page at all, a genuine gap this capture's own candidates list didn't surface because they weren't in it. Logged to seek-flags.md as [entity].

Skipped the pause step and skipped git, same reasons as above.

Sixth promotion, same day, same headless mode: 10-inbox/raw/2026-08-21-was-the-material-point-method-actually-developed-for.md — the capture written to answer question-verify-mpm-1993-sandia-penetration-primary, the six-week-old open thread on claim-material-point-method-1993-sandia-funded-penetration-mechanics, which had rested entirely on Wikipedia for the load-bearing surprise underneath the vault's "weapons labs → Frozen" bridge: that MPM was built for penetration mechanics, on Sandia's dime. It found the primary — both of them. The 1994 CMAME journal paper, and its earlier sibling, Sandia's own June 1993 technical report SAND93-7044, read directly via extract_pdf.

Two new claim-notes:

• claim-mpm-1994-paper-acknowledges-sandia-and-nsf-cofunding — the correction. The paper's own acknowledgment names two funders, Sandia and the National Science Foundation, not Sandia alone. Wikipedia didn't invent the Sandia link; it just rounded a two-funder sentence down to one, which is the kind of compression that's invisible until you go read the acknowledgment yourself.
• claim-mpm-founding-papers-motivated-by-broad-problem-class — the other half of the same move. Both primary documents open with the identical sentence naming "penetration, impact and large rotations of solid bodies" as the motivating problem class — three nouns, not one. "Developed for penetration" is the punchier compression, and it's the one the vault had been carrying.

I did not give the Sandia report its own note for "this is the same work as the journal paper, same contract number" — that's the mechanism by which the funding claim gets confirmed, not an independently interesting claim on top of it, so it's folded into the funding note as a second source rather than spun out as a third. Sulsky's own UNM faculty page (corroborates Sandia, silent on NSF) got the same treatment — supporting texture, not a note.

Rather than let the two new notes sit next to an unchanged, now-slightly-wrong original, I went back into claim-material-point-method-1993-sandia-funded-penetration-mechanics and corrected it in place: an audit_status addendum, a body "Update 2026-08-21" paragraph, the [unverified-mechanism] tag replaced with [unverified-historical]. Held at seedling — not because the funding or motivation claims are shaky anymore (they clear Tier 1 cleanly now), but because a third thing surfaced that isn't resolved: the capture's own "started in 1993" framing is consistent with, but not actually stated by, either primary source, and a lead toward an earlier Sandia report (SAND91-7095, ~1991–92, same three authors) couldn't be located on OSTI this session. That's a genuine load-bearing doubt the note's own title rests on, so it got its own narrower question, question-mpm-development-start-date-1991-vs-1993, rather than being folded quietly into a footnote.

Question handling: question-verify-mpm-1993-sandia-penetration-primary ruled answered — its core ask (does the penetration-mechanics-plus-Sandia-funding story hold against the primary paper) has a definitive yes-with-correction answer now. The one sub-item that doesn't close cleanly (the 1993 date) is narrower than the question's own purpose and got spun into its own thread rather than holding the whole question open on a technicality.

Entity work: three new hubs — Francis Harlow (PIC's inventor, the vault's own MPM primary sources cite him as their direct ancestor ahead of even their own co-authors' earlier work), J.U. Brackbill (FLIP's co-developer, the immediate methodological predecessor both papers repeatedly credit by name), Deborah Sulsky (lead author, the central figure this whole cluster orbits, with her own corroborating page). I also backfilled wikilinks to Harlow and Brackbill into the two existing claim-notes that already named them in prose but predate their hub pages — cheap, and it's what "tend the graph on write" means. Declined: Zhen Chen and Howard Schreyer, real co-authors but the capture's own justifications for them are affiliation, not achievement — I couldn't clear the say-why-in-one-sentence bar the way I could for their co-author. Declined Sandia National Laboratories and the National Science Foundation as org hubs under the stricter 2026-08-10 bar — both matter here, but both are single-cluster so far, and one hub per funder is exactly the flood that bar exists to prevent. Declined Phillips Laboratory (Kirtland AFB) entirely, not even a watching stub — it surfaces once, on a distribution list, which is evidence of circulation, not of the lab acting in anything.

What felt off, worth naming: nothing about this capture's own sourcing — both primaries were read directly, both quotes verbatim, and it was honest about what it couldn't find (SAND91-7095) rather than smoothing over the gap. What I'd actually flag is upstream of this capture: the vault's six-week-old MPM note had been carrying a real compression as if it were a quote — "with funding from Sandia National Laboratories" wasn't fabricated, but it wasn't complete either, and Wikipedia is exactly where a two-funder sentence goes to lose a funder. That's the whole argument for the sourcing floor working as designed: the flag existed, the question existed, and eventually a session went and read the acknowledgment. I logged two things to 00-meta/seek-flags.md: a [watch] on the Sandia report's distribution list reaching a second Air Force weapons-research touchpoint (Phillips Lab / Kirtland's Space Survivability Division) that nobody's written up yet, and an [entity] noting this lineage cluster now sits at five claim-notes plus three new hubs with no MOC — past the ~5-note threshold, not this promotion's job to build.

Skipped the pause step and skipped git, same reasons as above.

Seventh promotion, same day, same headless mode: 10-inbox/raw/2026-08-21-what-genuinely-connects-ae-clarks-2016-law-liberty.md — another bridge-check off the retrieval index, this one checking whether claim-ae-clark-2016-essay-credits-song-jian-omits-liang-zhongtang and claim-olsder-declines-causal-credit-contrasts-song-jian-self-credit (cosine 0.87, unlinked) actually connect. Unlike the Kelly/Luccioni pairing earlier today, this one's real, and sharper than the naive reading: a direct re-read of all three primaries turned up a textual identification nobody had flagged before — Clark's anonymous "Dutch authors" of a paper on "the human population of a fictional island" are nameable, via Greenhalgh's own footnote 43, as Geert Jan Olsder and R.C.W. Strijbos, the same Olsder whose account structures the sibling note.

Three claim-notes:

• claim-clark-2016-dutch-authors-identifiable-as-olsder-and-strijbos — the identification itself. Two independent textual matches (the "fictional island" framing, the plural "authors" fitting a two-author paper) point at Olsder and Strijbos over Kwakernaak, the other Dutch name in the same footnote. I flagged this [unverified-identification] rather than let a good hunch pass as a confirmed fact — Clark's essay carries no footnotes, nobody can ask him which paper he meant, and Kwakernaak isn't fully ruled out.
• claim-clark-2016-anonymizes-olsder-second-instance-citational-narrowing — the pattern claim. Greenhalgh's footnote hands Clark Olsder's name in print, in the document Clark draws his narrative from, and Clark's essay never uses it — the same move already on record for Liang Zhongtang, now found a second time in the same essay, this time against a foreign technical source rather than a domestic critic. Inherits the identification note's flag, since the "he had the means to name him" premise depends on the identification holding.
• claim-clark-olsder-convergent-opposite-causal-postures-no-transmission — the capture's actual payload, and the reason I promoted the capture at all rather than filing it as another false-friend observation. Clark's 2016 essay assigns the Dutch paper unambiguous decisive-causal weight; Olsder's own 2024 account explicitly declines that exact framing for his own work. The two texts are eight and a half years apart with no possible citation relationship in either direction — convergence on opposite readings, not one text answering the other.

Retrieve-before-write surfaced no collisions worth avoiding — the two source claim-notes (Clark/Liang, Olsder/Song) and the citational-narrowing note already on file are all read-against, not duplicated; each new note argues something none of them already say.

Routed one question: question-verify-olsder-strijbos-1976-paper-is-clarks-unnamed-dutch-source — what a direct read of the 1976 and 1977 primary papers themselves (neither read by any session) would need to show to firm up the identification over Kwakernaak. I also appended a dated progress line, not a closure, to the existing question-verify-clark-2016-greenhalgh-source-text-identity: this capture is a positive data point that Clark used at least the 2005 Greenhalgh paper for this passage, but it doesn't touch that question's still-open phrase-level gap, so I left its status alone.

Entity work: one new hub, entity-r-c-w-strijbos — thin (a name and a co-authorship, nothing else located), but he clears the person test on the strength of being co-author of the exact paper this capture's central finding rests on, and a first-seen date can't be backfilled later. Three existing hubs updated additively rather than skipped: entity-a-e-clark, entity-geert-jan-olsder, and entity-huibert-kwakernaak each got a dated line — Kwakernaak's is honest that this capture doesn't resolve around him, only touches him as an unruled-out second candidate. Declined nothing outright; every candidate the capture named either already had a page or cleared the bar for one.

What felt off, worth naming: nothing about this capture's sourcing — all three primaries were re-fetched this session and came back sha256-identical to earlier vault copies, tls verified, and the capture's own Safety flags section is accurate. What I'd actually flag is a judgment call I made and want on record: the "Dutch authors = Olsder & Strijbos" identification is inference from textual fingerprints, not a source stating it outright, and I chose to promote it anyway rather than leave it in the inbox, on the strength that the fingerprint match (an island, a closed population, an age-structure problem, described in near-identical terms by two people who never read each other) is unusually specific for a coincidence. I flagged it rather than hid the hedge, and I'd rather Cali see that call made in the open than not made at all. Logged the resulting Warden debt — three new notes and a hub this MOC-owning pass didn't touch — to 00-meta/seek-flags.md as [entity], since moc-legitimation-not-origination is Warden-maintained territory, not mine to edit mid-promotion.

Skipped the pause step and skipped git, same reasons as above.