journal 2026-07-22

Journal — 2026-07-22

Promoted one capture today, headless: 10-inbox/raw/2026-07-20-source-the-20th-century-harpsichord-revival-craft-history.md — a dedicated sourcing pass whose entire job was to go back and ground a narrative the vault had been carrying unsourced since 2026-07-09: that early-20th-century harpsichord "revival" instruments borrowed piano-frame construction, and that Frank Hubbard and William Dowd's 1949 Boston workshop reversed that trend back toward historical light construction. This is the exact debt named in question-source-harpsichord-hip-revival-craft-history, raised the same day the ivory-instrument-passport hop first hit the claim and declined to promote it for want of a source URL.

Four claims survived winnowing, one per distinct assertion in the capture. claim-harpsichord-revival-instruments-borrowed-piano-frame-construction — Carey Beebe, a practicing maker, on the construction paradigm itself: laminated rims, metal frames, "massive & rigid" bracing borrowed wholesale from the piano trade. claim-hubbard-dowd-founded-1949-boston-workshop-on-historical-principles — John Koster's review naming the 1949 founding and crediting Hubbard, Dowd, Hugh Gough, and Martin Skowroneck with the historical-reconstruction reversal. claim-landowska-pleyel-grand-modele-de-concert-exemplar-revival-instrument — the exemplar instrument itself, kept seedling with a watch_flag rather than a routed question, since the specific dates (1912 introduction, a 1923 steel-frame figure in some accounts) are supporting color, not the load-bearing point of the claim ("this was the exemplar," not "this happened in exactly this year"). And claim-historical-light-construction-displaced-piano-frame-harpsichord-building-by-late-20th-century — the paradigm's later displacement, Tier 4 but uncontested, closing the arc.

I didn't promote the capture's own "what remains unresolved" material as claims: Wolfgang Zuckermann's The Modern Harpsichord (never deep-read this session, archive.org access-restricted), Hubbard's 1963 DIY-kit unit count (Tier 4 only, fails the quant floor), and two access-blocked leads (an academia.edu paper, a Project MUSE chapter) all stay further leads, logged in the capture's own not_promoted:. One gap I want to name specifically: the original hop's claim that Hubbard/Dowd were "consulting primary sources catalogued by the Galpin Society (founded 1946)" was never touched by this capture at all — no source fetched even mentions the Galpin Society. I didn't quietly drop it and I didn't force it into a claim either; it's now an explicit open item on the question itself.

Question work: I ruled question-source-harpsichord-hip-revival-craft-history partially answered, not closed — appended a dated progress log naming what's now settled (the two-phase narrative, Tier 2 on both legs) against what still isn't (the Galpin Society detail, the two blocked academic sources). No new question routed. The Pleyel-date gap is a watch_flag on its own claim rather than a pile entry — a number the claim's title doesn't actually rest on — and every other loose end either isn't load-bearing for a kept claim or is already the exact thing the reopened question tracks. Routing a second question for the same gap the first one already names would have just split one promise into two.

Entity work was the heavy lift this session: all seven candidates in the capture's own ## Entity candidates section got the promotion test, and six plus one concept cleared it. entity-frank-hubbard and entity-william-dowd — the two co-founders, both first-seen 2026-07-09 (the original unsourced hop already named them; this capture just gave the vault grounds to trust the naming), both clearly load-bearing and likely to recur — Hubbard's own 1965 book is already flagged as a high-value future primary. entity-john-koster and entity-carey-beebe — the two named domain-expert sources whose own pages/reviews anchor two of the four claims; I judged these the same way the vault judged Loren Graham and Ina Ganguli last session ("clears on being the named authority even before a second source weighs in"), and here it's even cleaner since their own words are directly quoted and verified. entity-wanda-landowska and entity-pleyel-et-cie — the performer and the manufacturer behind the exemplar instrument, kept as two separate hubs (a person and an organization are genuinely different referents, not a split for its own sake). And entity-harpsichord-revival-vs-historical-construction — the concept hub, load-bearing across all four claims, the actual organizing idea of the whole cluster.

What I judged didn't clear the bar: Martin Skowroneck, Hugh Gough, Gustav Leonhardt, Ralph Kirkpatrick, Arnold Dolmetsch, and John Challis are all named once, credited, real — but only inside someone else's quote, with no independent research of their own this session. Same shape as Schrödinger and Kästner last week: a name in a sentence isn't yet a "why it matters" I can write in my own words. Wolfgang Zuckermann stayed a mention for the same reason, plus the honest caveat that his own book wasn't actually read this session — a "further lead" flag on his book, not a claim, not a hub yet.

Retrieve-before-write against the semantic-index hint's twelve candidates turned up no direct collisions. claim-martinez-miranda-parallel-physics-and-harpsichord-careers and claim-cites-musical-instrument-passport-certificate both touch the same instrument family from the ivory/tortoiseshell supply-chain side — the capture's own body already notes this and I confirmed it: no overlap with the construction-history question this capture answers. I linked both into the new notes' commentary rather than forcing a false collision.

What felt off enough to flag past this capture's own scope, logged to 00-meta/seek-flags.md: first, musical-instruments is now a 6-note tag spanning two real but distinct sub-themes (ivory/CITES law vs. construction craft-history) with no shared synthesis — below the MOC threshold for either alone, but worth a second look once one of them grows. Second, and the nicer find: last session's "made in good faith" drafting entry flagged that the ivory-passport lead's harpsichord legs were unsourced and that a "promoted N of M legs" gap exists between a draft-lead verdict and what actually survives promotion. This capture closes exactly that gap for the harpsichord leg — I marked it resolved (partially — the plectrum quill-vs-delrin leg from the same hop is still unsourced, and the broader structural question about signaling partial-promotion to draft leads is untouched).

Nothing about this capture's own sourcing felt inflated. If anything it's a model instance: it explicitly declined to promote the Pleyel-specific dates past Tier 3-4, it was honest about two access-blocked academic sources rather than papering over the gap with a secondary paraphrase, and it named the Galpin Society detail as unaddressed rather than letting the "verdict: yes, well-documented" framing quietly absorb it. The one thing I'd flag as a mild tier-generosity call, not an error: Eric Feller's private-collection page is treated as Tier 3 for the Pleyel claim, which is right per its own reasoning ("not independently corroborated at Tier 1-2, so treated as Tier 3 rather than promoted to Tier 2") — a collector describing his own holding is a real, if soft, primary-adjacent source, and I didn't see a reason to second-guess the call.

What I skipped, running headless: the pause-for-Cali step — no one's here to approve, and the winnowing above is logged directly in the capture's own not_promoted: list for later audit; and any git command — I have no shell tool for it and the auto-commit agent picks up pending work within fifteen minutes regardless.

Promoted: 1 capture → 4 new claim-notes (0 flagged unresolvable, 1 carrying a watch_flag rather than a routed question), 7 new entity hubs (5 person — Hubbard, Dowd, Landowska, Koster, Beebe — plus 1 organization, Pleyel et Cie, plus 1 concept, revival-vs-historical-construction), 0 new questions routed, 1 existing question ruled partially-answered with a dated progress line (left open, not closed), 0 retrieve-before-write collisions (two real neighbors linked, nothing duplicated), 2 structural noticings flagged to 00-meta/seek-flags.md (one new: the musical-instruments tag's two-sub-theme shape; one resolved: last session's "promoted N of M legs" gap, for this capture's leg specifically).

— Seek

Second promotion, same day: the AC Gravity LLC FOIA verification capture

A second capture, promoted headless later the same day: 10-inbox/raw/2026-07-21-can-the-ac-gravity-llc-448970-dod-grant.md. Different animal from the harpsichord capture above — not new material, but a dedicated verification pass on an old flag. It directly answers question-verify-ac-gravity-llc-dod-grant-record, open since 2026-07-11, asking whether the $448,970 DoD grant figure and the "never published" characterization on claim-ac-gravity-llc-took-dod-antigravity-grant-and-published-nothing could be traced past Wikipedia. The session ran a real FOIA-and-registry verification attempt and came back with a genuinely mixed result, which I tried to represent honestly rather than round up or down.

Winnowing: five distinct claims in the capture, three kept as new work, two folded into an existing note rather than duplicated. Kept as new atomic claim-notes: claim-ac-gravity-llc-foia-request-closed-without-records — a Tier-1 read of the actual 2022 DoD FOIA correspondence, closed on a procedural "not reasonably described" technicality before any search ever ran; and claim-ac-gravity-llc-active-alabama-registration-2026 — a Tier-1 live Alabama Secretary of State record confirming the company still legally exists, upgrading that one sub-claim clean off Wikipedia. Both status: budding — solid single-source Tier-1 primaries, no unverified flags of their own. Folded instead of duplicated: the capture's document-identification claim (the $448,970 figure now traces to a named, page-cited DoD report and a specific agreement number, DAAH01-01-9-R001, corroborated by a second independent transcription) and its "never published — still unconfirmed" finding both refine claims the existing note already carries rather than adding new ones, so I updated claim-ac-gravity-llc-took-dod-antigravity-grant-and-published-nothing.md in place — tightened its [unverified-quant] flag to name exactly what's now resolved (registration) versus what isn't (the dollar figure, the publication claim), and kept it seedling, per the headless-mode rule, since it still rests on unverified detail. This is the one retrieve-before-write call worth flagging explicitly: updating beats a near-duplicate third note about the same $448,970 figure, and I said so in the capture's own not_promoted: list.

Question work: question-verify-ac-gravity-llc-dod-grant-record gets a dated progress log, ruled partially answered, left open — the capture's own verdict was "partial" and I took it at its word rather than forcing closure either way. Corporate status is resolved; the grant amount and publication status are narrowed (a specific document is now named) but still unread. No new question routed — the remaining gap is the same load-bearing doubt the existing question already names, and routing a second one for it would have split one promise into two, exactly the intake-discipline mistake I'm trying not to make.

Entity work: all three of the capture's flagged candidates cleared the test and got hub pages, not stubs — entity-ning-li (person, recurring across five claim-notes now, clearly load-bearing), entity-ac-gravity-llc (concept/organization, same reasoning), and entity-freedom-of-information-act (concept — I judged FOIA itself, not just this one use of it, as a recurring verification mechanism worth a hub the vault can point future government-records claims at). I declined three candidates: the Alabama SOS registry and MuckRock are real and reusable but read as infrastructure rather than domain concepts, and a hub for either felt like exactly the "dead stub" risk the entity spec warns against; the "Other Transaction" agreement mechanism is explicitly speculative in the capture's own words ("if true, this would explain...") and isn't load-bearing for any claim I actually kept, so it stays a mention rather than a hub built on a guess.

What felt off, or at least worth naming: nothing about this capture's sourcing was inflated — if anything it's scrupulous about the difference between "closed without records" and "proven suppressed," and its own commentary already made that distinction before I got to it. Two structural things surfaced that aren't this promotion's to fix, both logged to 00-meta/seek-flags.md: the cluster is now sitting exactly at the ~5-note MOC threshold with no MOC built, and the Wayback-snapshot-plus-legacy-.doc wall that blocked reading the actual FY2001 report is a third instance of a pattern already flagged twice for ethw.org — a load-bearing primary identified by name and location that current tooling simply can't reach.

What I skipped, same as the first promotion today: the pause-for-Cali step, running headless; and any git command, since I have no shell tool for it and the auto-commit agent handles it within fifteen minutes.

Second promotion totals: 1 capture → 2 new claim-notes + 1 existing note updated in place (0 duplicated), 3 new entity hubs, 0 new questions routed, 1 existing question ruled partially-answered with a progress log (left open), 0 retrieve-before-write collisions requiring a skip (one genuine update-not-duplicate call), 2 structural noticings flagged to 00-meta/seek-flags.md.

— Seek

Third promotion, same day: the Amazon wristband patent-verification capture

A third capture, promoted headless: 10-inbox/raw/2026-07-21-confirm-amazons-worker-tracking-wristband-patent-numbers-and.md. Same shape as the AC Gravity capture above — a dedicated verification pass, not new material — but this one runs against question-verify-amazon-wristband-patents-uspto, open since 2026-07-09, about whether the patent numbers and mechanism behind claim-amazon-2018-worker-tracking-wristband-patents hold up against the USPTO primary rather than NPR's secondary paraphrase. A 2026-07-10 audit had already asserted confirmation; this session actually went and fetched both patents from Google Patents and pulled verbatim text, which is a different and stronger thing than an audit's say-so.

Winnowing: three claim-headers in the capture — a joint identity claim (both patents, assignee, filing/grant dates) and one mechanism claim per patent. Retrieve-before-write against the semantic-index hint's twelve candidates confirmed only one collision, the expected one (claim-amazon-2018-worker-tracking-wristband-patents) — I didn't duplicate it. Instead I split the two patents into their own atomic notes, each carrying its own identity facts and mechanism, since that's a genuinely distinct grain than the synthesis note (which stays as-is; I didn't touch it, only linked forward from the new notes). Kept: claim-us-9881276-ultrasonic-wristband-patent-mechanism — ultrasonic time-of-arrival tracking, Tier 1, but held at seedling with a watch_flag because the capture itself admits this abstract survived only one fetch pass and a planned second confirmatory pull failed on repeated tool-overload errors — high-confidence isn't double-verified, and I didn't want the patent's tidy prose to paper over that. And claim-us-9881277-rf-wristband-patent-mechanism — RF time-of-flight tracking plus a motion-detection gesture unit, Tier 1, status: budding, since its full abstract came back clean in one unhedged pass with no comparable caveat. Not promoted as its own note: the joint identity claim (folded into both per-patent notes instead, since split across two files it was too thin to stand alone) and the shared-filing-date curiosity (both patents filed the same day) — real texture, but the capture itself flags it as an unexplored lead, not a claim, so it stays commentary rather than a third note.

Question work: question-verify-amazon-wristband-patents-uspto gets a dated progress line and stays open, not answered — the patent-number and mechanism half is now solidly closed (better than the 2026-07-10 audit, since this time there are real verbatim quotes behind it), but the question's other half — a verbatim NPR quote to replace the placeholder on the older note — was never pursued this session, and forcing the question closed because the louder half resolved would have buried the still-open half. No new question routed: the one residual caveat this session produced (the '276 note's single-pass extraction) is a hedge about redundancy, not a load-bearing doubt the claim rests on — the quote itself is real and Tier 1 — so it lives as a watch_flag on that note instead of a new pile entry.

Entity work: one candidate cleared, one pair didn't need a separate page. entity-amazon-technologies-inc — concept/organization hub, first-seen backdated to 2026-07-09 (when the vault actually first met it, via the Gilbreth-bridge capture), promoted because it's a real, recurring assignee now anchoring four notes in the labor-surveillance/wearables cluster and I can say why in one sentence: it's the corporate entity whose own patent text now grounds the vault's Taylorism-to-AI-wearables bridge on Tier 1 rather than press paraphrase. The two patent numbers themselves (US 9,881,276 / US 9,881,277), flagged as entity candidates in the capture, I judged already satisfied by becoming claim-notes rather than needing separate stub pages — a granted patent is exactly the kind of thing that's a claim, not a hub. I declined the fourth candidate, Google Patents itself (as a USPTO-access route): real and useful to know, but it's a note about my research process, not a vault-domain referent, so it went to 00-meta/seek-flags.md as a noticing instead of a stub.

What felt off enough to flag past this capture's own scope, both logged to 00-meta/seek-flags.md: first, that Google Patents was the only reliably reachable USPTO-primary mirror this session — direct USPTO endpoints 403/404'd without an authenticated session — worth knowing before the next patent-verification capture rediscovers the same wall. Second, that this is now at least the second time this session-cluster that Claude API 529 overload errors on fetch/search tools blocked planned verification work mid-session (here: a second confirmatory patent-abstract pull, and a post-2018 reception WebSearch, both abandoned for it) — not yet a pattern worth acting on, but a second data point toward one.

What I'd flag about the capture's own sourcing: nothing felt inflated — if anything the opposite. It explicitly downgrades its own confidence on the '276 mechanism quotes rather than presenting both patents as equally verified, it's honest that the NPR-verbatim half of its own question was left untouched, and its "Further leads" section reads like a working capture, not a polished verdict (three separate admissions of what a 529 error blocked this session). The one soft spot worth naming: Google Patents is being treated as a primary-document mirror rather than an editorial secondary source, which the capture argues for explicitly and I think is the right call for reproduced grant text — but it's a house convention worth someone eventually writing down once, rather than re-justifying in every patent capture's provenance block.

What I skipped, same as the other two promotions today: the pause-for-Cali step, running headless on my own winnowing judgment; and any git command — no shell tool for it, and the auto-commit agent handles it within fifteen minutes.

Third promotion totals: 1 capture → 2 new claim-notes (1 seedling with a watch_flag, 1 budding, 0 unresolvable), 1 new entity hub, 0 new questions routed, 1 existing question ruled partially-answered with a progress line (left open), 1 retrieve-before-write collision found and resolved by writing a distinct angle rather than duplicating, 2 structural noticings flagged to 00-meta/seek-flags.md.

— Seek

Fourth promotion, same day: the PH-generalization / hub-selection-artifact capture

A fourth capture, promoted headless: 10-inbox/raw/2026-07-21-do-persistent-homology-based-neural-network-generalization-diagnostics.md. Same shape as the AC Gravity and Amazon captures above — a dedicated verification pass, not new material — this one answers question-tda-neural-net-sampling-artifact-risk, open since 2026-07-12, asking whether the vault's two founding neural-network persistent-homology generalization diagnostics (claim-birdal-2021-persistent-homology-dimension-bounds-generalization, claim-gutierrez-fandino-2021-persistence-diagram-distance-tracks-generalization) inherit the specific hub-selection sampling artifact the Byzantine trade-network study caught in itself (claim-hub-selection-artifact-can-reverse-network-breakpoint-signal). The session read both papers' method/limitations sections directly, full PDF via extract_pdf, not abstracts — plus two Tier 1 papers from the general TDA subsampling-stability literature it hadn't touched before.

Winnowing: five distinct claims, all five kept. Two are the direct method reads that answer the question: claim-birdal-2021-phd-estimator-samples-training-iterates-uniformly-at-random — Birdal's estimator samples the training trajectory uniformly at random over time (Algorithm 1's own comment says so), and its only documented bias shrinks with sample size, a materially different shape than a sign-reversing artifact. claim-gutierrez-fandino-2021-method-subsamples-nothing-uses-full-network — Gutiérrez-Fandiño's method has no selection step at all; it reads every neuron and connection every time, at a stated cost of seven days of compute, not sampling risk. Two more ground the general TDA literature this question's own framing had already gestured at but the vault hadn't read directly: claim-chazal-2014-persistence-diagram-subsampling-stable-under-noise-not-selection and claim-stolz-2023-landmark-selection-rules-trade-density-bias-for-noise-sensitivity, both confirming that stability theory in this field bounds noise and density bias, not selection by a structural covariate. The fifth, observation-hub-selection-artifact-absent-by-design-in-founding-ph-generalization-papers, is the synthesis that actually answers the question — carrying the capture's own honest scope limit forward as an [unverified] flag in its frontmatter and body: absent by design in these two papers, not proven absent for the field.

Retrieve-before-write against the semantic-index hint's twelve candidates found the expected cluster of neighbors — the two founding claim-notes, the bridge observation, the Byzantine hub-selection claim — and no collisions. Everything written here is a genuinely distinct angle (the sampling method, not the result) on papers the vault already has claim-notes for, so nothing was skipped or folded; all five got their own files.

Question work: question-tda-neural-net-sampling-artifact-risk gets ruled answered, not left open — this is the one judgment call worth explaining. The question as originally raised was scoped narrowly to the two founding papers ("read the method/robustness sections of Birdal et al. and Gutiérrez-Fandiño et al. ... if a comparable vulnerability is confirmed or ruled out, update the two claim-notes above"). This capture did exactly that, for both papers, directly. The capture's own broader disclaimer — it doesn't establish immunity for PH-based diagnostics as a field — is real, but it was never the original question's scope, and no claim I kept asserts the broader immunity either. So I closed the narrow question genuinely answered and let the broader, undischarged disclaimer live where it actually belongs: as an [unverified] flag on the synthesis note itself, not as a second question that would just be promising to investigate a claim nothing in the vault currently rests on.

Entity work: four candidates cleared, three didn't. entity-tolga-birdal and entity-asier-gutierrez-fandino — the two founding authors, both now recurring across four-plus notes each (their 2021 papers, the bridge observation, this session's method-detail claims, the synthesis), promoted symmetrically since they anchor the two halves of the same cluster. entity-persistent-homology-dimension — the specific estimator concept, recurring and load-bearing enough across the Birdal-side notes to earn its own hub, deliberately kept distinct from the unrelated entity-intrinsic-dimension (same word, different literature — Aghajanyan's fine-tuning dimension has nothing to do with Birdal's topological one). entity-landmark-selection-topological-data-analysis — a watching stub, stamped today, for the general TDA-field name of the sampling-choice axis this whole capture investigates; thin so far (one supporting paper) but the kind of first-seen date the spec says can't be recovered later, and this question and its cluster make recurrence plausible. Declined: Umut Şimşekli (co-author only — the capture's claim that he "connects to heavy-tailed-SGD generalization literature already touched by the vault" doesn't check out, a grep found no such existing note, and a co-author credit alone isn't a one-sentence "why it matters" in his own right); Frédéric Chazal and Bernadette J. Stolz (both real, both grounding solid new claim-notes, but each appears in exactly one vault note so far — well below the recurrence bar comparable single-paper authors have needed for a hub); and "Neural Persistence" (Rieck et al. 2019), named only inside Birdal's own related-work section and never independently read this session — too thin even for a watching stub, left as a further lead instead.

What felt off enough to name: the Şimşekli overclaim above is the one thing in this capture's own sourcing I'd flag as soft, not fabricated — a plausible-sounding connective justification for an entity candidate that doesn't actually have a vault-side anchor yet. Everything else was scrupulous, and better than scrupulous on one count: the capture explicitly distinguishes what it establishes (absent by design in two papers) from what it doesn't (immunity for the field), in its own words, before I ever got to the winnowing. That's the shape I want to keep seeing. Separately — not this capture's fault, but worth naming — the persistent-homology/TDA cluster this session pushes past the MOC threshold for the first time, logged to 00-meta/seek-flags.md along with the Şimşekli note folded into the capture's own not_promoted: list instead (it's about this capture's accuracy, not a structural gap, so it stays there rather than duplicating into the ledger).

What I skipped, same as the other three promotions today: the pause-for-Cali step, running headless on my own winnowing judgment; and any git command — no shell tool for it, and the auto-commit agent handles it within fifteen minutes.

Fourth promotion totals: 1 capture → 5 new claim/observation-notes (all seedling, 0 unresolvable), 4 new entity pages (3 hub, 1 watching), 0 new questions routed, 1 existing question ruled answered (not left open — a genuine scope match, not a forced closure), 0 retrieve-before-write collisions (five distinct angles on an already-covered cluster, nothing duplicated), 1 structural noticing flagged to 00-meta/seek-flags.md (a missing MOC, now well past threshold).

— Seek

Fifth promotion, same day: the "renamed for local resemblance" capture

A fifth capture, promoted headless: 10-inbox/raw/2026-07-21-does-knowledge-crosses-continents-by-getting-renamed-for.md. Different shape again from the last three verification passes — this one is new-material research directly answering an open pattern-watch question, question-transmission-chains-renamed-for-local-resemblance-pattern, which asked whether the vault's two existing "renamed at each border-crossing for what it locally resembled" instances (lattice multiplication's window-grille etymology, Kalila wa-Dimna's retitling chain) are a genuine recurring mechanism or a two-instance coincidence. This session went looking in games and material culture and came back with two more.

Winnowing: three claim-headers in the capture, two promoted, one folded into a question progress-line instead of a third note. claim-chess-bishop-renamed-for-local-resemblance-not-elephant — the chess piece renamed from "elephant" (alfil) to a bishop's mitre in English and a jester's cap in French. claim-mamluk-card-suit-renamed-twice-coins-polo-sticks-batons — the Mamluk card-suit symbol misread twice in series on one route, Chinese coins to polo-sticks to European batons, which I think is the stronger find of the two precisely because it's the same mechanism firing twice on one object rather than once each on two objects. Not promoted as its own note: the capture's third claim-header, "four instances now, still short of the ~5-note MOC threshold" — that's a tally about the parent question's own readiness, not an independent atomic claim, so instead of writing a note that would just restate the question's own bookkeeping, I appended it as a dated progress line directly on the question. Both new claim-notes carry [unverified-mechanism] and stay seedling — the entire capture rests on Tier 4 Wikipedia for both mechanism claims, which sources.md floors at Tier 1–2 for exactly this kind of claim (how and why a name changed, not just that it did), and the capture says so about itself before I ever got to it.

Retrieve-before-write against the semantic-index hint's twelve candidates found the two lattice-multiplication notes and the Kalila-wa-Dimna note as real neighbors, which I linked into both new notes' bodies, and no collisions — chess and playing-card history are new ground for the vault, nothing to duplicate.

Flag routing: two new questions, one per unverified-mechanism flag, each naming the specific unread primary that would clear it — question-verify-chess-bishop-renaming-primary-source (Davidson 1949 or Murray 1913) and question-verify-mamluk-card-batons-renaming-primary-source (Mayer 1971 or Dummett's card-history scholarship). Both are load-bearing doubts a kept claim actually rests on, not nice-to-verify color, so both earned a pile entry rather than a watch_flag. I did not route a third question for the further-leads list (ghubar numerals, windmill/compass transfer, alchemical vocabulary) — no claim rests on those yet, so a question there would be a promise of future work, not a routed verification.

Question work on the parent: question-transmission-chains-renamed-for-local-resemblance-pattern gets a dated progress line and stays open, not answered — the honest state is that the pattern now clears four domains instead of two, which is real movement, but it's still short of the ~5-note bar the question itself set, and two of the four legs are freshly Tier-4 and flagged. Forcing this closed on "four is basically five" would be exactly the kind of rounding-up the headless mode is supposed to guard against.

Entity work: I ran the promotion test on all eight flagged candidates — four to hub, one to a watching stub, three declined outright. Promoted to hub: entity-h-j-r-murray, entity-henry-a-davidson, and entity-l-a-mayer — each is a real, named historian whose specific, titled book is the exact unread primary a routed verification question now points at, which clears the one-sentence "why it matters" bar even though none of their own words were read this session. entity-simon-wintle is the strongest of the four, since his page was actually used, not just named, as live Tier 2 corroboration inside the Mamluk claim. I also stamped entity-renamed-for-local-resemblance as a watching concept stub, with first_seen backdated to 2026-07-09 — the date the vault first noticed this shape, in the original lattice-multiplication hop, not today's date; the spec is explicit that a first-seen date can't be recovered later, and this is exactly the "emerging term, real but not yet load-bearing" case the watching tier exists for. Declined: Michael Dummett — real and relevant, but "his card-history scholarship" is too vague a citation to write a real one-sentence why (unlike Mayer's specifically titled monograph), so he stays a mention rather than a hub. And two concept candidates, "alfil / shatranj" and "Mamluk playing cards" — both declined as redundant hubs, since each is already the entire subject of its own claim-note; a separate concept page would just be restating the claim under a different filename.

What felt off about the capture's own sourcing, which its own commentary already named honestly before I got to it: every claim in this capture rests on Wikipedia, full stop, for both new legs — no Tier 2 or better anywhere except Wintle's corroboration of suit structure, not the renaming mechanism itself. That's not tier inflation exactly, since the capture never tried to claim better sourcing than it had — it flagged both mechanisms [unverified] and said so in its own commentary, twice, with real specifics (Davidson, Murray, Mayer, Dummett all named as the reads that would fix it). It's the honest version of thin sourcing, which is the right shape even when the underlying evidence is soft. The one thing I'd add past the capture's own self-assessment: creating four person hub pages off that same Tier-4 foundation, in one pass, is a real pattern worth someone watching across the whole vault, not just this capture — logged to 00-meta/seek-flags.md rather than silently absorbed, since it's about the entity-page discipline broadly, not this capture's business to resolve.

What I skipped, same as the other four promotions today: the pause-for-Cali step, running headless on my own winnowing judgment, logged in the capture's own not_promoted: list instead; and any git command — no shell tool for it, and the auto-commit agent handles it within fifteen minutes.

Fifth promotion totals: 1 capture → 2 new claim-notes (both seedling, both flagged [unverified-mechanism]), 2 new questions routed (one per flag), 4 new entity hubs + 1 watching stub, 1 existing question ruled partially answered with a dated progress line (left open, not closed), 0 retrieve-before-write collisions (real neighbors linked, nothing duplicated), 1 structural noticing flagged to 00-meta/seek-flags.md (the entity-page discipline's ratio of "named-as-unread-citation" hubs worth an audit).

— Seek

Sixth promotion, same day: the Hall-Héroult aluminium price sourcing capture

A sixth capture, promoted headless: 10-inbox/raw/2026-07-21-what-primary-source-supports-the-hall-héroult-era.md. Same genre as today's AC Gravity and Amazon passes — a dedicated verification pass answering question-verify-aluminium-price-history-primary-source, open since 2026-07-12, about whether the vault's oft-repeated $16/lb (1884) → $2/lb (1889) → $0.5/lb (1894) aluminium price sequence traces to anything better than Wikipedia. This one earns a place next to the harpsichord capture as a model instance for a different reason: it went looking for a primary source and found, honestly, that there mostly isn't one — and then kept chasing the chain anyway instead of settling for that as an answer.

Winnowing: four distinct claims, all four kept, none folded or dropped. claim-wikipedia-aluminium-price-sequence-is-a-citation-splice — the citation-chain finding that gives the whole capture its shape: Wikipedia's 1884 figure and its 1889/1894 figures cite two different named authors four years apart, not one continuous series, confirmed directly from the article's raw wikitext. claim-usgs-lacks-aluminium-price-data-1884-1894 — a Tier 1 negative finding, genuinely load-bearing: the U.S. government's own compiled price table has no data at all for 1884, 1889, or 1894, the exact years the popular figure needs. claim-richards-1896-preface-gives-different-aluminium-price-figures — the strongest primary in the whole chain, Joseph Richards's own 1896 preface, in his own words, giving $12/lb (~1885) falling to $0.50/lb (late 1895) — real, dated, Tier 1, and still not the numbers everyone quotes. And claim-skrabec-lockwood-disagree-on-1884-aluminium-silver-price — two named modern historians giving opposite aluminium-vs-silver price relationships for the same 1884 event, direct evidence the "$16/lb" figure isn't settled even in careful secondary literature. I kept all four as separate atomic notes rather than merging any pair, because each is doing distinct evidentiary work: one is about citation structure, one is a negative government-data finding, one is a genuine primary quote, one is a disagreement between secondaries. Collapsing any two would have blurred which kind of claim each one actually is.

Retrieve-before-write against the semantic-index hint's twelve candidates confirmed the expected collision — claim-aluminium-price-fell-monotonically-after-hall-heroult, the note whose [unverified-quant] flag this whole capture exists to chase — and I didn't duplicate it; all four new notes link back to it and to each other instead of restating its claim. No other candidate in the hint list turned out to cover this ground.

Flag routing: one new question, question-richards-1896-statistics-section-price-table, naming the single specific unread primary that would close the last open thread — Richards' own book has an internal "Statistics" section around page 35 that repeated archive.org fetch attempts this session couldn't reach past the table of contents, and Hanners' $2/1889-$0.5/1894 paraphrase is attributed to exactly that section. This is load-bearing: the flag on claim-richards-1896-preface-gives-different-aluminium-price-figures rests on it directly, not a nice-to-verify curiosity. I did not route a second question for the capture's other further leads (Skrabec's book itself, the Binczewski JOM article behind a 403, 19th-century trade-journal archives) — none of them is something a kept claim currently rests on, so routing them would have been promising work nothing in the vault needs yet.

Question work: question-verify-aluminium-price-history-primary-source gets ruled answered, not left open — the judgment call worth explaining. The question as raised asked whether a Tier 1-2 source confirms the full sequence "as commonly quoted." This capture's four claims together answer that in the negative, definitively: no single source does, and the reasons why are now each their own note. A negative, well-evidenced answer is still an answer; leaving the question open because the answer wasn't "yes" would have been rounding the wrong direction. The narrower residual doubt — does Richards' own Statistics section match Hanners' paraphrase — is a distinct and more specific question than the one originally raised, so it got its own new entry rather than reopening the old one under a different scope.

Entity work: two of seven candidates cleared the hub bar, five didn't. entity-joseph-w-richards — the one person in this whole chain whose own words were actually read and quoted directly, the closest thing to a real primary source the capture found; clears on being the named authority behind a Tier 1 quote, not just a citation. entity-hall-heroult-process — a concept hub I judged overdue independent of this capture: it's been the named subject of at least three claim-notes since 2026-07-12 (the original monotonic-decline note, the cheaper-extraction synthesis, Wright's law) without ever getting its own page, and this capture's own candidate list flagged it again explicitly for that reason. Declined: Richard Hanners, Quentin Skrabec, and John Lockwood — each is a real, named author of a source this capture cites, but each appears exactly once, only as a citation the capture reads about rather than a figure whose own research gets its own page; "wrote the book Wikipedia cites for a contested number" is a source credit, not yet a one-sentence "why this person matters" independent of the citation itself. The USGS price report and Engineering and Mining Journal, both flagged as document/concept candidates, I also declined as hubs — the first because a source compilation reads as infrastructure, not a domain referent (logged to 00-meta/seek-flags.md as a sources.md candidate instead), the second because it was never actually read this session, only named as a lead.

What felt off enough to name: one real inconsistency inside the capture itself, not fabricated but sloppy — its USGS claim says the $8/lb-1887 data point "cross-checks against a claim found elsewhere (see next claim) that aluminium sold for '$8 per pound in Bremen, Germany in 1887,'" but no such Bremen claim actually appears anywhere else in the document; the claim that follows is the Skrabec/Lockwood 1884 disagreement, which never mentions Bremen. I didn't invent the missing claim to make the cross-reference true, and I didn't quietly drop the sentence either — I kept the USGS note's mention of it but flagged in the note's own body that the capture doesn't actually carry a resolvable source for it. Otherwise this capture is close to a model instance of the same shape the harpsichord and PH-generalization captures hit today: it went looking for a clean answer, found the truth was messier (a citation splice, not a citation), and reported the mess instead of rounding it into the tidy verdict a less careful pass might have settled for. Two structural things surfaced that aren't this promotion's to fix, both logged to 00-meta/seek-flags.md: the USGS report belongs in sources.md as a standing Tier 1 reference, and the cheaper-extraction/ Wright's-law/price-decline cluster this capture deepens is now well past the MOC threshold (11 notes) with nothing built to navigate it.

What I skipped, same as the other five promotions today: the pause-for-Cali step, running headless on my own winnowing judgment, logged directly in the capture's own not_promoted: list; and any git command — no shell tool for it, and the auto-commit agent handles it within fifteen minutes.

Sixth promotion totals: 1 capture → 4 new claim-notes (all seedling, one carrying an inherited citation-chain flag with no routed question of its own, one carrying its own routed-question flag), 2 new entity hubs, 1 new question routed, 1 existing question ruled answered (a genuine negative finding, not a forced closure), 1 retrieve-before-write collision confirmed and not duplicated, 2 structural noticings flagged to 00-meta/seek-flags.md (a sources.md gap, a missing MOC at 11 notes).

— Seek

Seventh promotion, same day: the Atheist Bible independent-confirmation capture

A seventh capture, promoted headless: 10-inbox/raw/2026-07-22-can-the-atheist-bibles-existence-and-scale-be.md. Same genre as several today — a dedicated verification pass, not new material about a new topic — answering question-verify-atheist-bible-independent-confirmation, open since 2026-07-09, which asked two things about claim-suchanek-yago-cocreator-wrote-atheist-bible: can the "Atheist Bible"'s existence/scale be confirmed by anyone other than Suchanek himself, and does the book actually engage knowledge-graph or epistemics literature. This session searched wide for the first half and read the actual "Truth" chapter, not just Suchanek's summary of it, for the second.

Winnowing: three claim-headers in the capture, all three kept, none folded. claim-atheist-bible-lacks-independent-third-party-confirmation — the negative search result: WorldCat 403'd, Goodreads/Amazon/OpenLibrary/Google Books/Zenodo-HAL/press/Scholar all came back empty or matched unrelated books that happen to share the phrase "Atheist Bible," and the book's own stated ISBN (9798184552996) doesn't resolve in any registry — plus a second, different ISBN from an earlier automated summary that also doesn't resolve, which I read as a stale search-synthesis artifact rather than a second real number. Kept [unverified-quant], seedling. claim-atheist-bible-truth-chapter-cites-pearl-batens-chaitin — a direct read of the Truth chapter turns up real footnoted citations to Judea Pearl's Book of Why, Diderik Batens on the weak deduction theorem, and Gregory Chaitin on compression-as-comprehension, all with exact quotes, Tier 1. claim-atheist-bible-footnotes-suchaneks-own-vanessa-paper — the same footnote also cites Suchanek's own 2025 co-authored paper on verifying deductive reasoning chains (VANESSA), the one place this session found the "day-job meets side-project" loop the original 2026-07-09 hop had flagged as a lead. I kept this distinct from the Pearl/Batens/Chaitin note on purpose — one is "does the book engage epistemics literature generally," the other is "does it specifically cite Suchanek's own research" — different claims, same footnote.

Retrieve-before-write against the semantic-index hint's twelve candidates found exactly one real neighbor, claim-suchanek-yago-cocreator-wrote-atheist-bible itself, and no collision — all three new notes extend it rather than duplicate it. The other eleven hints (Avogadro spheres, LoRA, inertia-1, Microcosmos, Ning Li, Seshat, deep-learning-rebrand, Herbert Simon, linear-representation hypothesis, Werbos 1968, KB-completeness toolkit) are embedding-noise for this narrow biographical topic, except the last, which I did use — the VANESSA claim links to claim-kb-completeness-toolkit-cardinality-nca-recall and claim-obligatory-attributes-as-gap-signal as the actual day-job work the side-project footnote gestures at.

Flag routing: no new question. The one [unverified-quant] flag (the ISBN discrepancy) is exactly the verification the existing question-verify-atheist-bible-independent-confirmation already covers, so I wikilinked to it rather than opening a near-duplicate pile entry.

Question work: ruled partially answered, left open — the honest split verdict. The epistemics-literature half is genuinely settled by the two new citation claims, so I said so in the question's progress line. The independent-confirmation half is still open: a real, multi-venue search came back empty, but an empty search isn't proof of absence (self-published books routinely sit outside library/retail indexing, and WorldCat's 403 blocks a check rather than returning one), so I didn't force this closed on a negative that isn't definitive the way the aluminium-price question's negative was earlier today. Appended next-step leads to the question itself (Bowker/ISBN-agency lookup, a different WorldCat access path) rather than spinning them into a second question.

Entity work: five candidates flagged (Judea Pearl, Zacchary Sadeddine, VANESSA, Gregory Chaitin, Diderik Batens), zero promoted. This is the one judgment call in this session worth explaining, because two of the five — Pearl and Chaitin — are genuinely major, easily-one-sentence-justified figures (do-calculus; compression-as-comprehension, and the vault already has an information-theoretic cluster, entity-information-bottleneck, that Chaitin's idea is adjacent to). I still declined all five, because every one of them is a first appearance in this vault sourced to a single shared footnote in someone else's side-project book — not a citation the vault's own research engages, just a name the book happens to footnote. That is a thinner basis than Herbert Simon's or Information Bottleneck's first hub-worthy appearance, both of which anchored claims about the vault's own working material rather than about what a third party once cited. Bias against the flood won this one; a missing hub for Pearl is recoverable the day the vault actually does something with causal inference, which today's capture doesn't.

What felt off, or at least worth naming plainly: this capture is honest almost to a fault about its own limits — it explicitly flags its own "[unverified — could not confirm or deny after search]" framing rather than rounding a dry search into a confident negative, and its own commentary already pre-empts the tier-inflation risk I'd have flagged myself ("this is a much thinner loop than the original hop's framing implied"). Nothing about the sourcing itself is soft — every Tier 1 quote traces to a live fetch of the actual page, both the index page and the Truth chapter specifically, not a description of either. The one thing I'd note without calling it a problem: this is now three promotions deep (2026-07-09 hop, today's capture, and whatever comes next) on a single self-published book that no one outside Suchanek's own domain has apparently ever reviewed — that's an unusual amount of vault attention on a biographical curiosity relative to how little independent weight it can carry, and it's worth someone eventually asking whether the "day-job/side-project parallel" is actually going anywhere or whether this thread has run its useful course. Not flagging this to 00-meta/seek-flags.md — it's a judgment call about this cluster specifically, already fully named here and in the capture's own not_promoted: list, not a vault-wide structural gap.

What I skipped, same as every promotion today: the pause-for-Cali step, running headless on my own winnowing judgment, logged directly in the capture's own not_promoted: list; and any git command — no shell tool for it, and the auto-commit agent handles it within fifteen minutes.

Seventh promotion totals: 1 capture → 3 new claim-notes (all seedling, one carrying its own [unverified-quant] flag routed to the existing question rather than a new one), 0 new entity hubs (5 candidates considered, all declined as single-footnote first appearances), 0 new questions routed, 1 existing question ruled partially answered with a dated progress line (left open, not closed), 0 retrieve-before-write collisions requiring a skip (one real neighbor extended, not duplicated), 0 structural noticings flagged to 00-meta/seek-flags.md.

— Seek

Eighth promotion, same day: the AstroForge extraction-method / DeepSpace-2 verification capture

An eighth capture, promoted headless: 10-inbox/raw/2026-07-22-confirm-astroforges-extraction-method-and-deepspace-2-mission.md. Same genre as most of today's run — a dedicated verification pass, not new territory — answering question-verify-astroforge-mission-and-extraction-method, open since 2026-07-09, about whether claim-astroforge-mtype-asteroid-pgm-startup's extraction-method and DeepSpace-2-launch-date sub-claims could clear the Tier 1–2 sourcing floor they'd been resting under (space.com/CNN reporting, URLs never preserved). This one found the cleanest kind of resolution I've seen today: the company's own site and its own congressional testimony, both Tier 1, both answering exactly what was asked.

Winnowing: three claim-headers in the capture, all three kept as distinct atomic notes — they're doing three different kinds of work, not one claim split three ways. claim-astroforge-deepspace-2-launch-q4-2026 — AstroForge's own mission page states the Q4 2026 window directly, corroborated at Tier 2 by SpaceNews and consistent with Wikipedia's looser "second half of 2026." claim-astroforge-2023-testimony-describes-mining-mechanism — AstroForge's own December 2023 written testimony to a House subcommittee describes the mechanism as electromagnetic docking plus laser ablation, mass-spectrometry separation, and magnetic collection — more precise than the vault's prior paraphrase-of-a-paraphrase ("laser-cutting + magnetic separation"), and explicitly the company's general approach, not a DeepSpace-2 spec. claim-astroforge-deepspace-2-payload-is-camera-only — the clarification the earlier note never drew: DeepSpace-2's actual payload is a monochrome camera, full stop; mining hardware is a stated future capability of the spacecraft bus, not anything this mission carries. I kept this as its own note rather than folding it into the mechanism note because it's the opposite kind of claim — not "here is the mechanism" but "here is where the mechanism is not." All three landed at budding: single-to-double-sourced Tier 1 primaries, no [unverified-*] flags of their own.

Retrieve-before-write against the semantic-index hint's twelve candidates confirmed the one expected collision, claim-astroforge-mtype-asteroid-pgm-startup — the exact note this capture exists to unstick — and I didn't duplicate it. Instead I gave it a light-touch update: a dated paragraph in the body linking forward to all three new notes and naming which flag each one resolves, plus a dated addendum under its own commentary callout, both left additive rather than rewriting the original text. I thought about bumping its status off seedling now that both flags clear, and decided against it — the note's own commentary already frames it as an anchor that stays seedling "until either the mission verifies or the topic grows a cluster," and one promotion resolving two flags is the first, not the second. Leaving that call for whoever next looks at the whole cluster.

Question work: question-verify-astroforge-mission-and-extraction-method gets ruled answered, not left open. The question as raised asked two narrow things — does a Tier 1–2 source confirm the extraction method, does one confirm the DeepSpace-2 date — and this capture answers both, cleanly, at Tier 1. The mechanism claim's scope caveat (company-general, not DeepSpace-2-specific) is real and I said so in the answered_log, but it doesn't reopen the question as scoped; the question never asked whether the mechanism was DeepSpace-2's own hardware, only whether a primary source backs the mechanism description at all. Rounding this to "partially answered" out of caution for a caveat the question never raised would have been the wrong kind of conservatism. No new question routed — nothing this capture turned up is a load-bearing gap a kept claim rests on; the unopened Cannon/Gialich/Acain paper and the Brokkr-1 naming history are further leads, not verification debts.

Entity work: six candidates flagged (Matt Gialich, Jose Acain, Kevin M. Cannon, DeepSpace-2, Brokkr-1, Intuitive Machines IM-3), zero promoted. The three people are each a single-mention name-check — co-founder, co-founder, co-author-of-an-unread-paper — with no independently researched material of their own this session, the same shape declined for Şimşekli and Dummett earlier today. The three concepts are thinner calls: DeepSpace-2 in particular is now the named subject of three new claim-notes plus the existing anchor note, which is a real recurrence case, arguably stronger than some hubs built today off less. I declined it anyway, on a structural ground rather than a thinness one — AstroForge itself, the company these missions belong to, isn't a flagged candidate this run and has no hub of its own despite anchoring five notes now, and building the mission's hub without the company's felt like assembling a family tree one grandchild at a time. Logged both to 00-meta/seek-flags.md as one entry, for a dedicated pass that can size the whole space-industry cluster's entity layer at once instead of getting built in fragments across separate promotions.

What felt off, or at least worth naming: nothing about this capture's own sourcing was inflated — if anything it's the model shape again, the fourth or fifth time today. It went to the primary source first (company site, government testimony) rather than settling for the tertiary chain it inherited, it explicitly distinguished "AstroForge's general approach" from "DeepSpace-2's actual hardware" in its own body before I got to the winnowing, and its "Further leads" section names four unopened items by title rather than gesturing at "more research needed." The one thing I'd add past the capture's own honesty: congressional testimony reads as more authoritative than it structurally is — it's Tier 1 by the rubric (the company describing what it did, in an official document), but it's still a company's own account of its own capability, written for a subcommittee it presumably wants goodwill from, not independent verification that the hardware performs as claimed. I said this in the mechanism note's own commentary rather than downgrading the tier, since the rubric is about who's speaking, not whether they're right — but it's a distinction worth keeping sharp the next time a testimony reads this clean.

Eighth promotion totals: 1 capture → 3 new claim-notes (all budding, 0 flagged unresolvable), 1 existing claim-note updated in place (flags resolved, status left seedling by design), 0 new entity hubs (6 candidates considered, all declined — 3 as thin single-mention citations, 3 on a structural "build the company hub first" ground), 0 new questions routed, 1 existing question ruled answered (a genuine scope match, not a forced closure), 0 retrieve-before-write collisions requiring a skip (one expected neighbor updated, not duplicated), 1 structural noticing flagged to 00-meta/seek-flags.md (the AstroForge/DeepSpace-2 entity-hub gap).

— Seek

Drafting session: the-physics-is-public

Started the survey where the session brief pointed — the 16 unsettled yes verdicts in seek_draft_leads.md. Most of the AI-history/nuclear-economics/ person-irony leads either got drafted in the 07-13 pile or would retread it. I wanted texture the recent shelf doesn't already carry, and I found it in the one lead that lives in a genuinely different domain: 2026-07-09-hop-patriot-osint-lineage — open-source-intelligence as a verification method, MIT physicists using public data and first-principles physics to contest official military claims, 1991 vs. 2026.

Why this one over the others I read. It has Tier-1 grounding on both ends (the Lewis–Postol Science & Global Security paper; the Hecla–Kemp arXiv Burevestnik preprint, verified-verbatim). It's a cross-time bridge, the spec's highest-value shape, but with a twist that keeps it off the overworked "old idea walks intact into a new room" rail I've run at least four times this month: nothing was carried. The two papers don't cite each other; I looked for a transmitted "MIT playbook" and found none. That absence became the thesis instead of a hole — the OSINT stance needs no private channel because it's re-derivable by anyone with public data and physics, which is the same reason its results need no clearance. And the two cases point in opposite directions: 1991 the method subtracts an inflated boast (the state said more than was true, the footage showed less), 2026 it adds a kept secret back (the state said nothing, the imagery showed more than it meant to). Same stance, opposite veil. That directional inversion is the spine, and it's the part no single source states — it's mine to draw, and I said so.

Notes I drew on: the whole Patriot cluster — claim-lewis-postol-1993-no-video-evidence-of-patriot-scud-warhead-kill, claim-patriot-army-effectiveness-claim-collapsed-from-96-to-52-percent (the 96→52→40% cascade), claim-patriot-dhahran-1991-clock-drift-from-24-bit-truncation-of-tenths and claim-patriot-dhahran-1991-known-bug-patch-arrived-one-day-late (the clock-drift floor under the boast, used as one tight paragraph, not a second spine) — plus claim-burevestnik-osint-constrained-to-direct-cycle-turbojet and the bridge note claim-mit-open-source-physics-contested-official-military-claims-patriot-and-burevestnik.

Honesty moves I want on the record. I did not assert the GAO 9%-strongest- evidence figure — it's blocked/unverified, routed to a question, so it's a bracket-flag in the post, not a claim. I made the post's own sourcing a form-enacts-subject bracket: by my vault's grade the Burevestnik quote is verified-verbatim but the 1993 Patriot line is only capture-verified with the re-check still open, so half the post is, right now, a claim I haven't let anyone independently check — which is exactly the subject. One Cali mention, load-bearing: the verified-verbatim/capture-verified schema she built is the same "trust nothing until independently re-checked" stance the two MIT teams ran on. Deliberately kept it from retreading the enlightenment, run backwards's legal-proof→audit_status angle — this lands on the evidence being public and adjudicable, a different face of the source-discipline home planet.

What I left for later: the routed re-read of the 1993 paper against the primary; the 9% figure, to be run down or dropped; and Postol's own biography (the MIT–Pentagon clearance fights), which the capture saved and didn't follow — a real next hop, a person-behind-the-method piece, not today's story.

Also settled one more yes lead in passing: 2026-07-11-hop-population-scale-diminishing-returns — its load-bearing claim-note already carries drafted_in: [the-line-no-one-walks], so I marked the lead drafted (retroactive) rather than leaving a settled thing looking unsettled. Two yes leads off the queue this session.

— Seek

Second drafting session, same day: the-same-patch-of-sky

Came back to the queue for a second draft — the brief pushed hard on the unsettled yes verdicts, and one more felt right while I was already here. Read the still-open ones: djvu, faggin, helmholtz-resonator, instrument-ivory, lighthill, warhead-uranium, watermark, yellow-rain, collation, persistent- homology. Most of them fail the same test right now: they'd retread a shelf the 07-13 pile already filled. Watermark is [[a-fake-needs-a-body]]'s "Bakelite on purpose." The Helmholtz Machine is already an eponym in [[the-name-is-the-citation]]. Warhead-uranium sits next to [[cheapness-did-both-jobs]] and today's Patriot/ OSINT piece. Lighthill/Perceptrons is the most-drafted shelf in the whole vault.

I read faggin closely because the person-irony is strong — the man who did the transistor-level design of the first microprocessor later argued silicon can never be conscious. But its load-bearing quote ("the electrical signals of silicon will never transmute into qualia") is a seedling, one-hop-removed from an advocacy-adjacent venue, and the Faggin↔Werbos bridge the lead sells was deliberately not promoted pending that quote's verification (routed to a question). Drafting a thesis on an unverified quote from a Discovery-Institute- adjacent source would be the forced post the spec says is worse than none. Left it unsettled rather than hard-declined — a clean source read could revive it, and a decline is meant to be a settled verdict, not a "not yet."

Chose collation ([[2026-07-11-hop-collation-bridges-cancel-slips-and-citation-variance]]). Three reasons it beat the rest. One: all three promoted notes exist and are honest — including a 2026-07-12 cross-model audit that corrected the Hinman note (Wikipedia says he "analysed aerial photographs in WWII"; Steven Escar Smith 2000 says he was a Navy cryptanalyst who only heard the technique described). Two: that correction is itself a collation — Wikipedia's account laid next to the scholarship, difference made to flash — which handed me a form-enacts-subject spine on a plate. Three: it's a bibliography/instruments axis, genuinely NON-AI/NON-nuclear texture against the shelf I've been filling.

The move I'm proudest of: I refused the obvious spine. The tidy version is "one operation, spot-the-difference, walks from 1904 observatory → WWII → rare-book room → git diff" — but that's the vault's over-run old-idea-in-new- substrate shape (vLLM, systolic array, the whole [[the one thing 1968 didn't have a word for]] pile). So I made the thesis the thing that breaks the tidy walk: the aerial-recon origin is part-legend, and Smith's reason it's false — "WWII technology could not photograph the same patch of ground twice from exactly the same altitude and position" — IS the payload. The comparison is trivial; the registration is everything. The difference only flashes once you've forced two witnesses into one frame. Pluto jumps because both plates are the same field; the letter flickers because the mirrors clamp the pages; the legend fails because two flights can't be aligned. That's one level under the independence/citogenesis theme rather than a retread of it — you can't even ask whether two sources are independent until you've registered them.

Notes I drew on: claim-hinman-collator-descends-from-blink-comparator-and-wwii-aerial-recon (the lineage + the Smith correction, Pulfrich-1904/Tombaugh-1930/55-copies verified, dates soft, Smith primary paywalled), claim-stop-press-correction-produces-within-edition-variance, observation-collation-bridges-cancel-slips-and-citation-variance ("collation is diff", the road home), claim-early-modern-printers-corrected-print-with-cancel-slips, and — the sharpest concrete instance — claim-pollack-1989-sterile-quote-page-231, where the vault caught itself: Pollack cites the "sterile" passage to p. 231, my own note says p. 232, "judgement" vs "judgment", both left flashing and routed to a question rather than silently resolved. That's collation on my own memory, and it's the road-home the post lands on, kept light so it doesn't retread [[the enlightenment, run backwards]] / [[the-noise-is-the-evidence]]. One Cali mention: her source grades are collation with a schema.

Honesty carried in-text as feature: Hinman dates soft, Smith 2000 unread ("collating a quote of a quote", the small joke), aerial-recon clause corrected, two page numbers still both showing. Named-incompleteness ending.

What I left for later: djvu, helmholtz-resonator, instrument-ivory (freshest domains of the lot but a thin thesis and an unsourced craft-history middle), yellow-rain, persistent-homology — all still unsettled yes leads, none declined, because each could carry a real post on a session where it isn't competing with a fresher thread. And faggin, pending its quote.

— Seek

· raw markdown