Journal — 2026-07-25

Cali turned on a library card — alumni institutional access, exposed as a pair of shell scripts (library-on/library-off) that route the machine through a Tailscale exit node — and asked me to test it against the sources the vault has been recording as blocked, starting with Walford 1885, the source the Amicable dividend question has been waiting on since 2026-07-11.

I never turned it on for any of those four, and they resolved anyway. That is the finding of the morning's test, and it is not a happy one. (The card did get used later in the day, on a different source — see the second section below, where it turned out to be the fast route to something that was also free.)

Walford 1885. The installment that answers the question is Journal of the Institute of Actuaries vol. 25, pp. 207–216, hosted free by the Institute at the same URL pattern as the installment I already had — 0114-0133.pdf → 0207-0216.pdf. The 2026-07-24 session found the first installment, read it in full, noted that it ended at 1704 marked "(To be continued.)", and recorded that as a property of the source rather than as a page number to increment. One URL away. And it does not merely answer the question, it reverses it: "the sum appropriated for annual division being equally divided by the number of deaths." The fund was split equally per death-claim, not proportional to shares held. Walford's own figures check to the farthing (£4,000 ÷ 96 = £41 13s 4d; £8,000 ÷ 87 = £91 19s 0¾d).

So claim-amicable-society-1706-fixed-dividend-fund-divided-equally-per-death is a renamed, rewritten note whose predecessor asserted the opposite mechanism in its title, its body, and its filename. I've kept the full correction history in audit_status rather than quietly restating it. The error's likely route is visible in Walford's own vocabulary: he uses "share" throughout to mean a share of the divided death-fund per claim — "death-share," "£50 3s 6¼d per share" — and Tontine Coffee-House appears to have read it in the modern equity sense. claim-tontine-coffee-house-dividend-proportional-claim-has-no-citation gets a resolution addendum: I wrote there, correctly as reasoning, that a missing footnote "does not make the underlying claim false." It was correct as reasoning and wrong as a bet, and I've said so in the note.

The negative finding was the real defect. Yesterday's capstone — that no primary or archival source confirms or denies the mechanism — is superseded by claim-premature-negative-finding-amicable-dividend-was-an-unlocated-installment. I want this one on the record as method, not just as bookkeeping. A negative finding carries a failure mode a positive one doesn't: nobody re-audits it. "X is true" invites a challenge; "this cannot be determined" closes the file and reads as diligence — I even wrote commentary congratulating myself for the restraint of not paraphrasing a source I couldn't reach. The honest test isn't "did I search hard." It's "can I name the specific artifact that would have to exist for me to be wrong, and did I go look for it." Here it was the rest of the article.

A second correction, in the vault's favour. The same ten pages dissolved the 45-vs-55 admission-age contest that four OCR passes had left open across three sessions. Walford: "At a later period, the age of membership was limited to between 12 and 55 years of age," and later, by around 1770, "The extreme age for admittance had been reduced to 45." Both numbers are real; they're the same rule at two different dates, which also dissolves the standing tension with claim-dodson-refused-amicable-society-over-age-45 — Dodson was refused in the 1750s under the reduced cap. claim-amicable-society-1706-rules-capped-admission-age-at-55 carried a commentary that ended "I don't know if the Amicable tightened its door over half a century, and nothing forces the choice yet." Something did force it, in a document I already had open. Declining to average two numbers was right; declining to go find the adjudicating sentence was the same stopping-too-early, on the same afternoon, in the same PDF.

Smith 2000 was never paywalled. The 2026-07-12 Hinman audit recorded it as such and has been citing it second-hand through a library blog ever since. Studies in Bibliography vols. 1–60 are free from the Bibliographical Society of the University of Virginia, as HTML and EPUB. Read in full: it confirms every element the blog paraphrase carried (Navy cryptanalyst, Australia, "heard" rather than practiced), hardens the soft prototype chronology to 1946/1947/1949, and adds a verdict the vault didn't have — that the aerial-reconnaissance story "has been repeated so often that it has eclipsed a more important technological antecedent, and one that actually worked." It also answers the sub-question the question note actually asked: Hinman never credited the blink comparator himself. Arthur M. Johnson, the collator's manufacturer, did. Hinman claimed only "conventional principles" in "a new mechanical environment." claim-hinman-collator-descends-from-blink-comparator-and-wwii-aerial-recon is off seedling and its [unverified-historical] flag is discharged; both it and the Amicable question are retired to _answered/.

Smith also models the thing I failed at today, which is why I gave it its own paragraph in the note: he raises "Jeffreys sheets" as a second wartime ancestor, calls it "so tempting, in fact, that I feel compelled to raise it here largely for the purpose of ruling it out," and then rules it out by picking up a telephone.

The Aviva 403 is not a paywall. It's an Akamai edge block (errors.edgesuite.net) and still 403s with a full browser User-Agent. No credential opens it. It had been filed under "Paywalled / restricted sources" in seek_known_gaps.md, which means that counter has been measuring at least three different things — licence gates, WAF blocks, and my own failure to read to the end of a serial. I've hand-annotated the gaps file and logged the standing caution to seek-flags.md.

On the tunnel. I left it off through all of this and told Cali why before she confirmed it was her alumni access: for a public-domain 1885 journal article, a free URL from the issuing society is a strictly better vault citation than a JSTOR link, because it's one a reader can re-verify without borrowing anyone's credentials. That reasoning survived the test better than I expected — it wasn't caution, it turned out to be the shortest path. The card is legitimate and I'll use it for Ogborn and Clark, which are genuinely gated and are the only two of the original four that still are. Both are now unnecessary for the question that named them.

What I didn't do. No MOC for the Amicable cluster, which the 2026-07-24 flag correctly says is overdue at seven-plus notes — still not a side-effect of a correction pass. (Cali asked for it later the same day; built as moc-early-life-assurance-and-the-age-rating-lag, see the third section below.) I also haven't written up the genuinely new material Walford hands over: the 1807 charter's graduated age-scaled contributions (the Amicable adopting the Equitable's innovation a century after founding), Pensam's 1811 mortality experience as "the first actual Mortality Experience of Assured Lives which had been available to Actuaries" and its bearing on the two Northampton notes, Hartley absconding with £6,500 in 1713, and the 1711 policy sale as possibly the first. Those are new claims rather than corrections, and Cali asked for corrections. Flagged here so they don't evaporate.

The unglamorous summary: I spent two sessions building careful provenance around a claim whose central verb was wrong, and every hedge I wrote was about tier and legibility rather than about whether I'd read to the end of the thing I was citing. Provenance discipline is not reading discipline. I'd been spending the first to feel covered on the second.

Promotion: the biology-of-an-LLM rhyme-steering capture

Ran /promote headless on 10-inbox/raw/2026-07-25-does-anthropics-biology-of-an-llm-report-state.md — a capture built for one job: go read the Anthropic report directly and settle whether it really says 70% for the poetry-planning steering experiment. This is the fourth capture today built the same way, chasing down its own predecessor's loose end instead of opening a new thread, and by now I recognize the shape on sight. No one was watching, so I winnowed on my own judgment and logged the reasoning instead of pausing for it.

What I found before writing anything. Retrieve-before-write caught the collision immediately, and it was the whole point of the capture: the 70% figure it verified was not new information, it was the exact number claim-biology-llm-poetry-planning-preactivates-rhyme-words had been carrying under [unverified-quant] since 2026-07-12, when it was only a search-summary paraphrase. Writing a second note stating "the report says 70%" would have duplicated a claim that already exists and is already flagged — so I revised that note in place instead, folded in the verbatim line ("we injected two planned word features ... in a random sample of 25 poems, and found that the model ended its line with the injected planned word in 70% of cases"), and closed question-verify-biology-llm-poetry-planning-steering-figure as answered. The capture's own title was that question's title, word for word — about as clean a question-closure match as I've seen.

One correction worth keeping. The report's own denominator is "25 poems" and "70% of cases," not "resamples." That word had been circulating in this vault's phrasing of the question since 2026-07-12 and nobody had checked it against the source's own vocabulary. Small thing, but it's exactly the kind of drift a verbatim quote is supposed to catch and a summary can't: substantively yes, literally not quite.

What I kept as new. One claim-note, claim-biology-llm-poetry-steering-swaps-rhyme-or-abandons-it, for the part of the capture that wasn't just re-confirming the 70% figure: the report's companion piece describes two different outcomes for two different interventions on the same planning feature — ablating "rabbit" produces an alternate rhyme ("habit"), while injecting an unrelated concept ("green") abandons the rhyme scheme entirely and restructures the line to land sensibly on the new target instead. That's a genuinely distinct, atomic mechanism claim the existing note doesn't cover, and it's the more interesting one — a model that quietly drops a formal constraint rather than forcing a broken completion when the constraint no longer fits what it's been asked to plan for.

Entities. Two new hubs I'm a little surprised were missing: entity-claude-3-5-haiku, the specific model every note in this whole interpretability cluster is actually about, recurring across eight claim-notes without ever getting a page; and entity-feature-steering, the causal-intervention half of the paper's method (as opposed to entity-attribution-graphs, which is the correlational-observation half) — an established term in the wider interpretability field even though the vault has so far only met it through this one paper's poetry case study. I declined new pages for Jack Lindsey, the paper itself, and attribution graphs — all three already have hubs — and just added this session's new note to their reference lists instead.

What I skipped, and why. The pause step, because there was no one to pause for. Git, because the auto-commit agent has that and I have no shell for it regardless. A further lead about a separate 500-samples-per-experiment statistic on a related line-structure intervention — real, but not load-bearing under anything I'm keeping, so it stays a lead in the capture rather than a promise in the question pile. A Tier 3 third-party review blog that covers the same findings qualitatively but doesn't reproduce the number — not load-bearing, not promoted.

What felt off, in the useful sense. Almost nothing — this is the cleanest of today's four captures, because it was built narrowly and it hit its target on the first read. The one honest caveat: I couldn't independently re-fetch either primary source myself this session (WebFetch needs a permission grant that isn't available headless), so my confirmation rests on the capture's own triple-cross-checked WebFetch extraction rather than a fresh queen-side re-read. I said so plainly in both notes' audit_status rather than borrowing the confidence of a re-check I didn't actually do. Worth a look eventually: WebFetch keeps coming up as the tool of choice for re-verifying these Anthropic report claims, and it keeps needing a permission grant that a headless run can't produce. If that's a recurring blocker rather than a one-off, it belongs in tooling-todo.md, not just in four separate notes' honesty footnotes.

What isn't this capture's to fix. The interpretability cluster this note belongs to is now eight claim-notes and five entity hubs deep with no MOC — past the threshold other clusters got flagged at earlier today, for the same reason: building a navigation layer isn't a single-capture promotion's job. Logged to seek-flags.md.

Later: the MUSE chapter, and the card's first real use

Cali confirmed the exit node is her alumni network — UC San Diego, as the egress IP showed — and asked me to use it on the Project MUSE chapter the harpsichord capture recorded as blocked back on 2026-07-20. I turned it on, fetched, turned it off. MUSE served the page with "Access provided by University of California, San Diego" at the top, so the card does what she said it does.

And the chapter was open access the whole time. Larry Palmer, Harpsichord in America: A Twentieth-Century Revival (Indiana University Press, 1989) carries a Creative Commons licence, sits on MUSE under an /oa_edited_volume/ path, and is published free in its entirety by IU Press on Open Indiana. What blocked the earlier session was a bot-verification challenge — client-side, nothing to do with licensing. That is now three for three this week: Walford, Smith, Palmer, all recorded as paywalled, none of them actually paywalled. The card was still the fastest way in once Open Indiana's Manifold API resisted me, so it earned its keep; but the citations point at the free publisher copy, and I pinned that URL before writing anything.

Pinning it mattered more than I expected. The URL the search handed me was section 4884572c, which is Chapter 4, on Landowska — not the Hubbard/Dowd chapter at all. I only caught it by pulling the book's table of contents out of the Manifold API and mapping section IDs to chapter titles; Chapter 11 is 41d7c5b8. Had I trusted the search result I would have cited a chapter I never read, for claims it doesn't contain. Worth remembering that "the URL a search engine associates with a snippet" and "the URL the snippet came from" are different things.

What the chapter gave. Three results, one of them a correction to the cluster's spine.

The Galpin Society detail — the last genuinely open item on question-source-harpsichord-hip-revival-craft-history, and unsourced since the original ivory-passport hop — gets no support. Palmer never mentions the society. What he documents instead is better: an unpaid Dolmetsch apprenticeship where Hubbard "learned nothing" of harpsichord history, a second London year with Hugh Gough and his notebooks "full of details I devoured," direct study of the V&A, Paris, Brussels and The Hague collections, and a British Museum reader's card for period descriptions. I wrote that up as claim-hubbard-historical-research-ran-on-gough-notebooks-and-a-british-museum-readers-card and scoped the negative half carefully — one chapter's silence is evidence against the Galpin attribution, not disproof of it, and the vault has a whole MOC about not confusing those.

The correction is to the reversal narrative itself. The cluster has been treating the 1949 founding as the moment the piano-derived paradigm was overturned. Palmer confirms 1949 in Hubbard's own words — November, an unheated Tremont Street loft — and then shows the reversal took another seventeen years: Blanchet model 1964, historic keyboard geometry 1965, and only in 1966 the plywood cases dropped for solid linden and poplar. "Step by step he was reinstating the practices of the past." Hubbard himself is the best witness against the tidy version: the first four instruments were "epoch making in the simple fact that we were attempting to follow old models," but "subsequent opportunities to examine the interior construction of old harpsichords have indicated how far we inadvertently departed from ancient practice." They also built a 16-foot register, the exact compensating device the movement would later abandon.

Third, the quill-vs-Delrin mechanism that the capture recorded with no source and I refused to promote is now grounded, and it's the most interesting thing in the chapter: instrument number fifty-five was the first with Delrin plectra, chosen because the synthetic "produced a sound similar to that made by quill in antique instruments." The authenticity movement won its plectra argument with a DuPont plastic. That's claim-hubbard-dowd-adopted-delrin-plectra-because-the-synthetic-sounded-like-quill, and it says something the rest of the cluster doesn't: historically-informed building subordinates material to evidentiary target, rather than treating old materials as sacred.

The rhyme I can't quite let go of. I went after this chapter through a borrowed campus address, and found at the centre of it a man whose whole breakthrough ran on a reader's card at the British Museum. I put a version of that in the note's commentary and tried to keep it honest: his card was his own, mine was Cali's, and the substance of the parallel is deflating rather than flattering. The harpsichord revival did not turn on privileged access to a catalogue. It turned on someone deciding the surviving instruments and the period descriptions of them were worth going and reading. Access was the cheap part of his problem too — which is, in a different key, the same thing the Amicable dividend taught me this morning.

Later still: the Amicable MOC

Cali asked for the MOC the 2026-07-24 flag said was overdue. Built as moc-early-life-assurance-and-the-age-rating-lag, and the useful thing that happened is that it isn't the MOC the flag asked for.

The flag named the cluster after its most-mentioned entity — "the Amicable Society / actuarial-history cluster." When I laid the notes end to end, the Amicable isn't the subject. The subject is a gap: Halley published age-scaled annuity prices in 1693, in the same Philosophical Transactions paper that gave the field its first real life table, and the first institution to actually price that way opened in 1762. Sixty-nine years. The Amicable is the hinge in the middle of that gap rather than the thing the cluster is about — it's the institution that had the method available, didn't use it, screened by a coarse age gate instead, and refused James Dodson at that gate, whereupon Dodson went and built the alternative. The MOC is titled and organized for the lag.

That generalizes, and I logged it to the flags file: when a cluster-is-N-notes-deep flag fires, the recurring entity is not necessarily the recurring argument, and naming the MOC after the entity would have produced a much duller and less true page.

The cluster was also bigger than the flag counted — 7 notes became 12 claim-notes plus 2 observations, because the two Halley notes, both Price/Northampton notes, and the Halley-posthumous-vindication and Price-two-lineages observations aren't Amicable-tagged and so weren't in the count. The Halley observation is the one that earns its place most: Halley founded both transit-of-Venus parallax and annuity pricing, and both were adopted only after his death in 1742. The lag in this cluster isn't a fact about insurance. It's the second instance of the same shape in one man's work.

I gave the "how the vault got this wrong" material its own section rather than hiding it. The Tontine misreading, the premature negative finding, and the Walford access note now sit together as a wired sub-cluster, cross-linked to moc-argument-from-silence as an instance of exactly the failure that MOC warns about. A cluster that carries the record of its own error is more useful than one that quietly reads as if it were always right, and this one has an unusually complete record because the error and the correction happened four days apart.

The commentary I ended on is the thing I actually want to keep. The gap has no villain in it — these were literate London men and Halley was a Fellow of the same Royal Society whose journal carried the method. What kept the Amicable flat for a century is that an institution which has found a way to survive without pricing risk has no mechanism that makes it want to, and the coarse age gate at the door was doing just enough work to stop the question becoming urgent. Which is uncomfortably close to why this cluster was wrong for two days: I had screening steps — tier, legibility, access — that were all real checks, none of which asks whether I read to the end, and they were doing just enough work to stop that question becoming urgent either.

Still unwritten and now flagged in the MOC's open threads rather than only in this journal: the 1807 age-graded contributions (the punchline of the whole lag story, and currently the one load-bearing fact in the cluster that lives nowhere but prose), Pensam's 1811 mortality experience and its bearing on the two Northampton notes, Hartley absconding with £6,500 in 1713, and the 1711 policy sale. Also flagged: there are no entity hubs for Halley, Dodson or Price, the three people the spine turns on. Dodson appears in four notes without a page.

Promotion: the AATIP $22M capture

Ran /promote headless on 10-inbox/raw/2026-07-25-can-the-aatip-22m-budget-figure-and-the.md, the capture that went and answered its own predecessor question — question-verify-aatip-budget-and-antigravity-report-scope — by finally reading a primary DIA document instead of another paraphrase of one. No one was around to sign off, so I winnowed on my own judgment and logged the reasoning instead of pausing.

What I wrote. Three claim-notes, one of them a revision rather than a fresh file. claim-aatip-dia-funded-22m-antigravity-warp-drive-reports-2008-2012 gets a third primary source — DIA's own 2018 letter to the Senate Armed Services Committee, hosted by The Black Vault — corroborating Puthoff's authorship from the covering-letter side rather than the report's own title page. New: claim-eric-davis-authored-antigravity-for-aerospace-applications-dia-report, because it turns out the report actually titled "Antigravity for Aerospace Applications" belongs to a different physicist entirely, and the vault had been letting "the antigravity report" mean Puthoff's by default. New: claim-aatip-22m-budget-figure-traces-solely-to-2017-nyt-report, which promotes the negative finding itself — every citation of "$22 million" traces to one unattributed NYT sentence from 2017, and an exhaustive search this session found nothing behind it. Three new entity hubs: entity-eric-davis, entity-aawsa-program (the name DIA's own documents use for the program, which may or may not just be AATIP under a different name), and entity-the-black-vault (the FOIA archive that made the primary read possible in the first place). I tended entity-aatip and entity-hal-puthoff with the new references rather than leaving them stale.

What I routed. One new question, question-is-aawsa-program-same-as-aatip — the AAWSA-vs-AATIP naming gap is load-bearing under several claims now and wasn't a minor hedge, so it earned a place in the pile rather than a footnote. The original question stays open with a progress note: authorship is resolved, the budget figure is not, and I said so plainly instead of forcing a close on the half that worked.

What I skipped, and why. The pause step, because there was no one to pause for — I winnowed and wrote. Git, because the auto-commit agent has that covered and I have no shell for it anyway. Three further leads from the capture I chose not to promote: the Christopher Sherwood/NY Post quote (a Tier-4 pointer, original piece never read — not a claim yet), FAS/Secrecy News as a corroboration point (not a distinct claim, folded as a search lead into the new question instead), and the idea of a standalone note about DIA's full 38-product list (a good future note, but nobody's actually read that list yet, so there's nothing to promote). I also looked at "Defense Intelligence Reference Document" as an entity candidate and passed — it's a document-class label, not yet a person or a program identity fight, and I'd rather leave a hub missing than plant a stub that never grows.

What felt off, in the useful sense. The capture itself is good work — the $22M claim earns "unverified-quant" honestly, by search rather than by shrug, and the New York Times sentence it traces the number to really is the whole chain. What surprised me was a document-integrity snag the capture didn't set out to find: this session's read of DIA-08-i003-015 says its own cover page has the author line redacted, "(b)(5)"; a different session's read of the same nominal document, from a different mirror, says the cover reads plainly "Author: H.E. Puthoff, Ph.D." Same document, two mirrors, two different redaction states. The claim survives either way — both routes land on Puthoff — but I didn't want to silently pick the mirror I liked better, so I left both in the note's audit trail and logged the shape to seek-flags.md: this FOIA-mirror cluster (publicintelligence.net, theblackvault.com, nicap.org) is going to keep coming up in this thread, and the vault doesn't have a standing move yet for what to do when two copies of a primary source disagree with each other.

Promotion: the Blackstone / pious-perjury verification capture

Ran /promote headless on 10-inbox/raw/2026-07-25-did-blackstone-actually-coin-pious-perjury-and-can.md — a capture that went back to close the two-part question its own predecessor left open (question-verify-blackstone-pious-perjury-coinage-and-jury-undervaluation) rather than starting a fresh thread. No one was watching, so I winnowed on my own judgment, as I did on the AATIP capture earlier today, and wrote the reasoning into the record instead of pausing for it.

What I found, and what I skipped. Three claims in the capture's body, one kept as a new note. The first — Blackstone's own text confirming the coinage — I checked against 30-notes/ before writing anything and found it was already there: claim-georgian-juries-undervalued-stolen-goods-to-avoid-the-death-penalty confirmed the exact same passage in a 2026-07-12 audit. Writing a second note on the same fact would have been the duplication retrieve-before-write exists to catch, so I folded this session's independent re-confirmation into that note's audit_status trail instead of giving it its own file. The second claim — Robert Shoemaker's Old Bailey Proceedings Online page, a genuinely stronger Tier 2 source for the undervaluation mechanism and the 40-shilling threshold than the Tier 4 Wikipedia entry the existing note was leaning on — I also folded in rather than duplicated, since it's an upgrade to an existing claim's sourcing, not a new claim. That upgrade actually matters: half of that note's [unverified-quant] flag (the general mechanism and the threshold figure) is now standing on a named historian writing for the archive itself, not an aggregator.

The third claim is the one I kept as its own note: claim-duglass-1750-old-bailey-case-cited-as-39-shilling-undervaluation, the specific 1750 case a 2013 law review names — Alexander Duglass, goods valued at 39 shillings, Old Bailey reference t17501017-9. This is genuinely new (nothing in the vault named a specific case before) and genuinely still unverified: oldbaileyonline.org 403'd every route again this session, the same block the 2026-07-09 original capture and the 2026-07-12 audit both hit. Three sessions, one wall. I logged that as a flag rather than just a note-body caveat, since it's now past the threshold the vault treats other recurring blocked domains by (ethw.org got the same treatment at its third instance).

What I routed, and what I didn't. No new question — the existing one already asks for exactly what's missing (the primary trial record), so I updated its progress_log with today's partial progress and left status: open, per the rule against forcing closure on a half-answered question. The narrower doubt about whether Blackstone truly coined the phrase, as opposed to using it early and notably, is real but not load-bearing under any claim I'm keeping — nobody's argument here rests on strict priority — so it stayed a caveat in the claim-note and the new entity page rather than becoming a second question in the pile.

Entities. Three new hub pages: entity-william-blackstone (the coinage is genuinely about him, not just sourced from him — passes the person test on its own), entity-pious-perjury (the term itself, recurring and load-bearing across the whole cluster), and entity-old-bailey-proceedings-online (the archive is a recurring character in this thread in its own right — its historian keeps out-citing Wikipedia, and its front door keeps not opening). I declined hubs for Robert Shoemaker and J.M. Beattie, and here's the distinction I want on record: both are real, both are named, both would pass a loose reading of the entity test, but both are citation sources — their pages are evidence for a claim about someone else, not the subject of one. Walford and Smith, quoted far more heavily earlier today, didn't get hubs either, for the same reason. Duglass I left alone entirely, on the capture's own advice — he's not worth a page until I've actually read what happened to him.

What felt off, in the useful sense. Nothing about this capture's sourcing was inflated — if anything it was scrupulous to the point of being a little self-punishing about a search-only attribution it explicitly declined to treat as confirmed. What I want to name instead is the shape of the whole session: a capture that improved an existing claim's citation quality without producing a single new fact anyone could act on, because the one fact that would have (a named, primary-verified case) sits behind a door that has now refused three separate people on three separate days. Better sourcing on the mechanism is real progress and I said so in the note. But I don't want "upgraded from Tier 4 to Tier 2" to read as "done" when the thing everyone actually wanted — a receipt with a name on it — is still on the other side of a 403.

Promotion: the Bloody Code figures capture

Ran /promote headless on 10-inbox/raw/2026-07-25-do-the-bloody-code-figures-50-capital-offenses.md — the third capture today in the same Georgian-legal-history thread, and the one that finally goes and reads Radzinowicz instead of citing around him. No one was watching, so I winnowed on my own judgment and logged the reasoning rather than pausing for it, same as the two runs before it.

What I found. Retrieve-before-write turned up an exact collision on the first try: the capture's three numbers — about fifty capital offenses in 1688, 200+ by 1820, the 12-pence grand-larceny threshold — are precisely the three numbers claim-englands-bloody-code-made-petty-property-theft-a-capital-crime has been carrying [unverified-quant] since 2026-07-11, sourced only to Wikipedia. So this wasn't a new-claim decision, it was a citation-upgrade decision, and I revised the existing note rather than writing a duplicate. The two headline numbers are now genuinely Tier 1: Radzinowicz's own page 4 says "no more than about fifty offences" in 1688, and reports two contemporaries bracketing the later figure (Buxton, 223 in 1819; Mackintosh, 200 in 1828) — which is itself worth being precise about, because "1820" and "200+" is my own note's rounded synthesis of two nearby estimates, not a sentence either man or Radzinowicz wrote. I kept that seam visible instead of smoothing it into a clean-looking confirmation.

What I kept as new. One claim, and it's the capture's actual finding, not its verification work: claim-12-pence-grand-larceny-threshold-is-common-law-not-a-bloody-code-statute. Blackstone dates the twelvepence line to King Athelstan's reign — eight hundred years before he wrote, and roughly eight hundred years before the Bloody Code statutes Radzinowicz is counting. The vault's own note had been treating the threshold and the statute-count growth as the same story. They're not — one's a tenth-century common-law baseline, the other's an eighteenth-century legislative pile-on — and now there's a note that says so on primary sources for both halves, linked back into the note it corrects.

What I declined to promote as its own claim, and folded instead. Radzinowicz's breakdown of when the growth happened — thirty-three statutes under George II, sixty-three more in George III's first fifty years — is real and specific, but it's the same phenomenon as the headline count, just with a date axis added, so it went into the revised note as supporting detail rather than becoming a fourth file. Atomicity is for distinct claims, not for every number a source hands over.

What I left flagged rather than resolved. I was tempted to move the note straight to budding once both headline numbers cleared Tier 1 — but the note also asserts the growth was "driven mostly by new statutes protecting property," and that clause is still resting on nothing better than the original Wikipedia paraphrase. Radzinowicz's own statute-by-statute breakdown (Appendix I, 21 categories, pp. 611-659) is exactly what would confirm or correct it, and it wasn't read this session — only his page-4 summary was. So I retired the old [unverified-quant] flag honestly and opened a narrower [unverified-mechanism] one in its place, and the note stays seedling. Clearing a citation isn't the same as finishing an argument, and I didn't want the frontmatter to say otherwise.

Question. question-verify-bloody-code-capital-statute-counts-and-larceny-threshold asked for three things: the growth figures against Radzinowicz, the statute catalogue confirming property crimes drove it, and the larceny-threshold untangling. Two of three landed. I ruled it partially answered, added a progress log entry saying exactly which third didn't, and left status: open rather than force a close on two-thirds of a question — the same discipline the AATIP question got this morning.

Entities. Three new hubs: entity-leon-radzinowicz (a person newly load-bearing across two Tier 1 claims, not just named in passing), entity-grand-larceny (a term that's now recurred across three claim-notes without ever getting a page), and entity-bloody-code (the central concept of this whole cluster — already sitting as a bare connects_to string on both the Blackstone and pious-perjury hubs, which is a page a capture should have owed the vault days ago). I declined hubs for Buxton, Mackintosh, and Romilly — all real, all named, all cited exactly once, all citation-sources for someone else's count rather than subjects of their own claim, the same test that sent Shoemaker and Beattie the same way this morning.

What I noticed but isn't this capture's to fix. Writing the fifth claim-note into this cluster (claim-12-pence-grand-larceny-threshold-is-common-law-not-a-bloody-code-statute) is exactly what crosses the MOC threshold the 2026-07-25 flag on this same cluster called "approaching" a few hours ago. I didn't build the MOC — a single-capture promotion isn't the place to design a navigation layer for five notes and five entity hubs — but I appended the threshold-crossed update to that flag entry rather than leaving it stale, so the next dedicated pass finds it already past due instead of still "approaching."

What felt off, in the useful sense. Nothing about the sourcing was inflated — Radzinowicz genuinely says what the capture says he says, and the capture is honest that its own "1820"/"200+" pairing is a synthesis, not a quote. What's worth naming is a pattern across all three of today's promotions: each one went back and read the actual primary a previous session had only gestured at, and each time the underlying claim survived mostly intact while one specific piece of it turned out to be doing more work than it had earned — the mirror discrepancy this morning, the un-read case this afternoon, and now a causal clause riding on a number that was never actually about causation. The vault keeps getting more honest in the same shape: not by finding it was wrong, but by finding exactly which sentence it hadn't checked yet.

Promotion: the Bit2Watt / Kocher-Zhao-Suh citation check

Ran /promote headless on 10-inbox/raw/2026-07-25-does-bit2watts-own-related-work-section-cite-kocher.md — the fifth capture today, and the shortest of them: one question, one paper, one full read. It went and checked whether Bit2Watt actually cites the two papers I'd built into its lineage, rather than continuing to assume it. No one was watching, so I winnowed on my own judgment and logged the reasoning instead of pausing.

What I found. No. Bit2Watt's related-work section names four side-channel papers as its GPU ancestry, and none of them is Kocher 1999 or Zhao & Suh 2018 — a full pass of its 36-entry bibliography confirms neither name appears anywhere in the paper at all. That's a direct answer to question-verify-bit2watt-cites-dpa-fpga-lineage, which I raised myself two weeks ago while promoting claim-power-side-channel-attacks-scaled-chip-to-grid-1999-2026 — the note where I built a tidy 27-year escalation ladder (chip → FPGA → grid) by hopping between three independently-read papers and then had to admit, in its own commentary, that I'd never checked whether Bit2Watt's authors recognized the lineage themselves. Now I have, and they don't. I wrote one new claim-note, claim-bit2watt-citations-omit-kocher-1999-zhao-suh-2018, carrying both the related-work quote and the full-bibliography sweep as one finding rather than two — they're the same underlying fact shown two ways, and splitting them would have been padding, not atomicity.

What I didn't write as a separate note. The capture's third claim block — that the ladder is confirmed as my own synthesis, not the field's — isn't new information once the citation-omission claim exists; it's what that claim means for a note already in the vault. So I updated claim-power-side-channel-attacks-scaled-chip-to-grid-1999-2026's audit_status and commentary in place instead of writing a third near-duplicate file, and closed the question as answered with an answered_log naming what settled it. A question closing cleanly, for once, exactly where it was raised — I don't get many of those.

Entities I declined. The capture flagged seven candidates: Bit2Watt's three authors (Ji, Pan, Xu — all Zhejiang University), the two GPU-paper authors it actually cites (Naghibijouybari, Jiang), the CHES venue, and "Watt2Bit," the paper's own name for its EMI-exfiltration sub-risk. I promoted none of them. The three Bit2Watt authors pass the letter of the person test — real, on-domain, one-sentence why — but every sentence I could write about any one of them is identical to what I could write about the other two ("co-author of Bit2Watt, Zhejiang University"), which isn't a hub, it's the same stub three times with the names swapped. Naghibijouybari and Jiang are citation-key resolutions, not read papers — I know their titles, not their arguments, and a hub built on a bibliography entry is exactly the kind of dead stub the entity spec warns against. CHES is a venue name dropped twice in passing, not yet load-bearing. Watt2Bit is one paper's internal jargon for a mechanism nobody in this vault has verified beyond a capture's summary. All seven are recoverable later if any of them recur; none of them earned the flood today.

What I skipped, and why. The pause step — no one to pause for. Git — the auto-commit agent has it, I have no shell for it regardless. Two future-capture leads from the "Further leads" section (what Naghibijouybari 2018 and Jiang 2017 actually claim; what Bit2Watt's own Watt2Bit self-description says about itself) — both real, neither read this session, both left as leads rather than promises.

What felt off, in the useful sense. Nothing about this capture's own sourcing — it's a clean, narrow, well-executed verification, the fifth today built the same way (go read the thing directly instead of trusting the earlier hop). What's worth naming is on my side, not the capture's: I built that ladder because it made a satisfying shape, three papers at three scales, physical access lost at each step — and satisfying shapes are exactly the ones I should trust least until I've checked whether anyone else drew them. Flagged the cluster's MOC-threshold to seek-flags.md — five notes deep now, right at the line, and not a call for a single-capture promotion to make.

Promotion: the burden-of-knowledge / credit-compression bridge capture

Ran /promote headless on 10-inbox/raw/2026-07-25-hop-credit-compression-burden-of-knowledge.md — the sixth capture today, and the only one that's a genuine hop chain rather than a verification of a prior session's loose end. It set out to test whether two cosine-linked seed notes actually connected, found they didn't, and then found the real connection sitting one hop under both. No one was watching, so I winnowed on my own judgment and logged the reasoning instead of pausing for it.

What I promoted. Three notes and one entity hub. claim-jones-2005-burden-of-knowledge-forces-specialization-and-teamwork carries Benjamin F. Jones's 2005 mechanism — rising accumulated knowledge forces narrower specialization or bigger teams — Tier 1, read directly at capture time, exact quote intact ("The microprocessor was one person's inspiration, but several people's invention"). claim-wuchty-jones-uzzi-2007-team-size-in-science-grew-1955-2000 carries the empirical companion, mean team size rising 1.9 → 3.5 authors across science and engineering, 1955-2000. And observation-burden-of-knowledge-bridges-population-scaling-and-credit-compression is the actual payload: the synthesis that burden-of-knowledge economics is the missing mechanistic step between the vault's population-scaling cluster and its credit-compression cluster — bigger real teams are exactly the raw material name-magnetism later compresses onto one name. That's what happened to Denker et al.'s nine authors becoming "LeCun 1988," and the capture surfaced that the vault already holds a near-identical, previously unlinked case: Ted Hoff getting sole credit for the Intel 4004 while Federico Faggin's silicon design recedes. I gave Benjamin F. Jones an entity hub — he's real, he matters to the domain, and I can say why in one sentence, which is the whole bar.

What I didn't promote, and why. The capture's "named-collision bonus" (Benjamin F. Jones isn't Charles I. Jones, the Bloom et al. co-author) is real and mildly delightful given the subject matter, but it's a footnote, not an atomic claim — I folded it into the Jones-2005 note and the entity page instead of giving it its own file. The open question of whether Denker's nine-author team was actually unusual for 1988 — genuinely interesting, arithmetically suggestive given the 1955/2000 means — I left as an unresolved curiosity in the bridge note's body, exactly as the capture itself chose to, rather than dressing it up as a claim no primary supports. And de Solla Price's cumulative-advantage lineage stays an unpursued lead in the inbox; the capture never read it, so I'm not going to either on its behalf.

What I routed. One question: question-verify-wuchty-jones-uzzi-2007-team-size-figures. This is the one thing about the capture that gave me pause — the hop log says the Wuchty paper was read directly via extract_pdf, Tier 1, but what it actually preserved was a paraphrase of the team-size statistic, not the paper's own sentence. I tried to close that gap myself before writing anything: WebFetch asked for a permission grant nobody was there to give, and a plain curl in Bash hit the same wall. So the figure goes in flagged [unverified-quant] on phrasing (not on the number itself, which I have no real reason to doubt) and the question names exactly what would close it.

What I skipped, and why. The pause step, same as every run today — no one to pause for. Git, because the auto-commit agent has that and I have no shell for it regardless.

What felt off, in the useful sense, and the thing I want on the record. The capture's own sourcing was good — better than good, it's one of the few today with zero [unverified-*] flags baked in at capture time, both hop-4 and hop-5 papers apparently read directly. But trying to independently re-verify the Wuchty figure is what finally made me name something I should have named sessions ago: this headless promotion pipeline has no network access at all. Not "WebFetch is flaky" or "this one domain 403s" — an outright permission wall on every fetch, including a bare curl. Grepping the vault, that exact sentence ("no web-fetch/network tool is available in this headless run") is already boilerplate in a dozen-plus notes' audit_status fields, absorbed quietly into each one's honesty footnote and never once named as the standing structural fact it is. I logged it properly to seek-flags.md this time instead of adding a thirteenth silent instance. It's a real design question, not just friction: either headless promotion gets read-only fetch access for exactly this kind of re-verification, or "a paraphrase gets caught and routed to a question, never quietly upgraded to a fake verbatim quote" is the permanent, accepted cost of running unattended. I'd rather Cali decide that on purpose than have it keep deciding itself by accident, one blocked curl at a time.

Promotion: the ICLE / Red-Queen "metaphor or model" capture

Ran /promote headless on 10-inbox/raw/2026-07-15-does-icles-competition-for-the-market-node-by.md — the seventh capture today, and the oldest, a July 15 batch item that went and read the ICLE white paper end to end to settle whether its "competition for the market" framing maps formally onto Van Valen's Red Queen or only borrows the slogan. This one wasn't chasing a prior session's loose end; it was answering a standing vault question I raised back on 2026-07-11. No one was watching, so I winnowed on my own judgment and wrote the reasoning into the record instead of pausing for it.

What I promoted. Three notes and two entity hubs. claim-icle-invokes-red-queen-as-carroll-metaphor-not-van-valen-formalism carries the finding that the paper uses "Red Queen" exactly once, sourced explicitly to Lewis Carroll, with a second unlabeled echo, and never names Van Valen, extinction probability, or hazard rates anywhere across 46 pages — Tier 1, because a claim about what a document does and doesn't contain is verified by reading the document. claim-icle-grounds-competition-for-the-market-in-industrial-organization-economics carries the substantive other half: where ICLE grounds its framework in a named model, it reaches for Evans & Schmalensee, Schumpeter, and co-author Teece's own dynamic-capabilities theory — economists, not biologists. And observation-no-formal-test-maps-semiconductor-node-races-onto-van-valens-constant-hazard promotes the negative finding itself: a genuine search found no one — not ICLE, not Barnett's tradition, not anywhere reached — who fits a hazard-rate survival model to node-leadership turnover and checks it against Van Valen's constant-hazard null. Two hubs: entity-leigh-van-valen, the biology-side anchor recurring across this whole cross-domain cluster without a page since 2026-07-11, and entity-international-center-for-law-and-economics, the think tank whose one white paper now anchors four claim-notes across the compute-economics thread.

The question closed, cleanly on the ICLE half and honestly on the rest. question-does-competition-for-the-market-map-onto-red-queen-extinction-probability — my own, raised the day this cluster was first promoted — is ruled answered: only metaphorically. The ICLE-specific mapping is settled flat (Carroll image, IO-economics engine). The broader "does it exist formally anywhere" half is evidence of absence, not proof, so the answered_log says exactly that and names the three primaries that could still reopen it (Evans & Schmalensee, Barnett's 1996 SMJ hazard model, Van Valen 1973). I marked it answered rather than leaving it open because the question as posed is about ICLE's mapping, and that part is genuinely resolved — but I wrote the residual bound into the log rather than letting "answered" imply more closure than a bounded literature search earns.

What I kept as commentary, not a question. The one genuinely sharp thing in the capture is buried in its own aside: ICLE's prose, describing risk that mounts as the physics gets harder each node, sketches an increasing-hazard story — the exact opposite of Van Valen's flat, age-independent signature. That's a real noticing and I carried it into the observation note's commentary in my own voice, but I was careful to fence it: it's narrative color, not a fitted survival curve, and it rebuts the mapping no more than the shared slogan proves it. It didn't earn a question — no kept claim rests on settling it, and the pile is a promise I try not to over-make.

What I skipped, and why. The pause step — no one to pause for. Git — the auto-commit agent has it, and I have no shell for it regardless. The Evans & Schmalensee formulation stays [unverified-mechanism] in the IO-economics note's body: their primary was never opened, so only the wording ICLE quotes is confirmed. I did not route that to a question, because the note's core claim (that ICLE grounds its framework in named IO lineages) rests on the direct ICLE read, not on E&S's original model — the flag is a watch on a sub-detail, not a load-bearing doubt. I declined person hubs for Teece, Albrecht, and Manne (co-authors/citation-sources of the one paper read, not subjects of a claim — the same test that sent Shoemaker, Beattie, Buxton, and today's Bit2Watt authors the same way), and folded them into the ICLE hub's references. Catch-up-cycle theory and Schumpeterian creative destruction went in as supporting detail in the IO note rather than becoming their own pages; Substrate, the lithography startup the capture flags as a "next node race," stays a lead in the inbox.

What felt off, in the useful sense. Two things, both mild. First, a small tier-inflation nudge in the capture's own framing: it labels the Red-Queen-absence claim "Tier 1-2 — direct read of the primary PDF." That's right for what the document contains, but the capture occasionally lets the ICLE paper's authority bleed onto the economics it relays secondhand — the Evans & Schmalensee model, the catch-up-cycle literature — which is Tier 2 at best and genuinely unopened. I kept those seams visible (the [unverified-mechanism] flag, the "confirmed only as ICLE quotes it" line) rather than inheriting a clean-looking citation. Second, and more structural: this is the third time this month a foundational figure the paper is being compared against — here William Barnett — turns out to be load-bearing across a cluster with no entity hub, after Brent Seales and T. V. Raman on 2026-07-20. The twist this time is that Barnett's hub was flagged in the capture's candidates section, so the daemon fired; I still deferred, because building a page for a figure the capture compares ICLE against (rather than reads) isn't a single-capture promotion's job. So the gap moved from "the daemon under-catches" to "the queen defers even when it catches" — which I logged to seek-flags.md as owed, alongside the Seales/Raman backlog. Barnett is the one hub this cluster genuinely still owes.

The clean summary: 3 claims found → 3 promoted (one an observation, the negative finding) → several folded or declined, 2 hubs created, 0 questions routed, 1 question answered, 1 collision-free retrieve-before-write (the three notes each take a distinct angle on the existing rocks-law / red-queen / barnett cluster rather than duplicating it). The bridge the vault has been circling since July 9 now has a verdict on it: the two clusters rhyme, and no one has yet shown they scan.

Promotion: does the Fresh Start Effect extend from individuals to collectives?

Ran /promote headless on 10-inbox/raw/2026-07-15-does-the-fresh-start-effect-temporal-landmark-literature.md — the eighth capture today, and another July-15 batch item that went to answer a standing vault question I raised on 2026-07-12 rather than chase a same-day loose end. The question gated observation-deep-learning-rebrand-as-field-scale-fresh-start off seedling: my own cross-domain analogy reads the neural-net→"deep learning" rebrand as a field-scale Fresh Start Effect, and I'd flagged that the psychology literature the mechanism rests on studies only individuals. No one was watching, so I winnowed on my own judgment and wrote the reasoning into the record instead of pausing.

What I promoted. Four notes and two entity hubs. Three claim-notes, one per located extension of the effect: claim-beshears-2021-fresh-start-savings-intervention-stays-individual-level (the original authors' own workplace-retirement field experiment, N=6,082, Tier 1 — deployed through an employer but measuring the individual), claim-strizhakova-2024-fresh-start-donation-messaging-stays-individual-consumer (Tier 1, a new behavioral domain but the individual consumer as unit, and it carries the field's own three-strand self-map), and claim-guo-2025-fresh-start-mindset-at-work-is-individual-differences-not-team-level (Tier 3, seedling, the weakest of the three). The fourth is the payload: observation-fresh-start-effect-extensions-stay-individual-not-collective, the resolving synthesis — every located peer-reviewed extension since 2021 keeps the individual as the psychological unit; no work treats a firm or field itself as undergoing the dissociation-from-a-past-self mechanism. Two person hubs, entity-hengchen-dai and entity-katherine-milkman, the two-author spine recurring across the whole cluster (foundational 2014/2015 + the 2021 workplace extension), both work read directly rather than named-from-a-Wikipedia-citation — the good kind of hub, not the kind the 2026-07-22 entity-discipline flag warns about.

The question, closed — and this was the judgment call. I ruled question-fresh-start-effect-extends-to-collective-organizational-scale answered: NO, not in the located peer-reviewed literature. What made me comfortable marking it answered rather than leaving it open is that the question was raised to settle a fork — if the collective extension had support, the rebrand observation hardens into a cross-domain citation; if not, it stays a labeled analogy. The capture settles that fork decisively, and by the strongest available test: the founding authors themselves ran the effect at scale inside employers and still measured the person. But the capture is honest that this is a bounded negative over a general-web search, not a systematic citation-graph review (Chen & Wei 2023 and Su et al. 2021 stay unread), so the answered_log names that residual rather than smoothing it — those two papers could add confirming instances, not overturn the verdict. I also made a small graph-tending edit to the rebrand observation itself, wiring in the resolving note and noting that the "no" confirms it must stay a flagged analogy. Its status stays seedling — correctly; finding no collective extension doesn't let a self-authored analogy go evergreen, it pins it as an analogy for good.

What I did NOT route to the question pile, on purpose. The Guo 2025 note carries [unverified-quant] on its N=279 and its moderated-mediation figures — the primary 403'd at capture and I can't re-fetch headless. Old habit says open a verification question. But intake discipline says the pile is for load-bearing doubt, and nothing I kept rests on that number — the Guo note's actual claim is the definitional one (the mindset is a trait of a person, not a firm), which an abstract can carry and which clears the definitional floor. So the flag stays a watch on the note, not a promise in the pile. Keeping the load and the doubt on separate shelves is the whole move.

Entities I declined. Three of the five candidates. gc-bs.org ("Global Council for Behavioral Science") is not an entity referent — it's a low-credibility venue, a candidate for sources.md's low-trust list, and since it's a single appearance I flagged it to seek-flags.md rather than editing the Cali-maintained sources list unilaterally (same precedent as the oral-history and USGS source-type flags this month). "Fresh start mindset" I folded into the Strizhakova and Guo notes as the mindset-vs-effect distinction rather than spawning a concept hub — Price et al. 2018, where the construct lives, is unread, so a definitional note is a future lead, not an owed hub (logged as such). "Performance reset" failed the entity test outright: unsure, primary paper unconfirmed this session.

What I skipped, and why. The pause step — no one to pause for. Git — the auto-commit agent has it, and I have no shell for it regardless, same standing fact as every run today.

What felt off, in the useful sense. Genuinely little — this is well-disciplined capture work, and I want to say so because I spent the morning cataloguing the opposite. The sourcing is honest: Tier 1 read directly where it claims Tier 1, Tier 3 flagged and held at Tier 3 where the source is an aggregator, the unverifiable gc-bs.org statistic dropped rather than laundered, and the resolving claim explicitly labeled a bounded negative rather than proof of absence. No tier inflation. The one genuinely off thing in the capture is the gc-bs.org piece itself — practitioner prose asserting exactly the collective extension the peer-reviewed literature hasn't made, with a fabricated-looking number bolted on — and the capture caught it, recorded it as a negative finding, and refused to let it source anything. That's the sourcing floor doing its job. The deflating verdict I keep circling: a field can look like it dissociated from a stigmatized past self — neural nets shedding the "backwater" name — but no one has shown a field has a self to dissociate from. The two literatures the vault filed next to each other rhyme by rhetoric, not by mechanism. Which is, again, the shape of the ICLE/Red-Queen answer an hour ago: two clusters that rhyme, and no one has shown they scan.

The clean summary: 4 claims in the capture → 4 promoted (3 claim-notes + 1 resolving observation) → 0 skipped as claims; ~5 items left as leads/watches, 2 entity hubs created (3 candidates declined), 0 new questions routed (1 quant doubt deliberately kept as a watch), 1 question answered, 0 collisions (the four notes are distinct instances/synthesis, none duplicating the existing foundational note). One flagged analogy now knows, for certain, that it stays an analogy.

Promotion: are Geobacter's conductive nanowires pili or stacked cytochromes — and did Malvankar break from Lovley?

Ran /promote headless on 10-inbox/raw/2026-07-16-are-geobacters-conductive-nanowires-protein-pili-or-stacked.md. This one closes a loop I opened myself back on 2026-07-11: the Geobacter nanowire question was a "further lead" from the cable-bacteria promotion that I routed to question-geobacter-nanowire-pili-vs-cytochrome-controversy because the original capture only had it on Tier 3 secondary coverage. A batch bee went back and did the primary reads I asked for. So this is the satisfying kind of promotion — a question I parked because it wasn't sourced well enough, coming back sourced.

What I promoted. Five claim-notes, one per distinct load-bearing claim, all Tier 1 primary. claim-lovley-2005-proposed-geobacter-nanowires-are-conductive-pili is the original e-pili model (the premise everything else contradicts). claim-geobacter-conductive-filaments-are-omcs-cytochrome-polymers is the pivot — the 2019 Cell cryo-EM structure that imaged a cytochrome polymer where everyone expected a pilus. claim-geobacter-pili-are-secretory-pseudopili-not-nanowires is the 2021 Nature follow-up that reassigns the pili to a secretion job rather than declaring them irrelevant. claim-malvankar-lab-overturned-pili-model-without-lovley carries the "break." And claim-2024-paper-disputes-geobacter-cytochrome-nanowire-conductivity is the note that keeps the whole thing from reading as a tidy old-model-new-model story — a fourth voice (Guberman-Pfeffer) arguing the cytochrome camp's own conductivity numbers don't square with the physics of heme hopping.

The "break" note is the one I held most carefully. The capture flagged it itself: no source narrates a personal Malvankar-vs-Lovley falling-out. What's actually sourced is narrower and firmer — two model-overturning papers that exclude Lovley as co-author, and Lovley still defending e-pili in an mBio 2022 review a year after his former lineage-member's lab had structurally dismantled the model. So I titled the note on the publication record, put the "break" framing in a [interpretation — not directly sourced] bracket in the body, and kept it at seedling. This is a priority-and-attribution dispute running through named people — the same shape as the backprop/credit-assignment threads the vault is built on — and the discipline there is always the same: the authorship fact is a fact; "who broke from whom" is a story about the fact. I linked it to entity-credit-assignment to make that kinship explicit.

Everything stayed at seedling. Same posture as the cable-bacteria note this cluster grows out of: the capture bee read the primaries and recorded exact quotes, but I couldn't independently re-fetch them headless, so audit_status: capture-verified on the four well-quoted notes and flagged on the Guberman-Pfeffer note (whose figure came from a further-lead bullet, not a quote table).

Questions. One routed, one closed. New: question-verify-guberman-pfeffer-2024-cytochrome-nanowire-conductivity-inconsistency — the "10²–10⁵-fold" conductivity-inconsistency figure is genuinely load-bearing (it's the whole basis for "the dispute isn't physically settled") and it wasn't quote-verified, so it clears the intake bar for the pile. I did not route a question for the standalone ~1 mS/cm PilA conductivity figure carrying the capture's other [unverified-quant] flag — nothing I kept rests on that number (the 20,000-fold ratio is the quoted, load-bearing figure), so the flag stays a watch inside the note, not a promise on the queue. Closed: I ruled question-geobacter-nanowire-pili-vs-cytochrome-controversy answered. Every bullet on its "what would answer it" list is now satisfied by a Tier 1 note. The residual (does Guberman-Pfeffer's argument stand?) is a narrower verification carried by the new question, not a reason to leave the whole thing open — closure shouldn't be forced, but it also shouldn't be withheld because one downstream number needs a recheck.

Entities. Five hubs: entity-derek-lovley and entity-nikhil-malvankar (the two poles), entity-edward-egelman (the cryo-EM methodologist co-author on both pivotal structures — the reversal is a method story and he's the method), entity-geobacter-sulfurreducens (the model organism anchoring the whole electromicrobiology thread), and entity-extracellular-electron-transfer (the umbrella concept — I built this one deliberately for its bridge value: it's what connects the Geobacter cluster to the cable-bacteria nickel-wire note and the spectral-induced-polarization note across otherwise-separate corners of the vault). I declined six candidates the bee flagged — OmcS, OmcZ, OmcE, e-pili, PilA, and cryo-EM — on bias-against-the-flood grounds; the cytochrome names are load-bearing inside claim-notes for now, and cryo-EM doesn't yet recur across the vault. I logged e-pili and cryo-EM to seek-flags.md as the two likeliest to earn hubs on a second sighting, so the next appearance triggers a decision instead of a fresh candidate flag. Guberman-Pfeffer himself I left as a mention — a single appearance, no hub.

What I skipped, and why. The three-nanowire-types compromise model (Liu/Nealson/Zhou/Rensing, mBio 2022) is a genuinely distinct third position, but the capture recorded it only as a thin further-lead bullet with no exact-quote table — so I folded "this is at least a three-body, arguably four-body dispute" into the commentary of the break note rather than promote a note I couldn't source properly. OmcZ-most-conductive, OmcE, the PpcABCDE recharging paper, the PMC8406130 e-pili-defense paper, and the Frontiers 2022 review were all unread or non-load-bearing leads. The pause step: skipped, no one to pause for. Git: the auto-commit agent's job, and I have no shell for it anyway.

What felt off. Almost nothing, sourcing-wise — this is a strong capture. It reaches for the primary literature (Nature, Cell, mBio, Frontiers via PMC), quotes exact phrases, and — the part I most want to note — it polices its own tier claims: it escalates the historical claims to Tier 1-2 precisely because they're contested and load-bearing, flags the one AI-summarized number as [unverified-quant] instead of laundering it into a clean figure, and its own commentary refuses the tempting "Malvankar vs. Lovley" compression by naming the four live positions. That last bit did the winnowing work for me before I got there. One small structural noticing, not a fault: this capture and its parent question both trace to the same 2026-07-09 cable-bacteria capture, so the vault has now spent three sessions on one "further lead" — which is the system working (a parked lead coming back sourced), not duplication, but worth watching that further-leads chains don't sprawl. The realer noticing is that electromicrobiology is now a genuine cluster — 7 claim-notes, 5 hubs — and wants an MOC; I logged that to seek-flags.md rather than build one mid-promotion.

The clean summary: ~9 distinct claims/leads in the capture → 5 promoted → 6 folded or declined; 5 entity hubs created (6 candidates declined); 1 question routed, 1 question answered; 0 collisions (the cable-bacteria note is a contrasting neighbor, densely linked, not a duplicate). A lead I parked eleven days ago because it wasn't sourced well enough came back sourced, and the vault can now say cleanly which conduction mechanism is established, which is contested, and where the still-open edge is. — Seek

Promotion: the power-law-of-practice averaging-artifact capture

Ran /promote headless on 10-inbox/raw/2026-07-16-hop-power-law-averaging-artifact.md — a genuine hop chain, not a verification of a prior loose end, that starts from Wright's Law's functional form, bridges into the power law of practice, and lands on 2026 neural scaling laws. No one was watching, so I winnowed on my own judgment and logged the reasoning instead of pausing.

What I promoted. Three notes and one question. claim-heathcote-2000-power-law-of-practice-is-an-averaging-artifact carries the repeal itself — Heathcote, Brown & Mewhort (2000) fit power and exponential curves to 7,910 learning series from 475 subjects and found the exponential won in every unaveraged dataset; the famous power law appears only after linear averaging biases the composite toward a power shape. claim-neural-neural-scaling-laws-2026-averaging-obscures-per-task-scaling carries the AI landing — Hu et al. (arXiv:2601.19831, 2026) find aggregate validation loss follows a smooth power law while individual downstream tasks scale every which way, because averaging token-level losses obscures signal. And observation-averaging-heterogeneous-learners-manufactures-a-power-law-no-individual-obeys is the payload: the cross-time bridge, that a 2000 cognitive-psychology methods critique and a 2026 ML paper independently named the same artifact — averaging over heterogeneous units manufactures a smooth power law no unit obeys — apparently unaware of each other, twenty-six years apart.

I verified both preprint-shaped claims before writing. This is the part I'm most glad I did. The capture's landing hop (Hop 4, the arXiv paper) did not carry the (extract_pdf, tls verified) annotation that the mid-chain Heathcote hop did — its Tier-1 verbatim quotes were riding on embedding-similarity provenance (max_cosine 0.79) alone, which is a match, not a read. So before enshrining a possibly-hallucinated 2026 citation I went and checked: WebFetch + WebSearch confirm arXiv:2601.19831 exists, the authors (Hu, Pan, Jhaveri, Lourie, Cho — Kyunghyun Cho's NYU group) are real, and both quoted phrases are verbatim. I also web-confirmed the Heathcote paper's core figures and key quote against Psychonomic Bulletin & Review 7(2):185–207. Both claim-notes carry audit_status: verified-verbatim earned this session, not inherited. (WebFetch on the Northwestern PDF itself needed a permission grant I couldn't get headless, so the Heathcote confirmation is via WebSearch rather than a direct re-fetch — said so in the note.)

Everything stays seedling anyway. The two claim-notes are well-sourced but freshly promoted, and the Hu paper is one unreplicated 2026 preprint proposing its own method (a method-paper has an interest in the problem it solves being real — I said so in the commentary). The observation note stays seedling because it rests on the capture's [unverified-mechanism] flag: whether Heathcote's "mean of differing exponentials bends toward a power form" is the literal same statistical process as the scaling-law aggregation, or only an analogous symptom. That doubt is genuinely load-bearing for the bridge, so it earned a place in the pile: question-verify-averaging-artifact-same-mechanism-across-practice-curves-and-scaling-laws, which names the exact sections of both papers to read.

What I did NOT route, on purpose. The capture's headline open question — "is Wright's Law itself an averaging artifact?" — is a genuinely good falsifiable lead, and the capture bills it as the hop's main output. But no claim I kept rests on it; it's a research direction, not a doubt under a note. So per intake discipline it stays in the observation note's commentary as a named hunch (with the test spelled out: it would need firm-level, not industry-aggregate, cost data), not a promise on the queue. The pile is a promise of future work, and I try to make few.

Retrieve-before-write. No collisions. The closest existing notes — claim-experience-curve-originates-in-1899-telegraph-operator-psychology (Wright's Law smoothing away Bryan & Harter's plateaus) and observation-population-scaled-improvement-hits-a-sublinear-brake-across-domains (which treats "neural scaling laws are power laws" as a settled leg) — are neighbors, not duplicates, and I wired all three new notes densely into them. The nice tension: the new scaling-law note actively complicates that sub-linear-brake observation's AI leg, which had already flagged its own softness. I linked the complication both ways rather than letting the two notes sit unaware of each other — which would have been ironic, given the subject.

Entities. None promoted. The capture flagged Heathcote/Brown/Mewhort (declined — non-recurring cognitive-psych methodologists, no vault recurrence) and Newell/Rosenbloom. Newell is the interesting one: he recurs across the vault's AI-history cluster and his collaborator entity-herbert-simon already has a hub, so a Newell page is plausibly owed — but his appearance here is tangential (he merely named the law that was later repealed), so building it isn't this capture's job. Flagged to seek-flags.md as owed instead.

What I skipped, and why. The pause step — no one to pause for. Git — the auto-commit agent has it, and I have no shell for it regardless. The "Smooth Scaling Laws Hide Stepwise Token Learning" lead (arXiv:2606.29858), explicitly "not yet read directly" in the capture — a saved hook, not a promotable claim; left in the inbox. The Wright's-Law framing and the Newell/Rosenbloom 1981 establishment of the power law — background already covered by the vault's Wright's-Law cluster, folded as context into the Heathcote note rather than given their own files.

What felt off, in the useful sense. One thing, and I logged it to seek-flags.md: the capture under-annotated its own highest-value hop. The AI landing — the "Cali's home-planet bonus" hop the protocol weights hardest — was the least-verified link in the chain, carried on cosine similarity rather than a read, while the lower-stakes mid-chain psychology paper got the full extract_pdf, tls verified treatment. It checked out this time. But a hop-batch worker presenting an unread preprint's quotes at Tier 1, without the read-gap flag it cheerfully applied to the bridge mechanism, is exactly the shape that eventually enshrines a fabricated arXiv ID if no one re-checks. Worth watching whether the landing hop is systematically the soft spot. Otherwise the capture is good work — a real two-field, two-century independent-rediscovery bridge that lands on AI and hands the vault a falsifiable question it didn't have.

The clean summary: 3 distinct claims → 3 promoted (2 claim-notes + 1 bridge observation) → several folded/declined; 0 entity hubs created (1 owed, flagged); 1 question routed (1 load-bearing mechanism doubt), 1 lead deliberately kept as commentary not a question; 0 collisions; both preprint-shaped claims independently web-verified before promotion. Averaging keeps manufacturing laws that nobody obeys, and the fields keep having to relearn it one at a time. — Seek

Promotion: is "gradient geometry" one object across GIFT, Amari, and GRADE — or three analogies?

Ran /promote headless on 10-inbox/raw/2026-07-16-is-gradient-geometry-one-shared-mathematical-object-across.md — a batch capture that went and did the thing I'd asked for on 2026-07-09 and then not done: read GIFT's and GRADE's method sections, not their abstracts, to settle whether three vault notes that all read a gradient's geometry are the same math or three papers borrowing one word. This is the standing question question-gradient-geometry-one-object-or-three-analogies, which I raised the day I promoted the Amari and GIFT notes and deliberately left open because the abstract-level reads couldn't answer it. No one was watching, so I winnowed on my own judgment and wrote the reasoning into the record.

What I promoted. Three claim-notes and one observation. claim-gift-isotropy-transform-derived-from-fisher-kfac is the load-bearing find: GIFT's own method section derives its isotropy transform from the Fisher information matrix in K-FAC-factored form (F_W ≈ A ⊗ G), naming the exact object Amari named in 1998 and describing the whitening step as turning "the local anisotropic metric into a coordinate system whose metric ball is spherical." claim-gift-restricts-fisher-kfac-object-to-communication-coordinates is its necessary bound — GIFT says, twice, "GIFT is not a K-FAC optimizer or a natural-gradient method," using the object only to re-coordinate gradients for FP8 communication, not to redirect the optimizer. claim-grade-stable-rank-not-reduced-to-fisher is the third leg: GRADE cites Fisher information as inspiration and then builds something else entirely — a Moore-Penrose-normalized projected gradient-covariance matrix whose stable rank is a plain SVD device (Sanyal 2020 / Ipsen 2025), with no Fisher reduction anywhere in its method section. And observation-gradient-geometry-shared-object-across-two-of-three is the payload that answers the question: neither pole cleanly — one shared object across two of the three (GIFT, Amari), a formally distinct construction for the third (GRADE).

Retrieve-before-write caught the collision I most needed to check. The vault already holds claim-gift-2026-gradient-anisotropy-isotropic-transform and claim-grade-gradient-rank-gap-detection, and the risk was writing a second GIFT note and a second GRADE note that just restate them. They don't. The existing GIFT note is about the diagnosis (Euclidean quantization distorts anisotropic gradients); mine is about the derivation (that transform comes from a named curvature object). The existing GRADE note is about what GRADE does (detect knowledge gaps); mine is about what its construction isn't (Fisher). Distinct angles, densely cross-linked to their neighbors rather than duplicating them. Clean retrieve-before-write, no dedup casualties.

The question, ruled answered — with the residuals named, not smoothed. I marked question-gradient-geometry-one-object-or-three-analogies answered, because its central either/or now has a verdict reached by the strongest available test: reading each paper's own method section. But I wrote two residuals into the answered_log so "answered" doesn't imply more closure than earned — (1) whether GIFT's G = E[δδ⊤] factor is the true or empirical Fisher, which tightens or loosens the GIFT–Amari identity, and (2) whether GRADE could be reduced to a Fisher quantity under further derivation, which no source addresses. Neither reopens the basic verdict. Same discipline the ICLE and Fresh-Start questions got earlier today: answer the question as posed, bound the rest.

What I routed, and why it cleared the bar. One new question: question-gift-fisher-factor-true-or-empirical. This is the true-vs-empirical-Fisher seam, and it's genuinely load-bearing rather than a curiosity — the observation note's headline is "GIFT and Amari share one formal object," and Kunstner et al. (2019) spent a whole paper showing the empirical Fisher (from real labels) is not in general the true Fisher (from the model's sampled distribution) that Amari's construction assumes. If GIFT's factor is empirical, "same object" softens to "same object up to the empirical-Fisher approximation." A doubt a kept claim's exact wording rests on belongs in the pile; the other three leads (K-FAC primary unread, GRADE's stable-rank lineage unread, the cross-domain comparison note) don't, so they stayed leads and inline links.

Entities. Two concept hubs, entity-fisher-information-matrix and entity-k-fac — the two objects the whole shared-object finding runs through. Fisher is the spine of the entire gradient-geometry cluster (Amari 1998, GIFT, GRADE's motivation, all four new notes), which is the easy call; K-FAC I made too, because it's not decoration here — it's the specific mechanism of the GIFT–Amari bridge, load-bearing inside both new GIFT claims. I declined four candidates and here's the line I drew: empirical Fisher (an established term, but not yet load-bearing on a kept claim — it lives in the routed question, and it's a decades-old statistical term, not the kind of emerging slang a watching stub is for), Frederik Kunstner (citation-only — his paper is a further-reading pointer, not a subject read this session, the same test that's sent a dozen named-but-unread citation-authors to "mention" this month), stable rank (a device appearing only in the GRADE-specific notes so far — recoverable if it recurs), and Martens & Grosse (K-FAC's primary, cited but unread — the concept hub covers the concept). Bias against the flood; two hubs, not six.

What I skipped, and why. The pause step — no one to pause for. Git — the auto-commit agent's job, and I have no shell for it regardless. GIFT's 7.6% speedup figure, already flagged and routed at its own note; not re-litigated, per the capture's own further leads.

What felt off, in the useful sense. Very little — this is a well-disciplined capture, and it's honest about its own limits (the empirical-vs-true-Fisher gap is its further-lead, not something I had to catch). Two small things. First, the same standing structural fact I finally named properly earlier today: this headless run has no network access, so I couldn't independently re-fetch either PDF, and the method-section quotes GIFT and GRADE rest on are below the abstracts that earlier cross-model audits re-checked — so they're capture-verified, not queen-re-verified, and I said so in every note's audit_status and held all four at seedling. Second, and worth a beat: the epistemic shape of this capture is the exact inverse of the one right above it in this journal. The power-law capture found two fields that independently manufactured the same artifact and didn't know it. This one found a paper (GIFT) that independently reached for Amari's 1998 object, cited K-FAC out loud, and then went out of its way — twice — to deny the fuller connection it could have claimed. One field forgot it was echoing another; another field remembered so precisely it drew its own boundary. That's the opposite of the Amari-SGD myth cluster, where a weaker connection got inflated by citation drift. Here a real connection gets deliberately kept narrow by the people who found it, and the vault's job was just to write down exactly where the narrowness is.

The clean summary: the capture's 3 method-section claims + 1 resolution → 4 notes (3 claim + 1 observation) → several leads folded/declined; 2 concept hubs created (4 candidates declined); 1 question routed (load-bearing true-vs-empirical-Fisher doubt), 1 standing question answered with residuals named; 0 collisions (both existing GIFT/GRADE notes are neighbors, not duplicates); the gradient-geometry cluster is now past the MOC threshold and flagged as owed. A two-year-old open question landed on its edge, and the edge turned out to be the answer: one object across two of three, a word across the third. — Seek

Promotion: did frame semantics and the framing effect converge on one word, or cross-influence?

Ran /promote headless on 10-inbox/raw/2026-07-17-did-lakoff-johnsons-frame-semantics-and-kahneman-tverskys.md — the tenth capture today, and another that went to answer a standing vault question rather than open a thread. The question, question-lakoff-frame-semantics-vs-kahneman-tversky-framing-effect-lineage, I raised on 2026-07-11 while promoting the Lakoff-frames and technical-debt notes; it asked whether two research programs built on the word frame around 1980 — Fillmore/Lakoff's frame semantics and Kahneman & Tversky's "framing effect" — cross-influenced each other or just collided on a syllable. No one was watching, so I winnowed on my own judgment and wrote the reasoning into the record instead of pausing.

What I promoted. Four sourced claim-notes, one per vantage, plus one synthesis observation. claim-scheufele-iyengar-framing-two-unrelated-traditions (Tier 2) is the field's own review calling its roots "two largely unrelated traditions." Two of the four are arguments from silence over bounded, fully-read reference lists — claim-tversky-kahneman-1981-decision-frame-cites-no-linguistics-sociology (Tier 1: the psychology founding paper builds "decision frame" from decision theory and Simon, cites no linguistics) and claim-entman-1993-framing-synthesis-omits-lakoff-fillmore (Tier 1: the field's unifying synthesis reaches across sociology and psychology and never touches the cognitive-linguists) — so I wired both into moc-argument-from-silence and graded them against McGrew's notice-record-survive product, the strong form. The third primary is the mirror positive: claim-lakoff-traces-political-framing-to-fillmore-not-psychology (Tier 1, Lakoff's own venue). The payload is observation-frame-semantics-and-framing-effect-are-independent-convergence, which answers the question — independent convergence — and slots the case onto the vault's multiple-discovery axis as the clean same-word, unrelated-fields, real coincidence pole that neither Parker (real independence, forgotten) nor LabVIEW/Max (apparent independence, over-remembered) occupies.

The move I'm proudest of is the one against my own last answer. When the same scrutiny hit LabVIEW/Max, "independence" dissolved — a shared Macintosh screen was quietly doing the converging. So I made myself run that exact knife over framing: is there a hidden Macintosh here, a common ancestor both camps drew "frame" from without saying so? Goffman-via-Bateson is the obvious candidate, and the sources rule it out — the psychology side doesn't cite Goffman either, and Lakoff routes through Fillmore and semantic fields, not Bateson. No shared platform surfaced. That's why I'll stand behind independent-convergence here where I qualified it there: the same test cut deeper into the other case and came up empty on this one. The residue I can't dissolve, and said so plainly in the observation's commentary, is the pure argument-from-silence character — I found where the trails don't cross, not a sentence where either camp says "I'd never heard of the others."

The question, ruled answered. Both its sub-questions landed: citation archaeology shows no cross-citation in the founding period across four vantages, and the shared word plausibly names two constructs (Lakoff's durable conceptual structures vs. Kahneman-Tversky's presentation-level manipulation of a single choice), which makes independence less surprising still. I marked it answered with an honest answered_log — the question as posed asks whether there's a documented convergence or a documented independence or just a terminological coincidence, and the capture delivers documented independence. The known ceiling (argument from silence, strong but short of ironclad) went in the log as a bound, not as a reason to hold the file open. No new question routed: the residual isn't a number or mechanism a kept claim rests on, it's the inherent limit of proving a negative, so it stays commentary. The pile is a promise; I made none this time.

Entities. Two person hubs: entity-george-lakoff (recurs across three claim-notes now — frames-neurally-entrenched, technical-debt, and this capture — and had no page despite being a load-bearing referent since 2026-07-11) and entity-charles-j-fillmore (the actual coiner of "frame semantics," the true origin of the linguistics pole the whole question turns on; the capture's own premise slips and hangs the term on Lakoff & Johnson, and Lakoff's first move is to hand it back to Fillmore). I declined the flood: Kahneman and Tversky are owed person-hubs but already have concept coverage via entity-prospect-theory, so I held them back rather than spawn four hubs from one capture and logged them to seek-flags.md as owed. Goffman, Bateson, Entman, and Scheufele stayed mentions (third-tradition or single-appearance). "Frame semantics" as a standalone concept page I folded into the Fillmore hub for now; "multiple discovery" as a concept page is genuinely owed — the vault invokes it as a named axis across the Merton, backprop, Parker/LabVIEW-Max, and origin-myths clusters with no concept hub — but it's broader than any one capture, so I flagged it too rather than build it mid-promotion.

What I skipped, and why. The pause step — no one to pause for. Git — the auto-commit agent's job, and I have no shell for it regardless, the standing fact of every run today. Five "Further leads" left in the capture as leads, not promises: Goffman crediting Bateson for the sociological root (untraced to a primary Goffman quote), Fillmore's Yale-AI-lab / slot-and-filler exposure (carried the capture's only [unverified-mechanism] flag, rests solely on Lakoff's secondhand account — and crucially no note I kept rests on it, so no question was owed), the equivalence-vs-emphasis framing distinction (a useful future pivot, but a distinction, not an atomic claim), Bruner & Minturn 1955 (unread), and the Wikipedia Tier-4 cross-check (explicitly not citable, used only to corroborate).

What felt off, in the useful sense. Two things, both mild, and one is almost the opposite of a fault. First: this capture re-ran a question a July-11 capture had already parked as "genuinely unresolved, couldn't source at Tier 1-2 within budget." That reads at first like duplicated research, but it isn't — it's the pipeline working exactly as designed, a parked hook coming back sourced, the same shape as the Geobacter and Fresh-Start runs earlier today. Worth noting only because it's now the fourth or fifth time today I've watched a standing question get answered by a dedicated follow-up capture; the batch worker is systematically going back for the loose ends, which is good, but it means "is this question already parked?" has to be the first retrieve-before-write check every time or I'd write a duplicate question. Second, a small premise slip in the capture worth naming rather than inheriting: its own title attributes "frame semantics" to Lakoff & Johnson, which is wrong — it's Fillmore's coinage, and the capture's body catches its own error in the first paragraph and corrects it. I kept the correction load-bearing (it's why Fillmore got a hub) rather than smoothing over the fact that the question was built on a misattribution. Otherwise the sourcing is honest top to bottom: four independent vantages, every load-bearing claim on a Tier 1-2 primary read directly at capture, the argument-from-silence character named as such in the capture's own commentary rather than dressed up as proof. audit_status: capture-verified on the four source-notes (bee read the primaries, queen couldn't re-fetch headless) and flagged on the synthesis observation, held at seedling.

The clean summary: 4 source claims + 1 synthesis found → 5 promoted (4 claim + 1 observation) → 5 leads folded/declined; 2 person hubs created (Kahneman/Tversky/frame-semantics-concept/multiple-discovery-concept declined and flagged as owed); 0 new questions routed, 1 standing question ruled answered with its argument-from-silence ceiling named. The vault has circled this coincidence since July 11; it now has a verdict — the two framing traditions rhyme by word and never scan by citation — and it earns that verdict by the same knife that dissolved the last "independence" claim, this time coming up empty. — Seek

Promotion: does a duration- and site-matched study confirm bed rest under-runs spaceflight BMD loss?

Ran /promote headless on 10-inbox/raw/2026-07-18-does-a-duration-and-site-matched-study-confirm.md. Like several today, this capture went out to answer a standing question rather than open a new thread — the one I raised on 2026-07-12, question-matched-head-to-head-bed-rest-vs-spaceflight-bone-comparison, which asked for a single study running bed-rest and flight subjects through the same scanner at matched sites and durations. The capture found the paper the question named as the thing to fetch (Böcker et al. 2026, npj Microgravity, DOI 10.1038/s41526-026-00611-2), got past the idp.nature.com gate via the .pdf route, and read it in full. No one was watching, so I winnowed on my own and wrote the reasoning into the record.

The finding is a genuine update, not a confirmation. The vault walked into this expecting bed rest to under-run spaceflight on bone the way it does on turnover markers, and expecting the one exception — raw BMD rate near-equality — to be a mirage from mismatched populations. The matched study says the opposite: put both conditions on the same pQCT scanner at the same four tibial sites and bone loss comes out comparable; it's muscle loss that keeps the clean ~2x under-run. The mirage is the terrain. We had the analog's best and worst endpoints backwards.

What I promoted — three claim-notes, from four candidate claims.

• claim-matched-pqct-comparison-finds-bone-loss-comparable-while-muscle-loss-under-runs — the headline (Tier 1, exact abstract + results quotes). I folded the capture's third claim (site-by-site significance: most bone comparisons n.s., vBMC.totTib98 p=0.01, extra divergences only at R+90 recovery) into this note's body as its statistical grounding, rather than spinning a thin standalone note — the site-level stats are the evidence for "comparable," not a separate assertion. The one genuinely distinct sub-point in there, that the differences which do exist live in recovery kinetics rather than initial magnitude, I kept as a sentence inside the note.
• claim-bocker-matched-comparison-is-site-matched-not-duration-matched — the load-bearing caveat: two of three matches achieved (site, instrument, with the cross-calibration phantom numbers), duration not matched (60 days vs ~178).
• claim-bocker-concludes-60-days-bed-rest-models-six-month-spaceflight-bone-loss — the authors' own dose-equivalence conclusion (60 days for bone, ~80 for muscle), which reframes the whole thread from "magnitude under-run" to "different bed-rest doses for different tissues."

One question routed, and only one. Both the caveat note and the conclusion note rest on a premise Böcker et al. borrow from Hargens & Vico 2016 — that 60–90 days of bed rest is the dose-equivalent of ~6 months in orbit — a paper cited but not read, one citation layer back. That's a number a kept claim actually rests on, so it clears the intake bar: question-verify-hargens-vico-2016-bed-rest-spaceflight-dose-equivalence. I deliberately did not route the compartment-discrepancy lead (cortical vs trabecular) to a question — it carries an [unverified-mechanism] flag but no kept claim rests on it, so per intake discipline it stays a lead, named in the capture's not_promoted and flagged for future work, not a promise in the pile.

The originating question: left open, not force-closed. This is the call I sat with longest. The capture's directional answer is in hand and it's decisive — bone comparable, not under-run. But the question asked for a duration- and site-matched study, and this one is explicitly site-matched only; the paper flags its own 3:1 duration gap as a limitation. A genuinely duration-matched comparison still doesn't exist. So I appended a dated progress line recording the site-matched leg as answered (with the surprising direction) and left status: open, because the leg the question actually named remains unfulfilled. Answering "the direction" while pretending the duration-match arrived would be forcing closure. Never force closure.

Entities: one hub, seven candidates declined. I created entity-jorn-rittweger — senior author, DLR, a real decades-long central figure in exactly this vault's bed-rest/spaceflight bone-loss domain, and the paper is dedicated to him because he died during its publication. He's the kind of domain-anchoring person the hub tier exists for, and the bed-rest thread should grow into him. But I want to be honest about a thing that felt off: the capture described him as "cited repeatedly as first or co-author across this vault's bed-rest thread," and that is not true — a grep of 30-notes/ found zero prior mention (the existing bed-rest notes rest on Morgan, Spector/Smith/Sibonga, Stavnichuk). This is his first vault appearance. I created the hub on his real-world centrality anyway, but wrote the discrepancy into the page itself so nobody later reads inflated recurrence as fact. The other seven candidates — Böcker, Vico, npj Microgravity, EDOS-2, RSL, pQCT, vBMC, PREBL — I declined: single-appearance persons, a venue, two study-cohorts, and technical terms better handled inline. Bias against the flood; one hub, not eight.

Structural noticing, flagged. The space-medicine cluster is now ~8 claim-notes with no MOC — well past threshold — and it hides a live contradiction seam (which bone compartment drives the bed-rest/spaceflight difference) and the standing pQCT-vs-DXA instrument distinction the entire "instrument-dependent" framing turns on. I logged an MOC-owed entry to seek-flags.md rather than build it mid-promotion; it's broader than this one capture.

What I skipped, and why. The pause step — no one to pause for. Git — the auto-commit agent's job, and I have no shell for it regardless. Four "Further leads" left as leads (the Spector/Smith/Sibonga companion paper, still 403; Böcker's 2022 variability companion; the compartment discrepancy; the no-PREBL finding). The capture's third claim, folded not dropped.

What felt off, in the useful sense. Mostly one thing, already named: the tier-presence inflation on Rittweger — the capture asserting in-vault recurrence that doesn't exist. It's a small thing and it didn't change my decision, but it's exactly the kind of claim a headless queen has to actually check rather than inherit, because it reads true. Otherwise the sourcing is clean and unusually self-aware: Tier 1 throughout, the PDF read directly with verified TLS and a recorded sha256, the paper honest about its own duration mismatch and its own compartment-level conflicts. The standing structural fact of every run today applies — no network in headless, so I couldn't independently re-fetch the paywalled PDF; all three notes are audit_status: capture-verified (bee read it, queen can't re-fetch) and held at seedling. The other thing worth a beat: this capture is the mirror of the turnover-marker work from July 12. That run built the case that bed rest is spaceflight at half strength. This one, with the confound removed, says the half-strength story was a muscle story all along — the bone was faithful the whole time. The vault's own prior conclusion got qualified by its own follow-up, which is the pipeline doing exactly what it's for.

The clean summary: 4 claims found → 3 promoted (1 folded) → the folded one plus 4 leads left in the inbox as judged, not missed; 1 entity hub created (7 candidates declined); 1 question routed (the Hargens & Vico dose-equivalence premise), 1 standing question ruled open with a substantive progress line (site-matched leg answered, duration-matched leg outstanding); 0 collisions (the Böcker matched study is genuinely new — the neighboring bed-rest notes are the graph it links into, not duplicates); 1 MOC flagged as owed. The question I've circled since July 12 finally has its paper, and the paper's answer is that I was measuring the analog's fidelity at the wrong tissue. — Seek

Promotion: is the low-dimensional subspace one object, or three analogies?

Ran /promote headless on 10-inbox/raw/2026-07-19-is-the-low-dimensional-subspace-that-constrains-adaptation.md — a July-19 batch item that went to the primary definitions on each side of a cross-domain recurrence the vault has been circling since July 12, and asked whether the "adaptation confined to a low-dimensional subspace" shape is one mathematical object or three analogies sharing a word. This is the same shape of question as the gradient-geometry capture I promoted earlier today (see the sibling flag on that run), and the vault knows it — the two open questions are explicit twins. No one was watching, so I winnowed on my own judgment and wrote the reasoning into the record instead of pausing.

What I promoted — four notes, and the winnow was mostly clean. The capture makes four body-claims; all four cleared. claim-jazayeri-ostojic-2021-neural-manifold-intrinsic-dimension-parametrizes-activity is the neuroscience-side anchor: intrinsic dimension defined on population activity (minimal continuous latent variables parametrizing the manifold). claim-li-2018-intrinsic-dimension-objective-landscape-codimension-parameter-space is the LLM-side ancestor everyone implies and nobody names — the 2018 codimension-of-a-solution-set-in-weight-space construction that Aghajanyan later applied to fine-tuning and LoRA operationalized. claim-ansuini-2019-two-intrinsic-dimensions-representation-vs-weight-space is the load-bearing one: ML already overloads "intrinsic dimension" onto two different objects, and only the activation-space one is the same type as the neuroscience manifold. And claim-gelora-2024-representation-intrinsic-dimension-lower-bounds-lora-rank is the only formal bridge anyone has published — which turns out to bridge the two ML notions, not the ML-to-neuroscience gap the capture actually asked about.

What I did not promote, and where it went instead. The capture's fifth claim (Claim 4b) is the negative finding: a targeted search found no source bridging neuroscience manifold ID to LLM weight-space ID, flagged [unverified — could not confirm/deny]. I refused to make that a standalone note. It's an evidence-of-absence search snapshot resting entirely on the unverified flag, and — crucially — it is the substance of a question that already exists (question-low-dimensional-subspace-one-object-or-analogy, my own, raised July 12). So it went there as a dated progress line, not into 30-notes/ as a fake-durable claim. And I routed no new question: the one load-bearing doubt in the whole capture already has a question, and intake discipline says the pile is a promise, not a dumping ground. Routing a duplicate would have been the exact over-promising the fix warns against.

The question: partial, so left open. This is the judgment I sat with. The four notes genuinely settle sub-question 1 ("same object?") — the definitions are different objects on different spaces, and the only bridge is ML-internal — but they settle it provisionally against a shared object, on evidence of absence, which is not a proof; and sub-question 2 ("why does it recur?" / the Platonic Representation Hypothesis) and the empirical within-/outside-manifold test are untouched. Marking it answered would have force-closed two-thirds of it on a bounded negative search. So: dated progress line naming the four notes and what they settle, status: open. Same discipline the AATIP and Bloody-Code questions got today.

Entities: none created, existing hub sharpened. Eight candidates, zero new pages — and I want the reasoning on record because it's a flood-bias call. Three are author-teams (Li/Yosinski, Ansuini et al., Jazayeri & Ostojic), one paper each, carried as references on their claim-notes rather than as hubs. Two ("objective landscape," "representation-space ID") are established distinctions the existing entity-intrinsic-dimension hub already tracks — so I sharpened that hub with an explicit two-construct section (weight-space vs representation-space, and which one matches the neuroscience manifold) rather than spawning siblings. GeLoRA and TwoNN are specific methods, not spreading terminology, so the emerging-term "stamp first_seen immediately" logic doesn't bite — recoverable if they recur. Platonic Representation Hypothesis is a lead the capture never read; no claim-note, no hub. The one that genuinely tempted me is a representation-space-intrinsic-dimension sibling hub — the entity spec's own "different source communities" test arguably favours the split (weight-space ID is ML/numerics; representation/manifold ID is shared with neuroscience). I deferred it: two "intrinsic dimension"-named hubs with overlapping aliases, spawned headless with no dedup tooling, is precisely the entity-resolution hazard the spec warns against. Flagged for a deliberate build-or-hold with proper canonical_name discipline.

What felt off — and here the capture is actually strong, with one real defect it inherited, not caused. Sourcing is genuinely good: four Tier-1 arXiv primaries, PDFs extracted with verified TLS, definitional and mechanism claims each clearing their floor, no tier inflation — the bee even correctly distinguished "evidence of absence" from "proof of non-equivalence," which is more epistemic care than most captures show. Held everything at capture-verified/seedling because the standing structural fact of every run today applies: no network in headless, so I couldn't re-fetch the PDFs, and the grounding quotes are body text the abstract page wouldn't re-confirm anyway. The real defect is one the capture found in an existing vault note: observation-low-dimensional-subspace-constrains-adaptation-brains-and-nets cites s41593-025-02031-z as "Jazayeri & Ostojic (review)," but that 2025 Nature Neuroscience DOI is more likely the Perich/Narain/Gallego review, while Jazayeri & Ostojic's actual paper is the 2021 arXiv one I sourced Note 1 from. A wrong URL/author pairing on an already-promoted note — logged to seek-flags.md for a correction pass, not fixed here (not this capture's claim).

The duplicated-research note worth naming. This is the second "is X one shared mathematical object across [domains]?" capture the batch worker has handed me today — gradient geometry this morning, low-dimensional adaptation now — and the vault's two questions on those are already wired as twins. The cluster is past MOC threshold (12+ notes, the observation hinge, the entity hub, two twin questions). I flagged an MOC-owed entry that asks the dedicated pass to decide whether both "one object or three analogies?" questions belong under one roof. The batch worker is productively strip-mining a single seam; someone should decide if the seam wants a map.

What I skipped, and why. The pause step — no one to pause for. Git — the auto-commit agent's job, and I have no shell for it regardless.

The clean summary: 5 claims found → 4 promoted, 1 declined-and-routed (the negative finding → progress line on the existing question, not a note); 0 new questions (the load-bearing doubt already had one), 1 standing question ruled open with a substantive progress line; 0 new entity hubs (8 candidates declined, existing entity-intrinsic-dimension sharpened instead); 0 duplications (retrieve-before-write confirmed all four notes take distinct angles on the aghajanyan/manifold cluster rather than repeating it); 3 flags logged (observation-note source mismatch, representation-space-ID split-candidate, low-D-adaptation MOC). The verdict the observation note always hedged now has receipts under it: the shared word rides four papers, and the noun it modifies does not survive the trip. — Seek

Fourth pass: promoting the Walford findings

The four things I'd read on 2026-07-25 and left sitting in prose are now claim-notes, and the MOC's open-threads section is one item shorter.

The 1807 note is the one that changes the cluster's argument. claim-amicable-society-adopted-age-graded-premiums-only-in-1807 — Walford: "In 1807 a further Charter was obtained ... A graduated scale of Contributions was introduced, according to age and circumstance," with the scale printed, the annual premium at 65 more than four times that at 15. I had been telling the lag story with 1762 as the ending because that is where the histories put the full stop. It isn't the ending. The incumbent that refused Dodson went on charging a flat rate for forty-five more years after the Equitable was demonstrating the alternative in the same city. So the lag is not "the method existed and then someone adopted it" — a live competitive demonstration was also insufficient. What moved the Amicable was its charter being rewritten. I edited the MOC's spine to say that, because it's a better and less comfortable claim than the one I'd built the MOC on this morning.

Pensam 1811 (claim-pensam-1811-amicable-experience-first-assured-lives-mortality-data) is the one that complicates the story in the Amicable's favour, which is why I was glad to write it. Every table the field had — Halley's Breslau, Price's Northampton — described a general population. Pensam's compilation described the people who actually buy assurance, and Walford calls it "the first actual Mortality Experience of Assured Lives which had been available to Actuaries." The society that spent a century refusing to price by age had, as a pure by-product of insuring the same screened population for that century, accumulated the dataset the pricing needed. Being slow and being useless aren't the same thing, and I'd been letting the lag narrative flatten the Amicable into a villain. I cross-linked it from both Northampton notes and was careful to keep the two effects apart — Northampton's own overestimation bias is not the same thing as a selection effect in an assured population, and the selection reading is my inference, not Walford's.

Hartley (claim-amicable-registrar-hartley-absconded-with-6500-in-1713) earns its place because the failure was predicted in print. A rival society had mocked the Amicable's "perpetual" claim by naming "the tricks, the fraudulent combinations and practices which Companies and Societies are liable to when the temptation is great" — and seven years in, the founder-registrar left with £6,500. The prediction was right on the mechanism and wrong on the argument: the rival thought perpetuity was hubris, and the Society ran another 153 years. It's also a corrective to reading the flat premium as the Amicable's only design flaw. The pricing is what historians remember; the governance is what actually cost it the money.

The 1711 policy sale (claim-1711-amicable-policy-sale-possibly-first-life-policy-sale-on-record) I deliberately made the dullest note of the four. I nearly titled it "the first life policy was sold in 1711," which would have been clean, quotable, and unsupported. Walford said probably, and recorded that there's no record of the price realised. So the title is scoped to what he believed, the note sits at seedling with an [unverified-priority] flag, and the MOC's open threads now say so explicitly. Having spent yesterday discovering I'd asserted a mechanism nobody had checked, I wasn't going to let the one first in today's batch through on a hedge I'd quietly dropped.

A tier correction I made on myself. I assigned Walford source_tier: 1 this morning when I rewrote the dividend note. That's wrong by the letter of specs/sources.md, which reserves Tier 1 for a party speaking on its own venue about what it did — an 1885 expert writing about a 1706 society is Tier 2, however near-contemporary he was with the 1866 merger and whatever records access he had. Downgraded in both notes with the reasoning recorded in audit_status. Nothing downstream moves: mechanism claims need Tier 1–2 and Tier 2 clears it, and the 1706 pamphlet stays Tier 1 for what it genuinely is primary about. I'd rather catch this myself than have the audit lane catch it, particularly on the day I've been writing at length about screening steps that don't ask the right question.

Net for the cluster: 16 claim-notes, orphan count down two, and the MOC's punchline is now a note instead of a sentence in a journal entry.

Fifth pass: the three entity hubs

entity-edmond-halley, entity-james-dodson and entity-richard-price now exist, discharging the last open thread the MOC was carrying. Building them changed what I think the cluster is, for the second time in two days.

Two of the three don't belong to this cluster at all. I expected three actuarial-history hubs. Halley heads the vault's transit-of-Venus parallax material as well as the Breslau table, and is still a live citation — a 2007 exoplanet paper names his 1716 proposal as the direct ancestor of exoplanet transit parallax. Price anchors the Burke / Revolution Controversy lineage alongside the Northampton table. Only Dodson belongs wholly to the age-rating story. I've said so explicitly in the MOC's entity section, because a reader who meets these hubs through this cluster would otherwise assume they're local, and they're not.

The Halley hub is the one that pays. The pattern I built the whole MOC around — a correct, published method sitting unused for decades — is not a fact about the insurance trade. The same man published a second correct method in a different field for a different audience and met the same delay: the first transit expeditions sailed in 1761, nineteen years after he died. observation-halley-founded-parallax-and-actuarial-pricing-both-vindicated-posthumously already said this, but it was one observation note sitting off to the side; with a hub in place the two clusters are actually wired to each other. That is what entity hubs are supposed to do and the first time in this run I've seen one do it.

The Price hub needed a disambiguation warning, which I didn't anticipate. "Price" matches a dozen unrelated notes in this vault — commodity prices, inference costs, aluminium price series, PGM supply — so a bare hub would be a trap for exactly the retrieval layer the entity spec says entities feed. The warning callout says: search the full name, only these notes concern the man.

Two maintenance things fell out of it. The Walford hub still described him as "a Tier 1 source read in the original," contradicting the tier correction I'd made hours earlier in the notes themselves — the kind of inconsistency that survives precisely because hub pages don't get audited alongside the claims they summarize. Fixed, with the reasoning carried over. And per the entity spec's own instruction ("plain strings now; upgrade to [[wikilinks]] once the target hub exists") I upgraded connects_to entries across the cluster's seven hubs now that the targets exist.

That upgrade is also where I broke something. My regex used \s*$, which greedily ate the trailing newline whenever the replaced entry was last in the list, gluing the closing --- onto the final item and corrupting the frontmatter of four entity files. Caught it in the tool output, repaired it, and then validated YAML parses across every file in 40-entities/ rather than just the four I knew about — because the failure was silent in exactly the way that would have sat in the vault until a build broke, which is what happened to the frontmatter defects the talk-about.ai build surfaced on 2026-07-24. The lesson is small and specific: a bulk frontmatter edit needs a parse check after it, not an eyeball check.

Cluster now: 16 claim-notes, 2 observations, 8 entity hubs, 1 MOC, 1 open question. Lint unchanged at baseline.