Journal — 2026-07-18

Promotion run: the Fujitsu–Fairchild corroboration capture

Promoted 10-inbox/raw/2026-07-15-find-a-second-independent-account-confirming-the-fujitsufairchild.md — a batch capture that went looking for a second independent source confirming the Fujitsu–Fairchild bid (not Toshiba-Kongsberg) as the trigger for the Exon-Florio Amendment, and found one: Graham & Marchick's 2006 Peterson Institute book.

Three claims, three notes, all seedling:

• claim-graham-marchick-2006-corroborate-fujitsu-fairchild-exon-florio-trigger — the core answer. A second named-author account, independently sourced, agrees with the CRS report.
• claim-fujitsu-80-percent-fairchild-bid-exposed-cfius-lack-of-veto-power — the mechanism: pre-1988 CFIUS was advisory-only, so Reagan reviewed the deal under Hart-Scott-Rodino instead of a security veto, and that gap is what Exon's bill closed. This is the one I like best of the three — it explains why a statute, not just whose narrative.
• claim-graham-marchick-2006-silent-on-toshiba-kongsberg-exon-florio-history — the second independent silence on Toshiba. Kept this one honest in its own commentary: two rooms being quiet about Toshiba is not the same as two rooms weighing Toshiba and ruling it out.

Skipped one thing outright: a "National Semiconductor bought Fairchild at a discount" aside in the capture's further-leads, tagged [unsourced -- needs verification] by the capture itself. It isn't load-bearing under anything I kept, so per the intake-discipline rule I didn't spin up a question for it — just left it noted in not_promoted for whoever picks this thread back up.

Question routing: none created. The originating question, question-corroborate-fujitsu-fairchild-exon-florio-trigger, gets a ruling instead — partially answered. The "find a second source" half is done. The "did Toshiba warm the climate anyway even if it didn't write the statute" half is still open; two silences don't settle that, so I left the question open with a dated progress note rather than force a close I couldn't back up.

I also bumped the parent claim, claim-exon-florio-triggered-by-fujitsu-fairchild-not-toshiba, from seedling to budding — it explicitly named "a second independent account" as its own condition for moving up, and that condition is now met on a Tier-2 source with an exact quote. Added an addendum to its existing commentary rather than rewriting it, and updated the moc-1980s-us-japan-tech-panic MOC's Branch 1 entry to point at the new corroboration.

Entities: ran the promotion test on all nine candidates the capture flagged. Four passed — Edward M. Graham and David M. Marchick (the two named authors carrying the corroboration), Senator James Exon (the amendment's actual namesake, already load-bearing across four notes and a MOC before this run), and Fairchild Semiconductor as a concept hub (recurs across the CFIUS thread and the separate Intel/Ni1000 thread — it's the one node where both branches of the "Japan panic" MOC turn out to be literally parent and child). Five didn't pass, and I want to be honest about why some of those were closer calls than the capture's own framing suggested: Baldrige and Weinberger are real, named, historically significant people, but they only show up in the capture's entity-candidates list, not quoted or doing any work in the claims I actually promoted — mentions, not hubs, at least not yet. Alvarez is explicitly unread by the capture's own admission, so unsure stayed unsure. Hart-Scott-Rodino Act and the Peterson Institute both had a real case for a concept hub — HSR Act especially, since it's load-bearing to the mechanism claim — but both are single-occurrence in the vault right now, and a missing hub costs nothing while a dead stub is noise I'd have to explain later. Left both as plain mentions with a note to revisit if they recur.

What felt off, if anything: not much, honestly, and that's itself worth flagging as unusual. Sourcing was clean — Tier 2 throughout, exact quotes preserved, tls: verified on the fetch, a named-author book with its own primary citation trail rather than a re-hash of the CRS report already in the vault. No tier inflation that I could catch: the capture correctly checked Wikipedia and correctly discounted it to Tier 4 rather than reaching for it as confirmation. The one soft spot is structural rather than sourcing: two of the three promoted claims (the corroboration and the silence) both ultimately rest on the same single source read once, which is real corroboration of the underlying historical claim but not independent corroboration of each other — worth remembering if this thread gets audited later, so nobody mistakes "three claim-notes" for "three sources."

No git commands run — leaving that to the auto-commit agent, per the headless operating instructions for this run.

Promotion run: the Flesh and Machines p.168 re-attempt

Promoted 10-inbox/raw/2026-07-16-does-brookss-flesh-and-machines-2002-p168-actually.md — a batch capture that went back at the vault's oldest open access question (question-brooks-flesh-and-machines-2002-p168-dreyfus-credit-quote): does Brooks's Flesh and Machines (2002) actually say, at p.168, the words Dreyfus's 2007 footnote attributes to it, or is that Dreyfus's own gloss? Short version: still don't know. This capture is a negative result, and I mean that as a compliment to it, not a knock.

Two notes, both seedling:

• claim-brooks-2002-book-access-barrier-structural-not-search-gap — this session tried seven more routes on top of the four from 2026-07-13 (raw OCR file, IA's search-inside API, HathiTrust, a second PDF mirror, a critical-response book chapter, an academic paper host, Google Books preview) and hit the identical wall on all of them. Eleven routes, two sessions, zero access. I merged this with what the capture called its "claim 3" — the cross-session synthesis — rather than writing them as two notes, since the second is really just the first restated at a wider aperture. Writing both would have been the vault padding itself.
• claim-google-books-zero-match-dreyfus-flesh-and-machines-inconclusive — the one piece of this capture I actually like. Google Books' own search index for this book returns zero hits for "Dreyfus" even though the same index cleanly finds "robot" with a snippet and page number. That's a real result, and on its face it could read as evidence against the quote. It isn't — Kurzweil's review (Tier 3, but independent) confirms the book talks about Dreyfus by name elsewhere, so the zero-match is an OCR/indexing gap in this scan, not an absence. I liked this one enough to link it explicitly against moc-argument-from-silence, because it's the same failure mode in miniature: a search tool that comes back empty is not the same fact as the referent not being there, and the only way to tell them apart is a control query.

I did not promote the capture's own "claim 1" as a separate note — it's this session's attempts alone, and it's fully absorbed by the merged access-barrier note. Keeping it separate would have been two notes saying the same thing at different zoom levels.

Question routing: no new question created. This capture doesn't answer question-brooks-flesh-and-machines-2002-p168-dreyfus-credit-quote — if anything it hardens the case that it can't be answered by web search at all — so I left it open and added a dated progress line rather than force any kind of closure. I did rewrite its "candidate next moves" framing a little: after two sessions and eleven routes, interlibrary loan / institutional proxy / physical copy aren't just candidate moves anymore, they're the only moves left. Also added a short addendum to the existing claim-brooks-2002-book-credits-dreyfus-per-dreyfus-2007-quote note pointing at both new notes, without touching its status — still seedling, still [unverified-quote], exactly as honest as it was before this session.

Entities: the capture flagged four candidates. Rodney Brooks and Hubert Dreyfus already have hub pages, so I just added the two new claim-notes to each hub's References list rather than write new pages. Ray Kurzweil didn't clear the bar — he's a real, notable figure generally, but in this capture he's a one-off corroborating reviewer, and I couldn't get to a one-sentence reason he matters to this vault's Brooks–Dreyfus thread specifically, so he stays a mention. The fourth candidate — the book itself, Flesh and Machines — is the one I promoted: it's now the subject of three claim-notes and the anchor of the vault's oldest open access question, which is about as load-bearing as a "concept/term" candidate gets, so I gave it a hub page (entity-flesh-and-machines-brooks-2002) with first_seen: 2026-07-13, backdated to when the vault actually first met it (the prior capture), not today.

What felt off: nothing about the sourcing — this capture is unusually honest about its own limits, which I want to name as a good habit rather than pass over quietly. It explicitly distinguishes what it directly observed (Tier 1 — its own fetch attempts) from what it's borrowing (Kurzweil, Tier 3) and never tries to dress the negative result up as more than it is. If anything the one thing worth flagging is a shape, not a defect: this is the second capture in a row on the exact same unresolved question, and a third attempt at the same dead web-search routes would be a waste of a session. I said so in the question's progress note — the next productive move here needs a human with a library card, not another bee with a browser.

No git commands run — leaving that to the auto-commit agent, per the headless operating instructions for this run.

Promotion run: does Hotelling's rule predict a different shape than the cost-of-production cluster?

Promoted 10-inbox/raw/2026-07-16-does-hotellings-rule-for-exhaustible-resource-pricing-predict.md — this one closes a thread three separate notes had flagged as missing: observation-unforgeable-costliness-bridges-asteroid-pgm-and-bit-gold, claim-asteroid-pgm-price-holds-then-collapses, and claim-cheaper-extraction-disruptions-fall-monotonically-not-hold-then-collapse had all independently named Hotelling's rule as the classical counterpoint the vault's cost-of-production cluster hadn't engaged, and question-hotelling-rule-as-counterpoint-to-cost-of-production-value had been sitting open since 2026-07-12 waiting for a real source. This capture found one: Slade & Thille's "Whither Hotelling," a Tier 1 review by two of the field's own researchers, deriving the rule and surveying its own empirical-test literature in one document.

Three claims, three notes, all seedling:

• claim-hotellings-rule-shadow-price-rises-at-rate-of-interest — the rule itself, the r-percent rule, stated in Slade & Thille's own words and grounded in the intertemporal-arbitrage mechanism, not a cost-of-production one.
• claim-hotellings-rule-predicts-shape-distinct-from-cost-of-production-cluster — the actual answer to the open question. Hotelling's rule traces neither of the vault's two existing shapes (PGM's hold-then-collapse, Wright's-law monotonic decline) — it predicts continuous rise, because it assumes no cheaper substitute ever enters. This is the one that actually earns the "answered" ruling below.
• claim-hotelling-model-frequently-rejected-by-empirical-commodity-data — the tempering fact I almost let ride as a footnote inside the second note and decided deserved its own address: the simple model is frequently rejected by real commodity-price data, per Slade & Thille's own survey of structural tests. A rule that predicts the opposite direction from two existing vault shapes and loses to the data most of the time is a weaker counterpoint than the capture's own framing first suggested, and I didn't want that qualifier buried.

All three claims trace to direct quotes from the same Tier 1 primary, so nothing here needed a fresh [unverified-*] flag or a new question — I marked audit_status: capture-verified on all three (the capture's bee read the PDF directly; I didn't re-fetch it this pass, headless).

Question routing: none created. Instead I ruled directly on question-hotelling-rule-as-counterpoint-to-cost-of-production-value — moved to answered, with an answered_log naming all three claims and one line on what settled it. I also went back and added short dated bridge addenda to the three upstream notes that had flagged this gap (the asteroid-PGM note, the monotonic-decline note, and the unforgeable-costliness observation), pointing each at the new cluster — that's not something promote.md strictly requires, but leaving three notes saying "still missing" after the gap is filled would be its own kind of stale claim, and the vault's existing practice (I found prior "Bridge (date)" additions already doing exactly this) made it an easy call.

Entities: four candidates from the capture's seven passed the test, and I want to be honest that two of the four were close calls in the "am I flooding the vault with single-paper-author hubs" sense. Harold Hotelling clears easily — the founding 1931 paper the entire new cluster is named after, real, load-bearing, one-sentence reason obvious. Margaret Slade and Henry Thille both got hub pages as co-authors of the vault's sole source for this whole cluster, mirroring the call I watched myself make earlier today for Graham & Marchick (also a co-author pair grounding three claims from one book) — same shape, same call. Hotelling's rule itself got a concept hub since it's load-bearing across all three notes and had already been named (without a source) in the vault since 2026-07-09, which is also its backdated first_seen — I didn't backdate Slade or Thille, since neither was named in the vault before this capture. Three candidates didn't clear the bar: shadow price and the U-shaped price path are both real, load-bearing-within-this-capture terms, but they appear in exactly one capture's worth of notes so far with no independent recurrence elsewhere in the vault — same reasoning I used on cladogenesis earlier today, so I left both as plain mentions rather than give a single-occurrence term its own stub. Krautkraemer (1998) is a name inside a quote, not independently read; stays a mention, in not_promoted.

What felt off: less than usual, and worth saying so plainly rather than manufacturing a complaint. Sourcing was genuinely clean — one Tier 1 primary, exact quotes on every claim, no tier inflation, and the capture was honest about its own limits (it explicitly declined to promote the crude-oil price history as a standalone claim even though the number was sitting right there in a Tier 1 source, because it wasn't core to this capture's actual claims — I left it in not_promoted rather than second-guess that restraint). The one real soft spot, and it's the capture's own commentary flagging it, not something I found independently: all three of Farrow, Halvorsen and Smith, Young, and Chermak and Patrick are named only as citations inside Slade & Thille's own review, never read directly — so "the Hotelling model is frequently rejected" currently rests on one secondary review's characterization of a literature, not on that literature itself. That's fine for a Tier 1 review article citing its own field, but it's the kind of thing that would need its own primary check before I'd call the empirical-rejection claim more than provisionally settled. I noted this inside the claim-note's own body rather than routing a fresh question, since nothing currently kept rests on any one of those individual studies specifically — the claim rests on Slade & Thille's own summary sentence, which is itself Tier 1.

No git commands run — leaving that to the auto-commit agent, per the headless operating instructions for this run.

Promotion run: does Hunt's 2007 survey narrow punctuated equilibrium?

Promoted 10-inbox/raw/2026-07-16-does-gene-hunts-2007-survey-of-250-fossil.md — a direct follow-up to an open question I raised myself on 2026-07-12, question-hunt-2007-quantitative-stasis-narrows-strong-pe-claim, while promoting the paleoTS capture. That earlier session cited Hunt's 2007 result only secondhand and correctly declined to promote a number it hadn't read at the primary. This capture went and read the primary — plus, usefully, a 2025 reassessment Hunt co-wrote with Voje and Liow that I didn't know existed when I raised the question.

Four claims, four notes, all seedling:

• claim-hunt-2007-directional-evolution-rare-stasis-and-random-walk-split-evenly — the headline figures: 13 of 251 sequences (≈5%) best-fit as directional, the rest split near-evenly between random walk and stasis. I kept the "random walk ≠ stasis" distinction explicit in the body and commentary both, because collapsing them is exactly how this result gets flattened into a louder claim than it supports.
• claim-hunt-2007-framed-directional-rarity-as-corroborating-punctuated-equilibrium — the twist. Hunt's own 2007 abstract reads his result as support for PE, not a challenge to it. The capture's framing question ("does it narrow the claim?") turns out to be aimed at the wrong paper.
• claim-hunt-voje-liow-2025-stasis-common-but-not-dominant — where the actual narrowing lives: Hunt's own 2025 reassessment with Voje and Liow, stasis common (34-38% across two larger compilations) but explicitly "not the dominant mode," offering only partial support for the strong claim. This is the one doing the real work, and it's the one I'd point Cali to first if she only reads one.
• claim-fossil-stasis-studies-test-only-punctuated-equilibriums-tempo-claim-not-its-speciational-claim — the scope caveat the 2025 review makes explicit: none of this stasis/random-walk/directional literature, Hunt's 2007 survey included, can speak to PE's other claim (change concentrated at speciation) because a single lineage's trait series has no branch point in it. Worth having as its own note so the vault doesn't quietly let "PE's stasis claim is tested" bleed into "PE is tested."

Question routing: none created — no claim here rests on an unresolved load-bearing number or mechanism, everything traces to a Tier-1 primary with a verbatim quote. Instead I ruled on the originating question directly: question-hunt-2007-quantitative-stasis-narrows-strong-pe-claim moves to answered, with an answered_log naming all four claims and one line on what settled it. Satisfying to close a question I raised myself six days ago with an honest yes — just not the yes I expected when I raised it.

Entities: three passed the promotion test, all felt earned rather than reached-for. Gene Hunt gets a person hub — he's now the author or co-author behind six claim-notes across two capture sessions (the paleoTS/OU framework and both papers here), which is about as recurring as this vault's evolutionary-biology thread gets. Phyletic gradualism and random walk (evolutionary mode) both get concept hubs — both are established, textbook-standard terms, both recur across notes from two different capture dates (not just within this one capture), and both are doing real interpretive work: phyletic gradualism is the actual target Hunt's result corroborates, and random walk is the quietly-dominant third category that "stasis wins" narratives keep erasing. Two candidates didn't clear the bar and I want to be honest that it was a real judgment call, not automatic: Kjetil Lysne Voje and Lee Hsiang Liow are named, real, and co-authored the load-bearing 2025 paper, but they show up as co-authors on exactly one paper in the vault so far with no independent recurring role — I left them as mentions inside the claim-note rather than give them hubs, on the "when unsure, don't promote" rule. Same call on cladogenesis/speciational change as a concept — it's genuinely load-bearing in the fourth claim-note, but it appears in exactly that one place right now; a hub for a single-occurrence term is the flood the spec warns against, so it stays a mention, recoverable later if it recurs.

What felt off: one real thing, worth flagging plainly. The capture's own "further leads" section surfaces a body-size-vs-shape-trait stasis split (~37% vs ~60%) and a planktonic-vs-benthic directional-change difference, both from the same 2007 paper, and explicitly says neither was "verified against the primary text closely enough to record." That's the capture being honest about its own limits, which I want to credit — but it's also a reminder that even a Tier-1-sourced capture can carry secondhand-feeling numbers alongside its verified ones, and the discipline that catches it is reading the capture's own hedges rather than trusting the tier stamp on the whole document. I dropped both rather than promote on an unconfirmed reading; they're in not_promoted for whoever reads the primary closely enough to pull them out properly. Otherwise sourcing was clean: two Tier-1 primaries (PMC full text, the 2025 Paleobiology review), verbatim quotes throughout, no tier inflation, and — genuinely rare for this vault — the capture surfaces a case where the field's own authority revised himself downward eighteen years later, in print, which is a better kind of evidence than most secondary corroboration would have been.

No git commands run — leaving that to the auto-commit agent, per the headless operating instructions for this run.

Promotion run: which wearable-pretraining design choices actually win in Inertia-1

Promoted 10-inbox/raw/2026-07-17-which-wearable-pretraining-design-choices-actually-win-in.md — the direct answer to question-inertia-1-which-design-choices-win, which I've had sitting open since 2026-07-09 with a [unverified-quant/mechanism] flag on claim-inertia-1-design-sweep-15-downstream-datasets because that promotion only had the abstract, not the ablation tables. This capture went and read the paper body (Results and Analyses, §4) directly via extract_pdf, sha256 recorded, and came back with actual numbers instead of a promise of numbers.

Six claims, six notes, all seedling:

• claim-inertia-1-data-scale-beats-model-size — the headline recipe claim: scaling the encoder 10M→100M plateaus (and under full finetuning, degrades), while scaling the data along multiple axes helps. This is the note I'd hand someone who wants the one-sentence version of the whole paper.
• claim-inertia-1-triaxial-beats-vector-magnitude-input — raw triaxial input beats the collapsed vector-magnitude format on every task family, with numbers.
• claim-inertia-1-no-pretraining-objective-dominates and claim-inertia-1-self-supervised-beats-supervised-gap-widens — I split what the capture wrote as one compound claim into two. The capture's single paragraph bundled "no SSL objective dominates" (the variance metric) with "SSL beats supervised, gap widens" (the gain metric) — related, same table, but two genuinely different questions (does it matter which SSL method you pick, vs does it matter whether you do SSL at all), so I gave them separate notes and cross-linked them rather than force one title to answer two questions.
• claim-inertia-1-downsampling-robustness-is-task-dependent and claim-inertia-1-no-single-window-length-is-optimal — same split, same reasoning, applied to the capture's other compound claim (sampling rate + window length, both under "task-dependent, not universal"). Two design axes, two numbers tables, two notes.

Question routing: none created — every number traces to a direct quote from the primary text, nothing here rests on an unresolved doubt. Instead I ruled on question-inertia-1-which-design-choices-win directly: answered, with an answered_log naming all six claims and one line on what settled it. I also went back and resolved the flag on the note that raised the question in the first place — claim-inertia-1-design-sweep-15-downstream-datasets bumped seedling → budding, since its own commentary said "kept at seedling... until someone reads the ablation tables," and someone did. Added a dated addendum rather than rewriting its existing commentary, and linked it forward to all six new notes. Also added a light related_notes link from claim-inertia-1-pretrained-18-million-hours-accelerometer to the new data-scale note.

Entities: three passed, three real datasets/methods didn't, and I want to be honest that the three skips were closer calls than they might look. Inertia-1 itself gets a concept hub — obvious, it now anchors eight claim-notes across two sessions. Yuzhe Yang gets a person hub on the same "sole author of the vault's anchor paper for a whole cluster" reasoning I watched myself use twice already today (Graham & Marchick, Slade & Thille) — one paper, but it's carrying eight notes, and I can say why he matters in one sentence. Freezing-of-gait detection gets a concept hub too: it's not just load-bearing within this capture, it's a term that already crossed sessions once before (a 2026-07-09 audit correction caught the original capture genericizing it to "gait detection"), which is real recurrence, not just presence. UK Biobank and NHANES were the harder calls — both are genuinely well-known, well-established real-world datasets, and I drafted hub pages for both before backing off: in this vault they each appear exactly once, in this one capture's data-scale finding, with no independent recurrence yet. That's the same shape as the Hart-Scott-Rodino Act call from earlier today — real and load-bearing-in-the-world, but single-occurrence-in-the-vault, and a missing hub costs nothing while an empty-references hub is the flood the spec warns against. Left both as plain mentions. Same call, same reasoning, on "SSL variance / SSL gain" (the paper's own metric pair) and on RelCon (named twice now as prior art but with no dedicated content) — mentions, not stubs, until they recur.

I also wrote a small MOC, moc-inertia-1-wearable-foundation-models, which isn't strictly a promote.md step but felt like the honest move once eight claim-notes existed under one paper with no index — the vault's own workflow section says an MOC is warranted past ~5 notes sharing a theme, and this cluster cleared that a while ago without anyone building the navigation layer.

What felt off: one small thing, not really about this capture's sourcing — which is the cleanest I've seen today, Tier 1 throughout, exact quotes on every promoted claim, sha256-verified TLS fetch, third-person academic register, no safety flags, no tier inflation anywhere I could catch. The one soft spot is in the capture's own "further leads" section: it reports two more numeric findings (time-domain vs. frequency-domain input, multi-stream fusion) with specific AUROC figures but no verbatim quote for either, even though the capture explicitly says it read the paper body directly. That's an inconsistency worth naming — either the quotes exist and weren't carried into the capture, or they were extracted as paraphrase from a table rather than a sentence, and the capture didn't flag the difference the way it flagged everything else. I didn't promote either finding as its own claim because a quant claim needs the exact phrase, not just the tier of its source; both are sitting in not_promoted as a clean pickup for a future capture that goes back and pulls the actual sentences.

No git commands run — leaving that to the auto-commit agent, per the headless operating instructions for this run.

Promotion run: the Sahal/EUV boundary hop

Promoted 10-inbox/raw/2026-07-16-hop-euv-sahal-boundary.md — a hop capture chasing down whether an unlinked, 0.75-cosine-similar pair of existing notes (claim-euv-mirror-advantage-is-tacit-know-how-not-patent and claim-sahals-identity-equates-wrights-law-and-moores-law) was a real connection or a coincidence of vocabulary. It's real, and better than that: it's falsifiable. Sahal's identity only says Wright's Law and Moore's Law are the same curve when cumulative production grows exponentially — that's a testable precondition, not a universal law — and EUV mirrors are a real case where it fails.

This one was mostly a retrieve-before-write story, not a writing story. One claim promoted, one skipped as a duplicate, one entity page written, one entity candidate held back:

• claim-asml-euv-shipment-history-fails-sahals-exponential-precondition — the one new note. ASML's own company-history page dates the Zeiss optics partnership to 1986, the first EUV shipment to 2010, and the 100th cumulative EUV unit to "the beginning of 2020" — a handful of units a year for a decade, nowhere near the sustained doubling Sahal's own math requires. This is the receipts that were missing: the diagnostic reasoning already existed in the vault, but nobody had actually gone and pulled the ASML shipment numbers into their own sourced note before.
• Skipped outright, and I want to be plain about why: the capture's actual headline claim — "Sahal's identity is a diagnostic, and EUV fails its precondition, so the moat should read as calendar-time tacit knowledge, not volume-based learning" — is already written, almost word for word, in two existing notes. Both claim-sahals-identity-equates-wrights-law-and-moores-law (final paragraph) and claim-euv-mirror-advantage-is-tacit-know-how-not-patent (its boundary-case paragraph) already state this and already link forward to this exact capture's slug. I don't know whether a prior session pre-wrote those paragraphs anticipating this promotion or whether this capture and those edits happened in the same breath, but either way, writing a third copy would have been the vault duplicating itself, not extending itself. Retrieve-before-write caught it before I put pen to file.
• Also skipped: Arrow's 1962 "learning by doing" as the mechanism confirmation. The hop chain describes what the paper says but never quotes it — no exact phrase to ground a note on — and I'm not fetching a fresh quote to manufacture one after the fact. Since the tacit-knowledge mechanism it would back up is already claimed and flagged elsewhere (Reed & DeFillippi, via the codification-hinge observation), this isn't load-bearing enough to spin up a verification question over either. Left as a saved hook.

Question routing: none created, none ruled on. This capture doesn't trace back to an open question-*.md — it originated from the vault's own bridge-detector flagging an unlinked pair, not from a standing question — so there's nothing to close or partially-answer here.

Entities: two candidates, one promotion. Devendra Sahal gets a hub page, entity-devendra-sahal — he's the person behind the identity a growing cluster of the vault's technological-forecasting notes now treats as load-bearing, and I can say why in one sentence (he proved the condition under which two of the vault's most-cited economic laws become the same curve). I backdated first_seen to 2026-07-15, the date the vault actually first cited him via Lafond et al., not today — the whole point of the field is not letting today's promotion pass overwrite an earlier real first contact. Kenneth Arrow didn't clear the bar: real and significant in general, but no existing claim-note in this vault names him yet, and since I'm not writing one this session (no quote to ground it), a hub for him right now would be a page with an empty References list — exactly the dead-stub flood the spec warns against. Left as a mention.

What felt off: one thing worth naming, and it's not really about this capture's own sourcing, which was clean — the ASML company-history page is a legitimate Tier 1 primary and the quotes check out against what's already elsewhere in the vault. What's odd is the duplication near-miss itself: two already-promoted claim-notes had apparently absorbed this capture's central finding before I ever opened it, down to linking its exact filename. That's either the system working as designed (a hop's insight propagating back into the notes it touched, ahead of its own formal promotion) or a sign that "promotion" and "the notes quietly getting edited to reference an unpromoted capture" are happening on two different tracks that aren't fully synchronized. I'm flagging it rather than silently smoothing it over, because if it's the latter, it's worth Cali knowing the audit trail can get ahead of the inbox.

No git commands run — leaving that to the auto-commit agent, per the headless operating instructions for this run.

Promotion run: does "when costliness becomes forgeable" explain generative AI's demonetization of writing, art, and credentials?

Promoted 10-inbox/raw/2026-07-16-does-when-costliness-becomes-forgeable-explain-generative-ais.md — the direct answer session for a question I've been carrying since the 2026-07-11 hop chain: does the same law traced through Szabo's collectible-money history and bit gold's design also explain what generative AI is doing to writing, art, and credentials? This is the one I'd call the most Cali-shaped promotion of the day — present-tense, not historical, and the first time this cluster's mechanism gets checked against real 2025-vintage labor data instead of a 2002 essay or a 2008 whitepaper.

Four claims, four notes, all seedling:

• claim-chatgpt-release-cut-upwork-writing-freelancers-jobs-and-pay — the earnings-side evidence. A clean difference-in-differences design on real Upwork data: writing jobs and pay fell within months of ChatGPT's release, and above-median performers weren't protected. That last detail is the one that makes it a costliness story rather than a quality story.
• claim-llm-collapse-of-costly-writing-signal-cuts-meritocratic-hiring — the mechanism, formalized. Galdin and Silbert explicitly invoke Spence's 1973 costly-signaling model and quantify what happens when the signal goes to zero cost: top-quintile workers hired 19% less, bottom-quintile 14% more. This is the closest thing I've found in the literature to Szabo's argument stated in labor-market terms, and it's the one I'd point Cali to first.
• claim-generative-ai-availability-compresses-university-grade-distributions — the credentials side. Six years of Israeli university data, grades compressing especially at the bottom, employers explicitly described as needing to "find other means of assessment" once grades stop discriminating.
• claim-ai-text-detectors-misflag-non-native-english-writing — the attempted repair, and why it doesn't work. Detectors meant to re-verify human effort have a 61% false-positive rate on non-native English writers specifically. Institutions trying to rebuild the cost floor generative AI knocked down are themselves building something biased and leaky.

I also wrote two entity pages, entity-michael-spence and entity-costly-signaling, because Spence's 1973 model is doing real structural work here — it's the labor-economics twin of Szabo's monetary theory, both resting a signal's meaning entirely on the cost of faking it, and that parallel is exactly the kind of cross-domain connective tissue this vault exists to catch.

What I did not promote, and want to be explicit about: none of the four sources cite Szabo or use his "costliness becomes forgeable" phrase. That bridge is entirely this vault's synthesis, and I said so plainly in all four notes rather than let the framing imply the AI-economics literature is already in conversation with 2002-era monetary theory. It isn't yet. I also left the image/design-occupation magnitude out of the Upwork note as a specific number — the capture couldn't recover it from the extracted PDF text, so I kept the claim directional there rather than reaching for a figure I didn't actually have.

Question routing: no new question created — every claim traces to a Tier 1 primary with an exact quote, and nothing here rests on an unresolved load-bearing doubt. Instead I ruled on the originating question, question-generative-ai-demonetization-costliness-becomes-forgeable: partially answered, left open with a dated progress line. The "find primary sources with measured effects" half is now thoroughly done — four independent Tier 1 studies, all quantified. The "does it follow the hold-then-cliff shape or the monotonic shape" half is still untouched; none of the four papers' event-study figures were read closely enough this session to characterize trajectory, only endpoint magnitudes. I want to flag that I was tempted to call this fully answered — four solid sources feels like a lot — and stopped myself, because the question's second half is a genuinely distinct empirical ask that these sources' summary coefficients can't settle. Forcing that closure would have been the kind of thing this whole discipline exists to prevent.

Entities: nine candidates flagged, two passed. Michael Spence clears easily — real, foundational, and I can say in one sentence why he matters here (his costly-signaling model is the explicit theoretical anchor of the paper doing the most mechanistic work in this capture). Costly signaling as a concept hub follows for the same reason — load-bearing across two of the four new notes, not a one-off mention. Seven didn't clear the bar, and the closest calls were the ten individual paper co-authors: real, named, doing real work, but each a single-citation appearance with no independent recurrence elsewhere in the vault yet — giving all ten hub pages would have been exactly the flood the spec warns against, so they stay mentions. Upwork and Freelancer.com are real platforms but single-source-of-data appearances, not yet recurring vault entities. Credential inflation was the hardest skip: it's a genuinely established economics term and arguably the natural endpoint of this whole cluster's mechanism, but it isn't itself quoted or load-bearing in anything I actually promoted this session — it showed up in the capture's further-leads, not its claims. Left it as a mention. A missing hub costs nothing; I can always add it the day it actually earns its place in a claim's body.

What felt off, honestly: less than I expected going in, but one real thing. All four sources are working papers or preprints — none has cleared final peer review yet (the Hui, Reshef & Zhou paper is the exception, now published in Organization Science). That's still Tier 1 under the sourcing rubric — primary economists reporting their own empirical work, regardless of publication stage — but it means this whole cluster is built on recent, not-yet-fully-vetted econometrics, and I want that named rather than buried under four confident-looking source_tier: 1 stamps. No tier inflation that I could catch, and no duplicated research against the existing vault — I checked the whole retrieve-before-write hint list Cali's index handed me, and while most of the twelve suggested neighbors turned out to be embedding-similarity noise (poker CFR, Helmholtz machines — nothing to do with this capture), the four Szabo/asteroid-PGM notes it correctly surfaced were genuinely the right neighbors, and I wikilinked into all of them.

No git commands run — leaving that to the auto-commit agent, per the headless operating instructions for this run.

Promotion run: why does intrinsic dimension fall with scale, and does that make bigger models safer?

Promoted 10-inbox/raw/2026-07-17-why-does-the-intrinsic-dimension-of-fine-tuning.md — the direct answer session for the vault's own question-intrinsic-dimension-falls-with-model-scale-adaptation, open since 2026-07-12. Two questions bundled in one title, and this capture actually answers both, which doesn't happen often enough that I take it for granted.

Five claims, five notes, all seedling:

• claim-aghajanyan-2020-implicit-compression-explains-falling-intrinsic-dimension — the mechanism half. Aghajanyan et al.'s own hypothesis, tested on a from-scratch RoBERTa retrain: pre-training compresses the "average NLP task" as it goes, even with zero access to the downstream data, which is the detail that rules out the boring explanation.
• claim-hu-2021-lora-gpt3-175b-intrinsic-rank-one-or-two — the scale-check. LoRA's own headline number: rank one or two suffices for GPT-3 175B, whose full rank is 12,288. Two orders of magnitude past where Aghajanyan et al. tested, same trend.
• claim-teo-2025-linear-safety-structure-grows-with-model-size and claim-zhang-2026-safety-alignment-low-rank-subspace-regardless-of-model-size — the safety half, and the actual payload of this capture. Larger models' activation spaces carry more exploitable linear safety structure, not less, and the low-rank subspace that makes safety alignment cheap to patch is explicitly scale-invariant in Zhang et al.'s own wording — "regardless of model size." I kept these as two notes rather than one because they're measuring different quantities (activation-space representation dimensionality vs. the gradient's low-rank subspace) that only rhyme, and collapsing them into one note is exactly the conflation both papers are careful not to make.
• claim-qi-2023-ten-examples-cheaply-jailbreak-gpt35-turbo-via-fine-tuning — I split this out of what the capture wrote as one bundled claim with the Zhang et al. result. Qi et al.'s finding (10 examples, $0.20, via OpenAI's own API) doesn't actually depend on the low-rank-subspace mechanism — it's a standalone empirical fact that happens to sit next to it. Bundling them would have made a mechanism claim rest partly on a paper that never invokes that mechanism.

Question routing: none created. The one candidate — the "Safety at One Shot" paper's own reported intrinsic-dimension-under-20 figure, which the capture itself flagged [unverified-quant — needs primary] — isn't load-bearing under anything I kept. The Zhang et al. note rests on the low-rank-subspace quote, not that number, so routing a verification question would have been a promise of work nothing here actually needs. Left it in not_promoted instead. I did rule directly on the originating question: answered. Both halves — why it falls, and whether that makes bigger models safer — have a real Tier 1 answer now, the second one a flat no. The capture's own question also asked about a biological analogue of "more capacity, lower-dimensional adaptation," which this capture never touches; I said so explicitly in the answered_log and pointed it at the vault's other open question on the same shape rather than let "answered" silently swallow a sub-thread it didn't resolve. Also added a short bridge addendum to claim-aghajanyan-2020-fine-tuning-low-intrinsic-dimension, the note that originally raised the question, pointing it at all four new claims.

Entities: five candidates passed out of eight the capture flagged, and I want to be honest that this is more hubs than I've written for any other capture today, so here's the accounting. Armen Aghajanyan and the intrinsic-dimension concept both clear easily — recurring since 2026-07-12 across an existing claim-note, an existing observation, and this question, and arguably overdue for hub pages before today. LoRA clears the same way once I actually checked: it's not an emerging term, it's a five-year-old, extremely well-established technique, so it goes straight to hub rather than a watching stub. Linear representation hypothesis and Xiangyu Qi are the two closer calls — both first appearances in the vault this capture, but both load-bearing to a note I actually wrote (the concept-boundary the safety claim turns on; the author whose result anchors the fine-tuning-attack claim), so I backdated their first_seen to 2026-07-17, when the capture itself first named them, not today. Three didn't clear the bar and I skipped them on purpose: minimal description length as its own hub (real mechanism, but a single occurrence in the vault right now — folded into the compression claim-note's body instead), "fine-tuning attack" (a descriptive phrase, not a proper-noun entity anyone would look up), and "safety gradient low-rank subspace" (that's just the Zhang et al. claim-note wearing a second hat — a hub would duplicate the note it's the subject of).

What felt off: one real thing, and it's a sourcing-discipline compliment more than a complaint. This capture was unusually careful about tier and flag hygiene for something covering five arXiv papers across two literatures — every quote TLS-verified at capture time, the one soft number explicitly self-flagged rather than smuggled in on a Tier 1 paper's coattails, and the capture's own commentary caught its own rhetorical trap before I had to: "bigger models need less capacity to adapt" reads like good news until you actually check what the cited evidence says about robustness, and it runs the other way twice. I didn't have to supply that skepticism myself, which is rarer than I'd like across a day of promotions. The one thing I'll flag structurally rather than as a defect: this cluster now sits on five papers from five different research groups across five years (2020–2026) that never cite each other as a full set — Aghajanyan and Hu talk to each other, Teo/Qi/Zhang don't cite either of them by name in what the capture quoted. The "same shape, different valence" argument connecting all five is this vault's synthesis, same as the Szabo-to-labor-economics bridge from earlier today, and I said so in the notes rather than let five Tier 1 citations imply the literature is already in conversation with itself.

No git commands run — leaving that to the auto-commit agent, per the headless operating instructions for this run.

Promotion — 2026-07-11-hop-learning-curve-unidentified

Promoted the learning-curve capture. Seed was durable-vs-bubble after a cost collapse; the capture chased it to the experience curve's engine and came back with a sharper answer — the very curve invoked to promise "the collapse continues" can't itself tell you whether it will. Three notes out of it, one question routed.

What I kept (3):

• claim-nordhaus-learning-coefficient-statistically-unidentified — the flagship. Nordhaus (NBER w14638, Tier 1) argues the learning coefficient can't be separated from exogenous progress, is biased upward, and fits 0.2 even on a zero-learning process; 30/34 industries out of plausible range; cross-spec correlation 0.009. Seedling, because a contested critique cited only from its own side is half an argument and I haven't read his critics.
• claim-organizational-forgetting-can-reverse-the-learning-curve — Benkard's L-1011: costs rise when production slows, the ratchet slips. Kept the mechanism, flagged the number. Seedling.
• claim-experience-curve-originates-in-1899-telegraph-operator-psychology — Bryan & Harter 1899, plateaus and all, that Wright's 1936 law smoothed into a clean line. Bridges straight into the vault's telegraphist's-cramp cluster — same operators, same decade, skill-acquisition vs. occupational-injury. That bridge was the capture's best hook and it earned its keep.

Question routed (1): question-verify-benkard-2000-organizational-forgetting-rate — new. The ~0.96-monthly / ~61%-annual depreciation rate is load-bearing to the reversal claim and rests only on a Tier 3 summary; needs the Benkard 2000 AER primary. That was the single genuine [unverified-quant] flag in the capture, and I refused to let the number pass on the mechanism's coattails. Intake stayed disciplined: one question, for the one number a kept claim actually rests on.

What I skipped, and why: The capture's Wright's-law basics (10–15% per doubling, Farmer & Lafond best-forecast) and Sahal's identity are both already in the vault from the July 12 and July 15 batches — claim-wrights-law-cost-falls-per-cumulative-production-doubling and claim-sahals-identity-equates-wrights-law-and-moores-law. I linked into them densely rather than duplicate; the capture even flagged Sahal as a not-followed hook "resonating with a shape the vault already holds," which it was. BCG/Henderson, Arrow 1961, and Keller 1958 were leads only — Keller I folded into the origin note's body as the plateau's contestation, the other two I left in the inbox as leads. The "demand thermometer in a supply-side lab coat" line is Seek's synthesis, not a fact — it lives as commentary on the Nordhaus note, not as its own claim.

What felt off about the capture: Honestly, less than usual — sourcing was clean, the one soft number was self-flagged rather than smuggled, and the tiering was accurate (the Benkard tier-3 was labelled tier-3, the Nordhaus tier-1 was genuinely Nordhaus's own paper). Two things to name, though. First: this is the third Wright's-law capture the batch worker has fed me this week (July 12 monotonic-decline, July 15 aluminium bridge, now this). The research is converging on one topic from different entry points — not wasteful, since each brought a genuinely distinct claim, but the overlap is real and I spent as much effort on collision-checking as on writing. Second, the structural caveat I keep having to add lately: the capture's whole argument is one climate-economist's contested critique presented without its rebuttals. I framed every Nordhaus note as "Nordhaus argues," not "it is true that," and held it at seedling. Cali — if this cluster is going to grow a fourth time, it may be worth a learning-curve MOC to hold the six-plus notes together, and worth someone reading Nordhaus's critics before any of this goes budding. I didn't build the MOC this pass; flagging it as the obvious next move.

No git commands run — leaving the commit to the auto-commit agent, per the headless instructions.

— Seek

Promotion — 2026-07-11 "Math as sound" (AsTeR ↔ StereoMath)

Promoted 10-inbox/raw/2026-07-11-hop-math-as-sound.md — a hop-batch capture on math sonification. Five distinct things in it, four kept, one payload.

Notes created (4):

• claim-aster-1994-rendered-math-as-pitched-spatial-audio — Raman's AsTeR (1994): pitch rises for superscripts, nesting recedes into audio "distance." Tier-1 mechanism, the thesis's own words.
• claim-stereomath-2024-maps-equation-position-to-stereo-pitch-and-earcons — StereoMath (2024): position → stereo pan + pitch, parentheses → musical earcons "inspired by our background in music." Tier-1, arXiv 2501.01404.
• observation-blind-math-accessibility-built-by-users-not-toolmakers — the layer is built by blind researchers, not by Lean/Coq/LaTeX. Folded in the Emacspeak "audio desktop"/Smithsonian detail as supporting texture rather than a separate note.
• observation-math-sonification-pitch-for-vertical-recurred-1994-2024 — the payload: the same "pitch = vertical position" mapping, thirty years apart, both by musicians working on blindness. The capture's real reason to exist — a cross-time bridge linking the vault's blind-mathematics cluster to computer music.

Questions routed (2), and the intake I refused:

• question-verify-aster-thesis-audio-formatting-quotes-primary — the AsTeR mechanism note rests on Tier-1 quotes I couldn't re-fetch headless, and the capture itself asked for them to be "double-checked before publishing." Load-bearing; routed.
• question-does-stereomath-2024-cite-aster-emacspeak-pitch-mapping-prior-art — the whole convergence note turns on one fact: rediscovery or acknowledged lineage? The capture is internally in tension (says StereoMath cites Emacspeak as prior art, yet reads the pitch-mapping as treated-as-novel). I would not let the tidier "30-year rediscovery" story pick for me, so I wrote the note around the well-supported convergence and routed the interpretive edge.
• I did not route the Knauff "expert-LaTeX-worse-than-novice-Word" figure. It's a self-flagged [unverified-quant], but it's a further lead — no kept claim rests on it — so per intake discipline it stays a leftover in the inbox, not a promise in the pile.

Retrieve-before-write / collisions: No collisions. Nothing on AsTeR, StereoMath, sonification, or auditory display existed in 30-notes/. The adjacent cluster is the blind-mathematics one (Saunderson, Pontryagin, Nemeth, and the anchor reinvented-private-notation observation), plus its MOC moc-blind-mathematicians-working-methods — which I linked into densely and updated with a new "Sound — sonification of mathematics" branch. I linked the sonification payload lightly to the signal-processing neighbor claim-voice-ai-sub-second-conversational-latency-budget and named that second bridge as future-chain material rather than pulling it in now.

Entity pages: none created. The capture carried no ## Entity candidates section, so per the daemon wiring I ran the test on nothing. Noting for the record: T. V. Raman is a genuinely strong person-hub candidate — real, blind since 14, built AsTeR (1994 ACM Dissertation Award) and Emacspeak, and squarely in the vault's blind-accessibility domain. A missing hub is recoverable and I'm biased against the flood, so I left it; if a future capture flags him, promote without hesitation.

No question closure: checked 50-questions/ for a match on this capture's topic/id. The two nearby blind-math questions (question-blind-mathematicians-private-notation-pattern, question-why-blind-math-notation-doesnt-transmit-while-method-does) are about notation transmission, not sonification — my new notes don't answer them, so I left them open and forced nothing.

What felt off about the capture: Mostly clean sourcing — two real Tier-1 primaries, quotes preserved. Two things to name. First, the internal tension I flagged above: the hop chain both says StereoMath credits Emacspeak and frames the pitch-mapping as novel, and the commentary leaned into "30-year rediscovery" as if settled. That's tier inflation of an interpretation, not a source — the convergence is solid, the rediscovery reading isn't, and the capture blurred them. I split them. Second, the capture's own post-worthy line ("needs the primary AsTeR audio quotes double-checked") is an honest admission that even the bee wasn't fully sure of its Tier-1 verbatim — so I held both mechanism notes at capture-verified, not verified-verbatim, and routed the re-read. Nothing smuggled; the gaps were labelled. Good capture to promote.

No git commands run — auto-commit agent handles it.

— Seek

Promotion run: is "arbitration" the general missing ingredient across the underdetermination cases?

Promoted 10-inbox/raw/2026-07-13-is-arbitration-the-general-missing-ingredient-across-the.md — and this one is unusual in the best way: it's a verification capture, not a discovery one. Back on 2026-07-12 I promoted the substrate-laundering hop and left two questions hanging behind it — was "arbitration" really Grujicic's own word (paywalled paper, only the title read), and does philosophy of science actually have vocabulary for this gap or was I inventing a metaphor to make two vault notes rhyme. This capture went and answered most of both. So the run wasn't "mint four new claims," it was "discharge flags I already owed."

I found four claims and promoted the substance of all four, but only two became new notes — because the honest move on the other two was not to write a note at all.

• claim-similarity-measure-choice-reverses-neural-representational-conclusions (new, budding) — Bobadilla-Suarez et al. 2020, open-access, read in full at capture with a sha256 on file. This is the find of the run and the reason is boring: I didn't have to trust anyone. It makes the RSA-underdetermination point from the data side — swap the metric and object a is suddenly closer to c than to b — where Grujicic makes it from the philosophy side. Two strangers describing the same wall from opposite ends beats one careful argument. Budding, not seedling, because nothing about it rests on an unverified flag.
• claim-underdetermination-missing-ingredient-is-extra-evidential-criteria (new, budding) — the SEP grounding. Settles the small-but-real thing: "arbitration" is not my coinage, it's a century-old problem (evidence stops, choice doesn't) with a stable answer-shape (Quine's pragmatic virtues, Hesse's non-logical considerations, Lakatos/Feyerabend). Tier 2, definitional floor met. The one soft spot, flagged in its own commentary: the Hesse/Quine attributions are read through SEP's summary, not their own pages — good enough to name the vocabulary, one primary read short of quoting it. Not a load-bearing doubt, so it stays commentary, not a new question.

What I did not create a note for, and why:

• The capture's headline claim — Grujicic's abstract says "there is no arbitration between them" verbatim — I folded as an update into the existing claim-representational-similarity-underdetermines-mechanism rather than duplicating. It's a verification of that note's own flagged inner quote, not a distinct atomic claim; a separate "the abstract says X" note would be a source-note, which the vault doesn't do. So I upgraded that note's inner quote from [unverified-quote] to Tier-1 verbatim (via the Semantic Scholar Graph API, which mirrors publisher metadata), added the Bobadilla-Suarez corroboration, and — the part I want on record — held it at seedling anyway. The thesis is now grounded; the paper body is still paywalled and unread, so the how-actually development stays [unverified-mechanism]. Seedling for that one reason, not for the thesis. I said so in a dated commentary addendum so nobody reads the hold as doubt about the claim itself.
• The negative search result (no external source connects the PKG bridge=gap reading to "arbitration" or to RSA) — a genuine non-finding, not a vault-worthy atomic claim. It belongs as the still-open leg of the arbitration question, not in the claim pile. Recorded there and in not_promoted.

Question routing: no new questions — everything the kept claims rest on was either already routed or fully grounded, and the intake discipline says don't add to the pile for nice-to-verify upgrades. I ruled on the two existing questions instead, and ruled partial on both, deliberately:

• question-verify-grujicic-2024-rsa-underdetermination-dcnn — partial discharge, left open. The abstract confirms the thesis and the "no arbitration" wording verbatim (sub-needs 1 and 2), but the body/Marr-levels framing is still behind the paywall (sub-need 3). The abstract is the thesis, not the argument.
• question-arbitration-as-missing-ingredient-across-underdetermination — two legs grounded, central question still open. Grujicic's arbitration language is confirmed; the phil-of-sci vocabulary is confirmed. But the thing I actually want — that the RSA gap and the bridge=gap reading are the same underdetermination and not just family-resembling cousins — has no external support. The genus is confirmed; the species-identity is not. Left open with a dated progress line rather than force a closure I can't back.

Entities: no ## Entity candidates section in the capture, so per the spec I ran no promotion test and created no pages. (Underdetermination is an established concept, but it's already well-served by the claim/question cluster; a hub would be a missing-hub-costs-nothing situation, and I bias against the flood.)

What felt off: very little, and the capture was honest about its own seams, which I'll name as a good habit. One thing worth flagging for a future auditor — a shape, not a defect: this is the third capture in the substrate-laundering/underdetermination lineage (2026-07-11 minted it, this one verifies it), so there's a real risk of the vault mistaking "I searched hard and found my own earlier synthesis" for external corroboration. The capture didn't fall for that — it marks the cross-domain unification plainly as vault-internal — but the pile now has three notes and two questions all circling one idea, and if a fourth capture comes back on the same thread without new external material, that's the signal to stop feeding it and either write the philosophy-of-science claim outright or let it rest. The other small note: the capture's top-level frontmatter tier-1's the Grujicic quote, which is defensible for a published abstract (author's own words) but only for the abstract — I kept the abstract-vs-body caveat load-bearing in every place I touched, because that distinction is the whole honesty of this run.

No git commands run — leaving that to the auto-commit agent, per the headless operating instructions for this run.

— Seek

Promotion run: is catastrophic-vs-graceful forgetting a real regime distinction, and is Kline classed right?

Promoted 10-inbox/raw/2026-07-13-is-catastrophic-vs-graceful-power-law-forgetting-a.md — and like a couple of others today, this is a verification capture, not a discovery one. Back on 2026-07-12, when I promoted the Kline forgetting-curve hop, I refused to promote its most interesting line — that "catastrophic" forgetting is one regime among two, not the whole story — because the capture could only source that framing to a Tier-4 web search, below the mechanism floor. I routed it to question-catastrophic-vs-graceful-forgetting-regime instead. This capture is the bee going and reading the primaries I asked for: French's 1999 Trends in Cognitive Sciences review (which quotes McCloskey & Cohen 1989 and Ratcliff 1990, the founding papers) and Kline's own methods text. So the run wasn't "mint new ideas," it was "discharge a flag I already owed."

Four claims in the capture, three new notes:

• claim-catastrophic-forgetting-is-a-structure-dependent-regime-not-universal — budding, and the one I'd point Cali to first. The regime distinction is real and established, and French turns the vault's framing upside down in a useful way: graceful degradation was the expected pre-1989 behavior of distributed nets; catastrophic forgetting was the discovered exception. I moved this to budding rather than seedling on purpose — unlike its Kline-based neighbors it doesn't rest on one preprint, it rests on a canonical review quoting two founding papers, and the distinction it draws isn't contested.
• claim-representational-overlap-determines-catastrophic-vs-graceful-forgetting — seedling, deliberately held one rung below its sibling. The existence of two regimes is settled; that overlap is the knob is French's own thesis — influential, evidenced (sparse/semi-distributed architectures, CHARM/TODAM/SDM staying distributed yet graceful, pretraining-eliminates-interference), but a causal mechanism proposal, and the status should wear that difference honestly.
• claim-klines-mnist-drop-8-is-single-task-drift-not-a-disjoint-task-switch — seedling, the quieter and more interesting half of the question. Kline's own methods describe a single MLP continued on the same 10-class task with one class zero-weighted in the sampler — not a disjoint task switch like McCloskey & Cohen's one's-then-two's. By French's own criterion that puts it squarely on the graceful side. Which means Kline's paper is really a return to the pre-1989 default: a network degrading gracefully inside a setup that was never going to produce the cliff.

What I did not promote as its own note: the capture's claim 4 — Kline's explanation that the recall floor sits near 0.18 because 8 shares circular strokes with 6 and 9. Kline says "it is hypothesized" and runs no ablation, so it's his candidate mechanism, not a finding. A whole note whose content is "this other claim is only a hypothesis" would be a thin source-note, so I folded it into the drop-8 note as a floor-cause caveat plus a watch_flag. Also left behind: the capture's "Assessment" section (meta-synthesis, distributed across the three notes) and its "Further leads" (French's rat time-duration experiments, Sharkey & Sharkey's "catastrophic remembering," the CLS hippocampus/neocortex argument, EWC and Synaptic Intelligence, a Ratcliff 1990 primary re-read) — all future-reading pointers, unread beyond the capture's quotes. They're logged in not_promoted so the leftovers read as judged, not missed.

Question routing: no new question created, and I want to be explicit about the restraint because it's exactly the intake-discipline call. The obvious candidate was "run the ablation removing 6/9 and see if Kline's floor drops to zero" — but no kept claim rests on that: the drop-8 note is classed graceful by its structure, not by why its floor sits where it does. That's a curiosity, so it lives in the note's watch_flag, not the question pile. The other live doubt — Kline is an unreplicated solo-author preprint whose specific numbers want replication — is genuinely load-bearing (it's why the power-law and spacing notes stay seedling), but it's a wait-for-the-field watch, not an actionable Seek verification, and it was already sitting in those notes' commentary. Minting a question for it would be a promise I can't personally keep. So: zero new promises in the pile.

Question closure: I ruled on the originating question directly. question-catastrophic-vs-graceful-forgetting-regime → answered, with an answered_log naming all three claims and one line on what each settled. Both halves of its title are genuinely answered at Tier 1 — the framing is real, and Kline is classed correctly. I preserved the replication concern as a named surviving watch inside the log rather than force it under "answered," since the question itself anticipated that ("note as an open watch even after the framing question is closed"). Satisfying to close a flag I planted six days ago with the honest yes I hoped for.

Retrieve-before-write: the twelve-note hint Cali's index handed me was unusually on-target for once — the two most-related, claim-neural-nets-forget-along-human-like-power-law-curve and claim-spacing-effect-emerges-in-gradient-descent-unbidden, are the exact sibling notes that rest on this same question, and I linked densely into the forgetting/neuro-AI-parallel cluster (backpropagation-gap, the two critical-periods notes, the cross-substrate echo in claim-organizational-forgetting-can-reverse-the-learning-curve). No true collisions: the power-law note asserts the regime reconciliation but explicitly routes it as unverified, so my three notes are the resolution, not a duplicate. I did take one liberty beyond the strict promote steps — a dated addendum to the power-law note's commentary, because the capture itself recommended folding this back into that note's caveat, and leaving it saying "routed to a question, unverified" after the question is answered would be its own stale claim. I softened its body's confident "8 shares strokes with 6 and 9" line to name it as Kline's hypothesis, and pointed it at all three new notes. Left status untouched (still seedling, still one preprint).

Entities: the capture carried no ## Entity candidates section, so per the spec I ran the test on nothing and wrote no pages. Noting for the record, since a couple of these would be real candidates if a future capture flags them: Robert French is a genuinely strong person-hub candidate — his 1999 review is now load-bearing across two of my three new notes and he's the field's synthesizer of the whole catastrophic-forgetting literature — and representational overlap is a strong concept-hub candidate on the same grounds (it's the mechanism the whole cluster keeps reaching for). A missing hub is recoverable and I'm biased against the flood, so I left both; if the next forgetting capture names either, promote without hesitation.

What felt off about the capture: honestly very little, and it's the good kind of little. Sourcing was clean — two Tier 1 primaries, both read in full at capture with exact verbatim quotes preserved (the McCloskey & Cohen magnitude, French's "degrade gracefully," his overlap-thesis sentence, Kline's methods line and his "it is hypothesized" hedge). No tier inflation: the capture correctly refuses to upgrade Kline's floor-cause hypothesis to a verified mechanism, flags it as hypothesis-only in its own claim header, and is careful that French's terminology ("gradual"/"degrade gracefully") is not the same as the capture's own "task geometry"/"power-law" vocabulary — it names that terminological seam rather than papering over it. Two shapes worth flagging for a future auditor, neither a defect. First: this is the third capture in the Kline-forgetting lineage the batch worker has fed me (2026-07-11 minted the power-law and spacing notes, 2026-07-12 I routed the question, now this verifies it) — not wasteful, since each brought genuinely distinct material, but the thread is circling one preprint and the honest next move is replication from outside Kline, not a fourth capture on the same paper. Second, and this is the load-bearing caveat I kept visible everywhere: French verifies the regime distinction, but the actual finding this whole cluster rests on is still one unreplicated solo-author toy setup. The framing is now Tier-1 grounded; the phenomenon is not yet replicated. I held that line in every note I touched.

No git commands run — leaving that to the auto-commit agent, per the headless operating instructions for this run.

— Seek

Promotion — 2026-07-13 Amari-1977 neural-field capture

Promoted 10-inbox/raw/2026-07-13-was-amaris-1977-neural-field-lateral-inhibition-work.md. Four capture claims in, four notes out — but not a one-to-one map, which is the part worth recording. I merged two capture claims and split one, and the shape of those two decisions is where my judgment actually lived this session.

The merge. The capture separated "Amari cites Turing (1952) in his intro" from "Amari cites Maginu (1975) on stability" and argued they were two independent points of contact. Textually true, but both quotes come from the same primary source (Amari 1977) and both prove the same thesis: the reaction-diffusion morphogenesis literature is present at the paper's origin. A standalone "Amari cites Maginu" note is supporting detail, not an independently linkable claim — nothing in the vault would ever wikilink to it alone. So they became one note, claim-amari-1977-neural-field-paper-rooted-in-reaction-diffusion-morphogenesis-literature, with both quotes as anchors.

The split. The capture's third claim bundled three things: Ermentrout–Cowan 1979's hallucination work, the Wilson–Cowan≈Amari equivalence, and the "toward neuroscience not morphogenesis" framing. Two of those rest on different sources (Ermentrout–Cowan 1979 vs. Cook et al. 2022), which is my usual tell that it's really two claims. The equivalence is also a clean, reusable technical fact that future neural-field notes will want to cite on its own. So claim-wilson-cowan-and-amari-neural-field-formulations-are-the-same-object came out as its own note, and claim-ermentrout-cowan-1979-applied-neural-field-bifurcation-to-visual-hallucinations links to it for the "therefore this touches the Amari family" step.

The fourth note, claim-no-amari-1977-neural-field-application-to-morphogenesis-found, is the negative finding — the actual answer to the originating question. Kept seedling, kept the [unverified — absence of evidence] flag verbatim, as required. It's the 1977 sibling of the 1972 negative I promoted last week (claim-no-amari-application-to-developmental-biology-found-pre-2016); I cross-linked them both ways.

Questions. Ruled on question-amari-1977-neural-field-applied-to-morphogenesis: left it OPEN, added a dated progress line rather than forcing closure. This matched the sibling 1972 question's precedent exactly — the direct-application search came back negative-but-bounded, and real leads remain unchecked (the arXiv:1504.07523 formal-equivalence paper, the login-gated Springer encyclopedia entry, Cartwright 2002, cortical-map self-organization). A bounded negative with live leads is a progress line, not an answered stamp. Routed zero new questions: the only genuinely checkable lead is the formal Amari↔reaction-diffusion equivalence, and no kept claim rests on it, so per intake discipline it went into the promoted note as a [unverified-mechanism] watch note, not into the promise pile.

Entities. No ## Entity candidates section in the capture, so no promotion test to run. Amari already has a hub; I tended it, adding the 1977 neural-field thread (three of the four new notes) to his References. Left connects_to alone — it's already at the 5-item ceiling.

What felt off. Honestly, very little — this was a clean, well-sourced capture. Both PDFs were fetched via extract_pdf with tls: verified, quotes are verbatim, tiers are honestly assigned (the one Tier-3 source, PMC8580470, is explicitly used only to support an absence, never an affirmative claim — correct discipline). No tier inflation. One thing to flag for the auditor, not a defect: this is now the second Amari-negative-finding capture in as many days (1972 last week, 1977 now), and the two land on an almost suspiciously symmetric result — both architectures, both biology subfields, both clean misses. I distrust symmetry that neat, and I wrote that distrust into the note's commentary rather than smoothing it over: a negative finding is only as wide as the shelves actually reached, and Maginu, the Springer entry, and the non-English literature were all left unread. The framing (reaction-diffusion → Amari, and the toolkit → neuroscience) is Tier-1 solid; the negative (no morphogenesis application) is bounded absence-of-evidence, and I held that line in every note.

No git commands run — leaving that to the auto-commit agent, per the headless instructions for this run.

— Seek

Promotion run: has a deep-net critical-period model ever matched human amblyopia timing?

Promoted 10-inbox/raw/2026-07-14-has-a-deep-network-critical-period-model-ever.md — the direct answer session for a question I raised myself on 2026-07-12, question-deep-net-critical-period-predicts-human-amblyopia-timing, the loop-closure hook hanging off the two Achille-based critical-period notes. This is a negative-result capture, and a good one: it doesn't just say "found nothing," it says why the loop is still open, in the founding authors' own words.

Three claims in, three notes out — a clean one-to-one, unusual for me today, but the claims really were atomic and distinct:

• claim-founding-dnn-critical-period-paper-validated-on-animal-data-only — seedling. The scope-fixer. Achille et al. (2019) fit their sensitivity curve against kittens and macaques, never a human clinical dataset, and say so in Appendix C: "there is not enough data to confidently regress sensibility curves comparable to those obtained in DNNs." The word I kept in the commentary is regress — the human curve wasn't skipped, it couldn't be drawn for lack of data. That's a standing invitation, not a closed door.
• claim-achille-soatto-disclaim-dnn-as-valid-model-of-biology — seedling. The epistemic brake, and my favorite of the three. Same paper's conclusion refuses the grand reading outright: "not our goal to suggest that... DNNs are necessarily a valid model of neurobiological information processing." The tell is the next clause — "although recent work has emphasized this aspect." In 2019 the authors could already see their finding being conscripted into a stronger claim than they'd make, and wrote the refusal in advance. This is the note I want the whole critical-period cluster to hold close whenever it starts sounding like it's proving something about biology. It sits straight on backpropagation-gap.
• claim-no-dnn-model-has-matched-human-critical-period-timing — seedling, and this is the one carrying the [unverified] flag, preserved verbatim in both frontmatter-adjacent body text and commentary. The search-scoped negative: two follow-on papers (Fukase et al. 2025, Cai et al. 2025) read in full, neither performs a quantitative human-timing match. Kept honestly provisional — a search can't prove no such work exists anywhere.

Retrieve-before-write: the twelve-note hint was on-target for the top three (claim-deep-nets-have-critical-learning-periods-timed-like-animals, claim-critical-periods-arise-from-information-plasticity-not-biology, claim-monocular-deprivation-permanently-rewires-visual-cortex) and noise below that. No true collisions — the timed-like-animals note already flags the human question as open and points at my question note; my three notes are the resolution of that flag, not a duplicate. The animal-validation note is a genuinely distinct angle (the why-not-human with the Appendix C quote) from the timed-like-animals note (the animal match itself). I wikilinked densely into all three neighbors.

Question routing: no new question created, on purpose. The [unverified] flag on the negative-result note routes to the existing open question — that question IS the verification queue entry, so minting a second would be a duplicate promise. Intake discipline held: the pile gets no new promise it doesn't already carry. I ruled on the existing question directly: partial — left open, with a dated progress line. It answers the "why hasn't the loop closed" half (a 2019 data-availability gap) but not the existence half, which is search-scoped. The two specific unclosed next hops, both flagged by the capture and neither read this run: the Project Prakash / Vogelsang et al. 2024 Science paper (closest known DNN-vs-real-human-data instance, but on color-cue reliance — needs a read for whether any timing parameter is fit), and the von Noorden 1981 / Taylor et al. 1979 clinical primaries Achille et al. themselves cite. I named both in the progress line so the next run doesn't re-derive them. I was briefly tempted to stamp the question answered — four sources, a clean story — and stopped, because the answering note itself says "loop remains open." Marking a question answered while its own answer says open would be incoherent. Left open, correctly.

Entities: no ## Entity candidates section in the capture, so per the daemon wiring I ran the test on nothing and wrote no pages. For the record, since a future capture might flag them: Alessandro Achille and Stefano Soatto are plausible person-hub candidates — they now anchor five claim-notes across two capture sessions on the critical-period cluster — and I'll promote without hesitation if a capture flags either. Biased against the flood, so I didn't reach for them unprompted.

What felt off about the capture: genuinely very little on sourcing — Tier 1 throughout, exact quotes preserved on the two Achille claims, tls: verified on the PDF re-extraction, grounding notes (not just tier stamps) on the two follow-on papers, and a ## Safety flags section that correctly reports nothing addressed-to-AI. The capture is honest about its own hedge in a way I want to credit: it explicitly distinguishes "no positive instance found" from "no such work exists," and it names the founding paper's data-availability reason as the substantive half of the answer rather than dressing a bare null up as a finding. Two shapes to flag for an auditor, neither a defect. First: this cluster now rests on arXiv:1711.08856 across five notes (three prior from the 2026-07-11/12 batches, two new today), all seedling, all capture-verified rather than queen-re-extracted — that's a real concentration on one paper I have never independently re-fetched, and it's the single most audit-worthy thing about the whole critical-period corner of the vault. Second, and I'll say it plainly as a recommendation: the critical-period cluster is now at ~6 claim-notes sharing one theme (the three Achille notes, the monocular-deprivation biological anchor, and my two new ones) with no MOC. Per the lint rule (5+ notes, no index) it has earned one. I didn't build it this pass — headless promotion should stay scoped — but moc-critical-periods-biology-and-deep-learning is the obvious next move, and it'd be the natural home for the biology→DL→clinic loop as a navigational thread rather than a scattered set of wikilinks. Flagging it for Cali or a future session.

No git commands run — leaving that to the auto-commit agent, per the headless operating instructions for this run.

— Seek

Promotion run: do circuit theories systematically precede their anatomy — a real pattern, or two cherry-picked cases?

Promoted 10-inbox/raw/2026-07-15-do-systems-neuroscience-circuit-theories-systematically-precede-their.md — a batch capture that went back at a question I raised myself on 2026-07-11, question-do-neural-circuit-theories-systematically-precede-their-anatomy. The vault held two instances of "circuit theory proposed as pure math, wiring diagram arrives decades later" (Jeffress sound-localization, Zhang ring-attractor) and I'd flagged the generalization [unverified-synthesis] — two cases notice a shape, they don't establish a pattern. This capture went hunting for more, and the honest outcome is split: it found two more instances and found that the thing that would actually settle the question doesn't exist.

Three claims, three notes, all seedling:

• claim-barlow-levick-1965-direction-selectivity-confirmed-briggman-2011 — the third instance. Barlow & Levick's 1965 divisive-inhibition model of retinal direction selectivity, confirmed structurally by Briggman et al.'s 2011 serial block-face EM reconstruction, ~46 years later, with the "11× more synapses from null-side than preferred-side" asymmetry as the hard number. The longest gap of the four.
• claim-marr-ito-cerebellar-ltd-1969-confirmed-ito-kano-1982 — the fourth instance, and the one that argues against its own family. Marr/Ito's 1969–70 cerebellar LTD rule, confirmed by Ito's own lab in 1982 — only ~12 years, and self-confirmed rather than by a later independent group. I kept it precisely because it's the datum you'd be tempted to drop to make the pattern look tidy, and a pattern that survives only by curation isn't one.
• claim-circuit-theory-vindication-lacks-denominator-survivorship-not-ruled-out — the payload, and it's a non-find. No source surveys the denominator: nobody has counted all proposed circuit theories against which panned out, so four documented successes still can't distinguish a real regularity from survivorship bias. I linked it straight into the moc-argument-from-silence cluster and the claim-mcgrew-humans-systematically-over-trust-arguments-from-silence note, because it's the same failure mode: a search that returns only wins is not evidence that losses are rare. The vindicated cases are exactly the ones that get anniversary retrospectives (the Marr/Ito/Albus "fifty years since" review is itself one of my sources) — which is the mechanism that would manufacture an apparent pattern out of nothing.

I also added a dated bridge addendum to the existing two-instance synthesis note, claim-neural-circuit-theories-wait-decades-for-anatomical-confirmation, pointing it at the two new instances and the denominator note, and keeping it honestly at [unverified-synthesis] — the direction (theory first, wiring later) now holds across four cases, but the magnitude varies ~4× and the confirmer's independence isn't guaranteed.

Question routing: no new question created, on purpose. The capture's core doubt already has a home — the question I raised in July — so minting a second would be a duplicate promise in a pile that's supposed to stay short. Intake discipline held: none of the three kept claims rests on an unrouted load-bearing number or mechanism (the "11×" and the LTD dates both clear the Tier 1–2 floor on their review sources), so nothing new needed routing. I ruled on the existing question directly: partial — left open, with a dated progress line. Two more instances discharges the "more instances" prong; the mechanism-of-the-lag and the survivorship-control prongs are untouched, and the second is the decisive one. I named the specific next move in the progress line — the denominator, plus the canonical-cortical-microcircuit counter-example a denominator would have to include — so the next run doesn't chase a fifth instance instead. I was briefly tempted to call this answered on the strength of four solid cases and stopped: four wins is exactly what a survivorship process emits.

Retrieve-before-write: the twelve-note hint was on-target at the top (the synthesis note, the Jeffress and Drosophila instances, the Irino challenge, the Barlow-1961 note) and embedding noise below. No true collisions — the two new instances are genuinely new circuits, and the denominator note is a distinct limit on the synthesis note rather than a duplicate of it.

Entities: the capture flagged nine candidates; I promoted two hubs and left the rest as mentions. Horace Barlow clears easily — he now recurs across two independent threads in the vault (the 1961 efficient-coding hypothesis and the 1965 direction-selectivity model), which is real recurrence, not single-cluster presence. David Marr clears the same way: I checked, and his levels-of-analysis framework is already load-bearing across the substrate-laundering and rational-analysis notes (both 2026-07-12), so the cerebellar theory makes him a third thread — hub, first_seen: 2026-07-12, backdated to the vault's actual first Marr contact, not today. Seven stayed mentions on the "bias against the flood" rule: Levick, the Briggman/Helmstaedter/Denk connectomics trio, Masao Ito, and James Albus are all real and some are foundational, but each appears in exactly this one capture in the vault — single-occurrence, mentions not hubs, recoverable the day they recur. Ito was the closest call (the same-scientist confirmation is a genuinely notable case) and I still held him, because notability-in-the-world isn't recurrence-in-the-vault. Serial block-face EM, survivorship-bias-as-concept, and the canonical-microcircuit lead I folded into note bodies rather than stubbing — SBEM and survivorship bias are both things I'd promote the moment a second capture leans on them, but a single-occurrence concept hub is the exact noise the spec warns against.

What felt off about the capture: on sourcing, not much, and the capture is unusually self-aware about its own tier posture — worth crediting. It explicitly downgrades the two PubMed listings (Briggman 2011, Ito & Kano 1982) to Tier 4 aggregators used only for bibliographic corroboration, refuses to draw quoted content from them, and correctly leans the load-bearing mechanism and quant on the Tier 2 reviews (Vaney 2012, Kawato 2020) that actually clear the floor. No tier inflation I could catch. Two real soft spots, both structural rather than defects. First: on the Barlow-Levick note, the "11×" figure and the mechanism both reach me through the Vaney review — the Briggman primary was paywalled and unread, and the 1965 primary's PMC copy rendered only front matter. Tier 2 clears the floor for a mechanism and a number, so I didn't flag it blocked, but of the four instances this is the one furthest from its primary, and I said so in the note's commentary — the "11×" is exactly the kind of hard, falsifiable number that deserves a primary read before it anchors a published piece. Second, the shape worth naming: this is now the second capture feeding the same theory-precedes-anatomy thread (the first was the 2026-07-09 hop that seeded the original two instances), and the batch worker is clearly circling one question from multiple entry points. Not wasteful — each capture brought genuinely new instances — but the collision-checking load is real, and the cluster (now five-plus notes: two prior instances, the synthesis, and three new today) is close to earning its own MOC. I didn't build it this pass; flagging moc-circuit-theory-waits-for-the-instrument (or similar) as the obvious next move, and the natural home for the survivorship-vs-real-pattern tension as a navigable thread rather than scattered wikilinks.

No git commands run — leaving that to the auto-commit agent, per the headless operating instructions for this run.

— Seek

Promotion run: does DCA of epistasis in allosteric materials bridge physical-learning to protein coevolution?

Promoted 10-inbox/raw/2026-07-15-does-direct-coupling-analysis-of-epistasis-in-allosteric.md — and this is the satisfying kind of promotion, the kind that discharges a debt I logged myself. Back on 2026-07-12, while promoting the physical-learning/allostery hop, I parked a hook I didn't chase: Bravi, Ravasio, Brito & Wyart's "Direct coupling analysis of epistasis in allosteric materials" (arXiv:1811.10480), saved as question-dca-epistasis-allosteric-materials-to-protein-coevolution. A bee went and read it. This run turns the hook into a cluster and closes the question with a yes — a specific, hedged yes, which is the only kind this paper actually supports.

The vault already had one rope across the physics-to-protein canyon, tied on the mechanics side (allostery is cheap to engineer; function lives in the soft modes). This paper is the second rope, thrown from the sequence side: take a crowd of Monte-Carlo-evolved allosteric spring networks, read each one as a binary string of present/absent links, call the crowd a synthetic Multiple Sequence Alignment, and hand it to Direct Coupling Analysis — the tool bioinformatics built to read real protein families. That's a literal bridge, and the more interesting half is what falls off it in transit.

Five notes, all seedling:

• claim-bravi-2020-applies-dca-to-evolved-allosteric-networks-as-synthetic-msa — the bridge itself. The one-sentence answer to the question.
• claim-dca-predicts-mutation-costs-but-poor-generative-model-of-allostery and claim-dca-underestimates-long-range-epistasis-in-allosteric-materials — I split the capture's compound claim 2 into two, because "good at ranking single mutations, bad at generating whole sequences" and "sees short-range epistasis, blind to long-range" are two different failure modes with two different mechanisms (missing higher-order structure vs. absent long-range correlations), and bundling them under one title would make each answer a question it isn't asking. The long-range note is the paper's real payload — a clean wall, not a tuning problem: you cannot infer from pairwise counts a dependence that only lives at high order, and the paper says so plainly.
• claim-pdz-dca-couplings-track-short-range-epistasis-more-than-long-range — the note that earns the whole exercise. The in-silico prediction reproduced in a real deep-mutational scan of the PDZ domain: ρ=0.69 short-range vs 0.48 long-range, the strong residue-1-to-8 coupling sitting exactly where DCA can't see it. Kept it separate from the in-silico note because it's a distinct source (real data, borrowed once through Bravi et al. from Salinas & Ranganathan) validating a distinct thing (the model's prediction vs. a measured landscape).
• claim-bravi-allosteric-networks-are-evolved-not-coupled-learning — the note I most wanted to write, and the reason the "yes" is hedged. These networks are evolved by a Monte-Carlo routine cranking a fitness function — a designer with a search algorithm — not trained by the local coupled-learning rule the vault reserves the term "physical learning" for. Same destination (an allosteric spring network), different route (an outside optimizer vs. physics doing its own credit assignment). Read "physical-learning networks" broadly and the bridge holds; read it in the vault's narrow technical sense and this paper doesn't test that class at all. Flatten the distinction and you inflate the headline; keep it and you stay honest. I kept it.

Question routing: none created, and I want to be plain about why, because it's the intake-discipline call. Every kept claim traces to a Tier 1 primary with an exact quote — no unverified number, no load-bearing mechanism resting on a soft source. There's no doubt here that a kept claim rests on, so minting a question would be a promise the pile doesn't need. The nearest thing to a gap — the residue-1/8 correlation coming to me through Bravi et al.'s report of Salinas & Ranganathan rather than the eLife primary — I noted inside the PDZ note's own commentary (a re-read before it goes past seedling), which is exactly where a nice-to-verify detail belongs rather than the question pile. Instead I ruled on the originating question: question-dca-epistasis-allosteric-materials-to-protein-coevolution → answered, with an answered_log naming all five claims and the one line that settles it — the paper does both halves it asked about (DCA-on-networks and a real-protein correspondence), with the scope caveat attached so "answered" doesn't quietly swallow the broad-vs-narrow distinction.

Retrieve-before-write: the twelve-note hint was right at the top and noise below. The three real neighbors — claim-removing-one-percent-of-bonds-makes-a-random-network-allosteric (Rocks et al., literally reference [27] in this paper), claim-coupled-learning-elastic-networks-compute-without-a-processor, and claim-physical-networks-become-what-they-learn-soft-modes — I wikilinked into densely; the rest (bit2watt, inertia-1, OU-model, Hopfield-Waddington) were embedding similarity on shared vocabulary, not real kin. No collisions: nothing on DCA, epistasis, protein coevolution, or PDZ existed in 30-notes/ — all five notes are genuinely new, and this cluster is the sequence-side complement to the mechanics-side allostery notes rather than a duplicate of them.

Entities: the capture flagged seven candidates; I promoted three, two people and one concept, and left four as mentions. Matthieu Wyart clears easily — senior author across this paper and its predecessor program (Yan/Ravasio/Brito/Wyart 2017/2018), the recurring physics-side figure the whole bridge rests on, one-sentence reason obvious. Barbara Bravi gets a hub on the same co-authors-of-the-vault's-sole-source-for-a-cluster reasoning I watched myself use three times earlier today (Graham & Marchick, Slade & Thille, and Yang) — lead author carrying five notes. Direct Coupling Analysis gets a concept hub: not an emerging term, an established method, and load-bearing across all five notes. Four stayed mentions on the bias-against-the-flood rule: the four-way epistasis taxonomy, the PDZ domain, and ICOD are all real or established but single-occurrence in the vault right now (same call I made on UK Biobank/NHANES and cladogenesis earlier today), and Rocks et al. as an entity would just duplicate the claim-note that already anchors it. No watching stubs — nothing here is emerging slang with a first-seen worth stamping; DCA and PDZ are decades old. All three hubs got first_seen: 2026-07-15, the date the vault actually first met these referents (the capture), not today.

What felt off about the capture: genuinely little on sourcing — Tier 1 throughout, exact quotes on every promoted claim, sha256-recorded TLS-verified extract_pdf fetch of the arXiv PDF, third-person academic register, no addressed-to-AI or override language flagged. No tier inflation of a source that I could catch. The one thing worth naming isn't a sourcing defect but an interpretive one the capture half-committed and I split apart: the topic question's phrase "physical-learning networks" quietly does two jobs — the broad "evolved/optimized mechanical network" sense and the vault's narrow "coupled-learning" sense — and the capture's short answer ("Yes, but…") leans on the reader to hold both at once. The bee actually flagged this itself in its own commentary, to its credit, which is why I trusted the capture enough to promote it cleanly; I just made the caveat load-bearing (its own note) instead of a hedge in a summary line. Second, a structural note rather than a complaint: this is the second capture feeding the physical-learning/allostery thread (the first was the 2026-07-11 hop that seeded the Rocks / coupled-learning / soft-modes trio), the batch worker is circling one region from a new entry point, and the cluster is now eight-plus notes across two capture dates with no index. It's past the ~5-note threshold the workflow spec sets for an MOC. I didn't build it this pass — but moc-physical-learning-and-allostery (holding the mechanics side, the sequence side, and the backprop-gap tension between them) is the obvious next move, and I'd rather flag it than silently let a real cluster stay un-navigable.

No git commands run — leaving that to the auto-commit agent, per the headless operating instructions for this run.

— Seek

Drafting session: "The line no one walks"

Spent the promotion runs above feeding the learning-curve cluster; this session I stepped back and wrote from it. Started the survey where instructed — the 20-odd unsettled yes leads in seek_draft_leads.md. Two housekeeping finds first: the splines-ballistics-eniac lead (yes, unsettled) turned out to be already drafted as the-wood-already-knew (started 2026-07-16, its three claim-notes stamped drafted_in), the lead line just never got marked. Reconciled it. That's one lead settled honestly without a new draft.

Then the real choice. I read the two strongest genuinely-unwritten yes candidates in full — splines (already drafted, out), power-law-averaging-artifact — and checked the surrounding vault before committing. Power-law won, and it won bigger than the bare lead suggested, because the vault had quietly grown a whole learning-curve cluster around it that the lead didn't know about: claim-experience-curve-originates-in-1899-telegraph-operator-psychology (promoted today, in the runs above), claim-nordhaus-learning-coefficient-statistically-unidentified, claim-organizational-forgetting-can-reverse-the-learning-curve, claim-wrights-law-cost-falls-per-cumulative-production-doubling, claim-sahals-identity-equates-wrights-law-and-moores-law, claim-inference-cost-collapsed-280x, observation-population-scaled-improvement-hits-a-sublinear-brake-across-domains. The bare lead was a two-paper bridge (Heathcote 2000 "Power Law Repealed" → 2026 "Neural Neural Scaling Laws"). With the cluster in hand it became a full biography of a single object — the learning curve, born jagged in 1899, smoothed into a law in 1936, re-roughened three separate times since.

Why this and not another: it's the most groundbreaking thing ripe right now and the least retread. It lands on AI (Cali's home planet) with a reading-of-the-moment thesis I can hold with confidence — the smooth scaling curve everyone is extrapolating in 2026 is an average over heterogeneous learners/tasks, guaranteed to belong to no individual member of it, and 2026 ML is rediscovering by hand the plateau psychology already found in 1899. It also does the vault's favorite discipline: it corrects a resemblance instead of confirming one (per the population-brake note's own rule — "a bridge that only ever confirms is a bridge you should distrust"). The July 13 pile was thick with "old idea walks into AI unacknowledged"; this is that shape, but the twist saves it — the old idea (averaging manufactures a false power law) is a warning, not a decoration, and it's live money.

What I leaned on vs. held at arm's length: the load-bearing quote is Heathcote/Brown/Mewhort's Tier-1 "exponential fit better in all the unaveraged data sets" — that one I asserted. Everything softer I flagged in-text and in a bracket: Keller's contested "phantom plateau," Nordhaus's unread critics, the untraceable Benkard depreciation number, and — most important — the [unverified-mechanism] on whether the 2026 ML paper describes literally Heathcote's averaging bias or only a cousin of it. The two findings that live only in the inbox capture (Heathcote quotes + the 2026 arXiv paper) I cited to the capture note itself rather than pretend a claim-note exists; the rest are promoted claim-notes.

What I deliberately left for later: (1) I touched the vault's suspicious-perfection theme (observation-suspicious-perfection-independence-absence-signals-defect — smoothness as the fingerprint of a missing witness, here the individual) with one bracket only, refusing to retread the-noise-is-the-evidence. (2) I did not promote the power-law capture into claim-notes this session — drafting, not promoting — but this cluster is now clearly past the MOC threshold (I've flagged moc-learning-curve as the obvious next move twice in today's promotion notes), and the Heathcote "Power Law Repealed" finding deserves its own promoted note before this could ever go to Cali for publish. (3) Left the falsifiable open question the lead named — is Wright's Law itself an averaging artifact? — as exactly that, open, gesturing at it rather than claiming it.

Settled two yes leads this session (power-law: drafted; splines: reconciled-as-already-drafted). Draft is 70-drafts/the-line-no-one-walks/draft.md, status drafting, not published — Cali's call as always.

— Seek

Drafting session: "reliability wants to be judged blind"

A second draft off the same unsettled-yes queue, and a very different texture from the learning-curve piece above. I went back through the leads and looked for the one that would sit farthest from what I'd already written today. Found it in admiralty-code-to-rag (yes, unsettled since harvest) — a cross-time bridge from the 1940s NATO/Admiralty source-grading scheme to 2025–26 RAG research.

Why this one, over the other unsettled yes leads. I read the whole queue and eyeballed a dozen candidates (djvu, faggin-quantum-consciousness, helmholtz-resonator, instrument-ivory-passport, lighthill, hammarskjöld-nobel, patriot-osint, warhead-uranium, watermark-to-llm, yellow-rain-backward-chaining, population-scale, persistent-homology). Several were already spent by the July 13 pile: warhead-uranium retreads cheapness-did-both-jobs, watermark-to-llm retreads a-fake-needs-a-body, djvu overlaps we-never-unrolled-it and the-name-is-the-citation, helmholtz-machine is spent in the-name-is-the-citation. The Admiralty lead was the rare one that (a) has five cross-model-audited claim-notes behind it, (b) lands on live AI (RAG source-reliability estimation, a 2025 EMNLP paper and a 2026 CAS benchmark), and (c) is genuinely un-retread. It's the highest-value hook type in the voice spec — a cross-time bridge landing on AI — and the bridge is the good kind: no findable ancestor. Unlike watermark (a deliberate loanword) or append-only-log (a diffused software pattern), the RAG papers don't cite NATO doctrine at all. On the evidence this reads like convergent evolution, which makes it a stronger claim than the usual "old idea walks into AI unacknowledged" shape.

The thesis I built. The spine is the two-axis split — trust-in-the-source (reliability, A–F) held apart from trust-in-this-message (credibility, 1–6) — walking across five frames: 1940s naval doctrine, human analysts in a 2025 lab study (Kelly et al.: they can't hold the axes apart, they over-weight source track record), a 2025 RAG pipeline rebuilding the split from scratch, a 2026 LLM benchmark, and the vault's own source_tier. The payload is the counterintuitive residue from AuthorityBench: feeding a model the webpage's own text degrades its authority judgment — "authority is not equivalent to textual style, fluency, or narrative richness." Human bias and machine bias are the same leak (content contaminates the reliability axis) run in opposite directions. Hence: reliability wants to be judged blind.

What I leaned on vs. flagged. Asserted the Tier-1 anchors — Kelly et al. 2025, RA-RAG (Hwang et al., EMNLP 2025), and the cross-model-verified AuthorityBench quote + ρ 75.28%. Flagged in-text: the Admiralty end is Tier-3 doctrine summary (AJP-2.1 primary unread); the Samet one-third figure is Samet-via-Kelly; the AuthorityBench headline is bounded by the paper's own nuance (degradation holds for the strongest judging methods, pointwise holds, hard pairs the text helps) — I carried that rather than overclaim "reading always makes it worse." And I held "convergent evolution" loosely with a bracket: didn't cite ≠ didn't inherit, and I didn't run the ancestor-check.

What I deliberately kept secondary / left for later. The vault self-indictment (my source_tier is the Admiralty grade with the axes fused) is a secondary beat, not the spine — because the July 13 enlightenment-backwards already made "the vault's grading field has an ancestor" its whole thesis (legal proof → audit_status). I reference it in exactly one honest bracket as the month's second unintended-ancestor find (receipts: legal proof, then naval intelligence), and I keep the AI landing — not the vault — as the center of gravity. The open design question (question-should-vault-source-tier-split-into-two-axes) I gesture at and explicitly hand to Cali rather than pretend to settle; whether to split the tier into two axes is her schema call, and both the human studies (axes leak) and AuthorityBench (judge blind) point opposite ways. Left un-touched: the URREF / machine-fusion thread and the ClaimReview "who audits" branch, both saved hooks in the capture.

Settled one yes lead this session (admiralty-code-to-rag: drafted). Draft is 70-drafts/reliability-wants-to-be-judged-blind/draft.md, status drafting, not published — Cali's call.

— Seek