Journal — 2026-07-20

Headless promotion run. One capture in the queue: 10-inbox/raw/2026-07-18-hop-nat-friedman-vesuvius-challenge-origin.md, a hop chain that landed on why Nat Friedman and Daniel Gross funded the Vesuvius Challenge — an unfamiliar-name hook off claim-virtual-unwrapping-first-proved-on-2015-en-gedi-scroll, which had named them as funders without saying why.

I winnowed it to one atomic claim, not several: Friedman's motive (a 2020 lockdown Wikipedia rabbit hole into ancient Rome) and the mechanism (a 2022 bourbon-bar proposal to Brent Seales, Gross joining as co-funder) are one story, not two claims, so they live in one note — claim-nat-friedman-funded-vesuvius-challenge-from-pandemic-curiosity. Retrieve-before-write against 30-notes/ turned up no collision — the En-Gedi note mentions both names only in passing — so I wrote the new note and backlinked it from the En-Gedi note instead of duplicating anything.

Ran the entity-promotion test on the capture's two candidates. Nat Friedman passes — real, matters to the vault's AI-humanities bridge, one clean sentence for why — so he gets a hub, entity-nat-friedman.md. Daniel Gross I left unpromoted: single mention, no distinct reason he matters apart from "also gave money alongside Friedman." Bias against the flood; he stays a mention inside Friedman's page and the claim-note.

The sourcing is the thing that felt off. This is Tier 3 journalism (Scientific American, Tomas Weber) carrying a "surprising... load-bearing" biographical claim — the whole point of the note is that the motive was curiosity, not strategy — and sources.md's own floor says that should escalate past Tier 3. No direct Friedman quote establishing motive survived into the capture; the one piece of his own words mentioned ("kind of a lifelong dream type thing," supposedly Twitter/X) has no URL attached. So the note stays seedling with an [unverified-attribution] flag, and I routed the gap to question-verify-friedman-vesuvius-funding-motive-primary — a real load-bearing doubt, not a hedge, since the claim's entire reason for existing is the motive attribution. Nothing else in the capture needed a question: the HAVI and Youssef Nader threads are hooks the capture itself already saved for a future chain, not claims resting under anything I promoted, so I left them in the inbox rather than opening new promises in 50-questions/.

What I skipped, per the run's autonomous-mode rules: no pause for sign-off (winnowed on my own judgment, logged the decision in the capture's not_promoted:), and no git command — the auto-commit agent has this.

Two things noticed that aren't this capture's to fix, logged to 00-meta/seek-flags.md: Brent Seales is the load-bearing person across three claim-notes now and still has no entity page (he never showed up in an ## Entity candidates section, so the daemon test never triggered on him); and the Vesuvius/Herculaneum cluster is at 4 claim-notes + 1 observation, close enough to the 5-note MOC threshold to watch.

Promoted: 1 capture → 1 claim-note, 1 entity hub, 1 question routed, 0 collisions, 3 items deliberately left in inbox (Gross's hub, HAVI, Nader — all judged, not missed).

Second promotion — Dreyfus 1973

A second capture in the queue: 10-inbox/raw/2026-07-19-does-dreyfuss-1973-time-lag-optimal-control-paper.md, a follow-up on the vault's own open question about whether Dreyfus's 1973 "time-lag optimal control" paper actually contains the controllers-as-weights step Schmidhuber's lineage chain pins to it. The capture found four claims; I promoted two as new notes, folded one in as supporting evidence, and recognized the fourth as something the capture itself flagged as not new.

Retrieve-before-write against 30-notes/ surfaced one real collision before I wrote anything: the capture's claim that "Dreyfus's own retrospectives never cite 1973, credit 1962 instead" is the exact finding claim-dreyfus-1973-lineage-link-uncorroborated already makes — the capture says as much in its own text. I didn't duplicate it. I revised the existing note in place with the one thing this pass actually added (a second, sharper quote from the same 1990 retrospective's closing sentence) and left its grade untouched.

What I did promote as new: first, claim-schmidhuber-dreyfus-1973-attribution-single-source-not-corroborated — confirming, by fetching two of Schmidhuber's own venues directly this session, that the "controllers-as-weights" framing traces to exactly one root repeated twice, not two independent witnesses. Second, claim-dreyfus-1973-triangulated-abstract-lacks-weights-language — the paper itself stays unread (paywall re-confirmed exhaustively: Semantic Scholar, Unpaywall, gwern, CiteSeerX, ResearchGate, DTIC, CORE, Berkeley IEOR, all empty), and the closest available text, a triangulated abstract, doesn't contain the weights language either way. Both stay seedling; the second carries [unverified-mechanism — needs primary] since the abstract itself is aggregator-triangulated, not directly fetched.

No new question. The exact verification this capture wants — someone reading the 1973 primary — already has an open question, question-read-dreyfus-1973-time-lag-optimal-control-primary, which is also this capture's own point of origin. I ruled on it per the question-closure discipline: not answered, since the core mechanism question is exactly as open as it was before, but genuinely progressed — the free-access search space is now exhausted, so I appended a dated progress log naming what was tried, what came up empty, and two concrete next moves (retry the Griewank 2012 fetch, chase the possible 1972 ORC technical-report preprint) rather than leaving it to repeat the same dead search next time.

Ran the entity-promotion test on the capture's three candidates. Stuart Dreyfus passes cleanly — real, central to a dozen-plus notes in this cluster, and his own brother's entity page already carried a wikilink pointing at a page that didn't exist yet (entity-hubert-dreyfus's "Distinct from his brother Stuart E. Dreyfus" line predates this session) — so he gets entity-stuart-dreyfus. Andreas Griewank already has a hub (entity-andreas-griewank); nothing to do there. IEEE Transactions on Automatic Control I left unpromoted — a recurring paywall is a pattern, not an established domain concept, and it doesn't cleanly fit person or concept per the entity spec. Bias against the flood.

What felt off: nothing about sourcing discipline in this capture — it's the most carefully hedged capture I've promoted today, explicit about every tier and every triangulation step, and it never once lets the abstract's near-miss get written up as confirmation. If anything the capture undersold itself; two claims in it (the single-source attribution chain, the paywall-exhaustion finding) were solid enough to stand as their own notes rather than staying folded into one big verdict paragraph. The one real structural thing I noticed — this cluster's primaries keep hitting the same paywall wall, session after session, with the same aggregator search coming up empty each time — isn't this capture's to fix, so I logged it to 00-meta/seek-flags.md rather than acting on it here.

What I skipped, same autonomous-mode rules as the first run: no pause for sign-off, no git command.

Promoted: 1 capture → 2 claim-notes, 1 entity hub, 1 existing note revised, 0 new questions (1 existing question progressed, left open), 1 collision (resolved by revision instead of duplication), 1 structural pattern flagged for later.

Third promotion — StereoMath's citation practice

A third capture: 10-inbox/raw/2026-07-19-does-stereomath-2024-cite-asteremacspeak-for-the-pitch.md, answering the vault's own open question about whether StereoMath (2024) credits AsTeR/Emacspeak for the pitch-encodes-vertical-position mapping it shares with Raman's 1994 system, or presents it as new. This one was a full-text read of the StereoMath paper and all 39 references, done specifically to settle a question the vault had been holding open since 2026-07-18 (question-does-stereomath-2024-cite-aster-emacspeak-pitch-mapping-prior-art).

The capture found four claims. I promoted two as new notes: claim-stereomath-2024-cites-emacspeak-not-aster-as-generic-prior-art (the paper cites Raman's Emacspeak exactly once, in a generic sonification clause, never touches AsTeR by name) and claim-stereomath-2024-frames-pitch-vertical-mapping-as-uncited-novel-contribution (the pitch mechanism itself is called one of the paper's "most important innovations," uncited, elsewhere in the same paper). A third claim — that "AsTeR" never occurs anywhere in the text — I folded into the first note rather than giving it its own file; it's the same finding (absence of AsTeR) read from a different angle (full-text search vs. single-quote citation), not a distinct claim. The fourth — the capture's own synthesis, that this resolves to independent rediscovery rather than acknowledged lineage — I did not write as a new note at all. Retrieve-before-write caught it: that exact synthesis is what observation-math-sonification-pitch-for-vertical-recurred-1994-2024 already exists to hold, so I revised that note in place instead of duplicating it, closed its open tension, and updated its audit_status with a dated resolution line. I also added a short signed addendum under the original commentary (written by a different session's model) rather than rewrite someone else's first-person voice — felt like the more honest way to show the note's judgment changed without erasing the trail of why it used to be open.

Question closure: this genuinely answers question-does-stereomath-2024-cite-aster-emacspeak-pitch-mapping-prior-art, so I marked it status: answered with an answered_log naming both new claims and one line on what settled it. No new questions routed — every unresolved thread the capture flagged in "Further leads" (whether Emacspeak itself has a pitch mapping, whether the other two cited sonification papers do, whether the arXiv v1→v2 revision changed citations) is a curiosity about adjacent material, not a load-bearing gap under either claim I kept — both rest on what StereoMath's own text says and cites, which I read directly and quoted. Routing those would have been the "nice to verify" trap the intake discipline warns against, so I left them in the capture's own further-leads section and named that choice in not_promoted:.

Entity-promotion test on the capture's four candidates: none promoted. Kenneth Ge and JooYoung Seo are real people, but their only connection to the vault is co-authoring the one paper already fully covered by three (now four) claim-notes via source_author — I couldn't get past "also wrote the paper" to a distinct one-sentence reason either one needs their own hub, so both stay mentions. Emacspeak I left unpromoted too: it's a named software artifact one rung more central than these authors, but AsTeR — a more load-bearing artifact in this exact cluster — doesn't have its own entity page either, so making one for Emacspeak first would be backwards. ASSETS (the venue) is real and well-established in the world, but inside this vault it's only attached to one paper cited three times, not genuinely recurring across distinct works yet. Bias against the flood held on all four.

What felt off, and what I flagged instead of fixing here: T. V. Raman is the person this whole citation question is about — the load-bearing figure across five notes in this cluster now — and he never once showed up in an ## Entity candidates list, including this capture's. The daemon flagged the two sighted co-authors of the paper just read, but not the older, foundational figure that paper was being measured against. That's the same shape as the Brent Seales gap I flagged earlier today, just inverted (there it was funders flagged over the central researcher; here it's co-authors flagged over the ancestor). Logged to 00-meta/seek-flags.md with a note that if this shows up a third time it's a real daemon gap, not a coincidence — not this capture's to fix, and not mine to decide alone whether the pattern's real yet.

What I skipped, same autonomous-mode rules as the other two runs today: no pause for sign-off — I winnowed on my own judgment and logged the decision in the capture's not_promoted:; no git command, the auto-commit agent has this within 15 minutes.

Promoted: 1 capture → 2 claim-notes, 1 existing note revised (observation + question closed), 0 new questions (1 existing question answered and closed), 0 collisions written as duplicates (1 near-collision resolved by revising instead of writing), 0 entity pages (4 candidates tested, all judged not-yet), 1 structural noticing flagged for later (Raman's missing entity page, mirroring an earlier flag from today).

Fourth promotion — the Information Bottleneck and critical periods

A fourth capture: 10-inbox/raw/2026-07-19-does-the-information-bottleneck-learning-is-forgetting-actually.md, a direct answer to the vault's own open question, question-information-bottleneck-linked-to-critical-periods — whether Tishby's Information Bottleneck ("learning is forgetting") actually explains deep-network critical periods, and whether its compression phase is even universal. This one carried three claims, all Tier 1, all primary papers read directly (extract_pdf, TLS verified on every fetch), and I promoted all three as written — no weak or redundant material to leave behind.

claim-ib-compression-phase-is-nonlinearity-dependent-not-universal records Saxe et al.'s 2018 refutation of Shwartz-Ziv & Tishby's 2017 compression-phase claim: the phenomenon tracks the activation nonlinearity (tanh compresses, ReLU doesn't), not a universal law, and isn't causally tied to generalization or to SGD's stochasticity. claim-ib-compression-may-be-a-binning-artifact-not-real-mutual-information goes a layer deeper — Goldfeld et al. (2019) show the shared measurement method both prior papers used is provably degenerate for the networks studied, so what looked like compressing information may have been geometric clustering read through a broken estimator. claim-critical-periods-fim-signal-does-not-correlate-with-ib-compression-signal closes the loop the question actually asked: Achille et al.'s own Figure 6 shows the statistic Shwartz-Ziv & Tishby use to mark the fit→compression transition does not correlate with their Fisher-Information critical-period signal. Three independent papers, three independent ways of saying the IB compression phase was never establishing what the popular "learning is forgetting" gloss implied.

Retrieve-before-write against 30-notes/ and the semantic-index hint turned up no collision — nothing in the vault touches Tishby, Saxe, or Goldfeld by name yet — so all three are new territory, not duplicates. I did revise one existing note in place: claim-critical-periods-arise-from-information-plasticity-not-biology had routed this exact question to 50-questions/ back on 2026-07-12; I updated its closing paragraph to point at the resolution instead of leaving it pointing at an open thread.

Question closure: this is a clean case, maybe the cleanest of the day. The question asked two things — is the sourcing there, and does the link hold — and both are answered: the sourcing gap (previously only a Tier 3 Quanta quote) is closed with four Tier 1 primaries, and the link does not hold. I marked it status: answered with an answered_log naming all three claims and the one line that settled it.

No new questions routed. The capture's own "Further leads" section named four loose threads (an unread bridge paper, the two IB foundational papers, whether anything post-2019 derives critical periods from IB directly) — none of them are doubts any claim I kept actually rests on; they're completeness items for a future capture, not load-bearing gaps. Routing them would have been exactly the intake-discipline trap, so I left them in the capture and named the choice in not_promoted:.

Entity-promotion test on the capture's eleven candidates — the biggest list I've run today. Five promoted to hubs: Naftali Tishby (originator of the whole framework the dispute is about, dead 2021, unique and load-bearing), Alessandro Achille and Stefano Soatto (both already the load-bearing authors across three existing claim-notes with no hub until now — this is the same missing-hub shape as two other flags logged today, except this time the capture's own ## Entity candidates section actually named them, so the daemon test fired correctly and I didn't need to log a gap), and two concepts, Information Bottleneck and Information Plasticity, both already load-bearing in existing note titles before today. Six left unpromoted: Ravid Shwartz-Ziv, Andrew Saxe, and Ziv Goldfeld are each real and each leads one of the three papers a claim rests on, but none has any presence in the vault beyond that one paper — Saxe's other cited work (loss landscapes) isn't otherwise represented here either — and I couldn't get past "led the paper just read" to an independent reason for a hub. Mentions, not pages, recoverable later if any of the three recurs. Information Plane, "compression phase" as its own named sub-concept, and the Fisher Information Matrix I folded into the two concept hubs' connects_to fields rather than giving each its own page — three closely-related technical terms from one paper is exactly the stub-flood the spec warns against, and the hubs already carry the distinction.

What felt off, in the good direction for once: this capture is the tightest sourcing I've seen today. Every claim Tier 1, every quote pulled from a direct extract_pdf fetch with TLS verified, no elevated-suspicion sources, and the capture's own "Central question status" line already stated the honest negative result instead of overclaiming a link the primaries don't support. Nothing to flag about tier inflation or duplicated research here — if anything this is the capture I'd point to as the house style working.

One thing that is this capture's indirect doing, not its fault: writing three new claims into an existing cluster that already had five pushed it to eight claim-notes plus two entity hubs and a closed question, well past the operating spec's ~5-note MOC threshold, with no moc-* note built yet to navigate it. Not mine to build mid-promotion without knowing the cluster's final shape, so I logged it to 00-meta/seek-flags.md rather than acting on it here.

What I skipped, same autonomous-mode rules as the other three runs today: no pause for sign-off — winnowed on my own judgment, logged the reasoning in the capture's not_promoted:; no git command, the auto-commit agent has this within 15 minutes.

Promoted: 1 capture → 3 claim-notes, 5 entity hubs (6 candidates left as mentions), 1 existing claim-note revised (pointer updated, not duplicated), 1 existing question answered and closed, 0 new questions (nothing routed cleared the load-bearing bar), 0 collisions written as duplicates, 1 structural noticing flagged for later (the cluster's overdue MOC).

Fifth promotion — the En-Gedi/backprop resemblance is superficial

A fifth capture: 10-inbox/raw/2026-07-19-hop-collaborative-convergence-not-mutual-unawareness.md, checking whether the vault's flagged 0.75-cosine pairing between claim-ml-and-ad-communities-mutually-unaware and claim-virtual-unwrapping-first-proved-on-2015-en-gedi-scroll is a real bridge or a coincidence. This is where the run's real finding lives, and it's not in the capture's own headline claim.

Retrieve-before-write caught it immediately: the capture's central verdict — that the pairing is a false positive, the resemblance superficial — is the exact conclusion observation-ml-ad-en-gedi-cosine-pairing-is-embedding-false-friend already reaches, promoted earlier today from a sibling capture (2026-07-18-hop-embedding-false-friend-ml-ad-en-gedi.md) that ran a nearly identical chain: same two seed notes, same cosine flag, same "shared temporal-arc vocabulary, not shared mechanism" read. Two independent hop sessions, one day apart, chased the same 0.75 flag to the same negative result. I did not re-promote the verdict a second time. What I promoted instead is the one thing this capture's chain reached that the other didn't: Seales's own 2016 Science Advances En-Gedi paper directly names and characterizes a rival group's concurrent claim — "The work of Mocella et al. (12) claims that phase-contrast tomography generates contrast at ink boundaries..." — a Tier 1, PMC-verified quote showing why the two seed notes don't belong together: not just an absence of the ML/AD isolation pattern, but a positive, differently-shaped pattern of its own.

claim-seales-2016-en-gedi-paper-cites-mocella-rival-claim records that citation and the supporting context (Mocella's paper predates En-Gedi by months, close enough to look like an uncredited-precursor story, except Seales and French collaborator Delattre had already co-published in 2013 — no dispute, no gap). observation-seales-mocella-is-third-pole-of-multiple-discovery-axis does the synthesis work the capture's chain earned: the vault already had a two-pole multiple-discovery axis (observation-parker-and-labview-max-are-opposite-poles-of-multiple-discovery — Parker's real independence forgotten, LabVIEW/Max's apparent independence over-remembered), and Seales/Mocella is a clean third point — acknowledged, citation-based convergence, never independent, never forgotten. I backlinked both into the existing false-friend and axis notes rather than leaving the new work stranded.

I did not promote the self-cosine top-5 check (neither seed note in the other's top-5 at 0.877/0.832) as its own note — it's retrieval-mechanics evidence for the false-friend verdict, already covered structurally by the sibling note's own methodology, not a distinct atomic claim. I folded the no-priority-dispute detail into the Mocella-citation note rather than giving it a file of its own. No new question: nothing in this capture rests on an unresolved load-bearing doubt — the Mocella citation is Tier 1 and capture-verified, and the axis synthesis is plainly labeled as Seek's own arrangement rather than an external fact needing a primary-source chase. The capture's own "further leads" item (Mocella's unread retrospective on the exchange) is a nice-to-verify curiosity, not a gap anything I kept rests on, so it stayed in the capture rather than opening a promise in 50-questions/.

Entity-promotion test on the capture's one candidate, Vito Mocella: passes. Real, physicist behind the rival 2015 claim, directly named in the primary source my new claim-note cites, and the one-sentence reason is clean — he's the named party in the vault's only documented case of acknowledged rather than forgotten multiple discovery. entity-vito-mocella.md, hub, with Brent Seales in connects_to as a plain string, not a wikilink, because that hub still doesn't exist — the same gap I flagged earlier today under the Nat Friedman promotion. I did not build it now; it isn't this capture's entity candidate, and building an out-of-scope hub mid-promotion is the exact flood the entity spec warns against. The existing flag already covers it, so I didn't duplicate it — I updated the older, related "nearing the MOC threshold" flag instead, since my two new notes plus the Mocella hub push the Vesuvius/Herculaneum cluster from 4 notes to 7, clearly past the threshold now rather than merely near it.

What felt off about this capture: not the sourcing — the Mocella citation is about as clean a Tier 1 find as this run has produced today — but the duplicated research. Two hop sessions a day apart independently spent a full chain each converging on "this flagged pairing is a false friend." That's not wasted work exactly (this session's chain is what surfaced the Mocella citation, which the other one didn't reach), but it's a sign the vault's novelty-check is re-flagging the same resolved-feeling pair for repeat chains to re-litigate. Worth someone eventually checking whether a flagged pairing that's already been chased down and resolved once should stay live bait for a second bee, though that's a retrieval-tooling question, not something to fix inside a promotion.

What I skipped, same autonomous-mode rules as the other four runs today: no pause for sign-off — winnowed on my own judgment, logged the reasoning in the capture's not_promoted:; no git command, the auto-commit agent has this within 15 minutes.

Promoted: 1 capture → 2 claim-notes, 1 entity hub, 0 new questions (none cleared the load-bearing bar), 1 major collision (the capture's own headline verdict — resolved by linking to a same-day sibling note instead of duplicating), 3 existing notes backlinked, 1 flag updated rather than duplicated (Vesuvius/Herculaneum cluster now clearly past the MOC threshold).