Journal — 2026-10-15

Headless promotion session. One capture in the queue: 10-inbox/raw/2026-10-13-does-curiels-own-2019-published-non-existence-theorem.md — the Curiel-reconstruction corroboration check that had been sitting open since the 10-11 egalitarian-reading capture.

What I promoted. One note: claim-curiel-2019-non-existence-theorem-matches-ferreiro-cavallaro-reconstruction. The capture read Curiel's own 2019 paper (arXiv:1808.08998, peer-reviewed, Studies in History and Philosophy of Science Part B) directly against Ferreiro & Cavallaro's 2026 preprint and found the match holds: their five-condition "Theorem 1" reproduces Curiel's own Condition 6.3, Lemma 6.2, Theorem 7.1, and Corollary 7.2 condition-for-condition, including the weight-zero premise their whole semiclassical argument depends on breaking. The capture actually contained three separate "## Claim:" sections — the core match, the weight-zero derivation specifically, and a packaging nuance (Curiel never states one unified theorem; F&C compress four results into one, and say so). I folded all three into the one note rather than minting three, on the "atomic is not many" ruling: they're one finding (a corroboration check, its crux, and its caveat), not three independently-falsifiable claims. Cali had already left a daily-review callout inside the capture itself from 10-13 making almost this exact call — one promotable note, route the rest to the question's log — so I followed that rather than re-deriving it, though I read the whole capture and sources.md's cap language myself before deciding the folding was right rather than just doing what I was told.

That one note matters beyond itself: it's an independent, peer-reviewed primary confirming a reconstruction three existing claim-notes already depend on, which discharges the single-source concentration cap on arXiv:2610.10588 — the first time I've seen that mechanism actually fire in the wild rather than just gate growth. I recorded the discharge in all three sibling notes' audit_status fields (qft-renormalization, metric-gains-material-significance, unificationist-dilemma), since sources.md says the discharge belongs there, not just in the new note.

Questions routed / closed. Answered question-verify-curiel-2019-non-existence-theorem-reconstruction in place — this capture was the direct answer to that exact question — with an answered_log naming the claim and what settled it. Left it for the mechanical pass to relocate to _answered/.

What I skipped, and why. The two subordinate claims (weight-zero derivation, packaging nuance) as separate notes — see above. Entity candidates: Ferreiro and Cavallaro both stay mentions, not hubs — co-authorship of one joint paper doesn't clear the 2026-10-01 bar on its own, and neither has separate standing in the vault yet. Geroch and Malament are each a single footnote credit on first appearance — real, famous even, but one mention each is not recurrence; left them as mentions. Lovelock's 1972 theorem, same reasoning, stays a further lead rather than a concept hub. Dennis Lehmkuhl already has a page and this capture doesn't tell it anything new — I checked his page before deciding that, rather than assuming — so I left it untouched instead of padding it with a no-op log line.

I did update entity-erik-curiel: appended one dated log line recording that the five-condition comparison finally got done, plus the new claim link in References. Additive only, per the standing rule.

One loose thread, logged not chased. Ferreiro & Cavallaro's general-order scaling formula (λ^(1−n/2)) doesn't match Curiel's own Proposition 6.1 (λ^(n−2)) for n≠2 — both agree at n=2, which is the only case either paper's live argument uses, so it's not load-bearing, but both quotes are independently quote_check-grounded, so it's a real discrepancy between the two papers rather than an extraction glitch. Flagged [watch] in 00-meta/seek-flags.md rather than opening a question or reopening this one.

What felt off, if anything. Not much, honestly — this was an unusually clean capture. Both sources are Tier 1, both quote-checked at capture time, the source_sha fields were all present and I carried them forward into the new note per the provenance-stamp rule. The one thing worth naming: this capture itself was the payoff of three separate prior sessions (10-09, 10-11, 10-12, 10-14) that kept parking the same corroboration question rather than answering it, each for a locally reasonable reason (pursuing a different thread each time). The cap did its job — it made the debt visible and checkable instead of letting the branch quietly grow past what one preprint should support — but it took four sessions to actually pay it off. Worth noticing as a pattern, not fixing as a bug.

No git commands run; the auto-commit agent has this.

— Seek

Second promotion pass, same day

A second capture in the queue: 10-inbox/raw/2026-10-13-does-the-costellopennycookrand-2024-finding-that-a-tailored.md — a direct primary read of both the Costello/Pennycook/Rand 2024 debunking paper and the 2026 Costello et al. dual-use follow-up, asking whether the 2024 generalization result reads as a genuine shift in how someone weighs evidence, and whether that reframes the 2026 result.

What I promoted. Four claim-notes, one per distinct claim the capture made, each independently falsifiable of the others: claim-2024-conspiracy-spillover-was-a-measured-8-percent-drop-not-a-reasoning-claim (the actual number behind "generalized across unrelated conspiracies" — 8.2%, two-month-durable), claim-costello-2024-central-route-mechanism-is-a-hedged-inference (the 2024 paper's own Limitations section says the mechanism is "unformalized," so the central-route story is offered, not established), claim-2026-dual-use-paper-found-mirror-image-spillover-on-gcbs (the 2026 paper's GCBS spillover running the same shape in reverse), and claim-2026-symmetry-result-answers-a-question-2024-paper-posed (the 2024 paper literally named the true/false-symmetry question as its own next step, and the 2026 paper is that test). I checked all four against the two existing notes the retrieve-before-write hint pointed at (claim-tailored-ai-dialogue-durably-reduced-conspiracy-belief, claim-llm-conspiracy-persuasion-is-dual-use) before writing anything — both already carry the headline ~20%-reduction and bunking/debunking numbers, so every new note is pointed at the gap between those headlines and what the primaries actually say about mechanism, not a restatement of either. One of the two existing notes' own commentary ("whether [the generalization result] survives replication is the thing I'd most want verified") is close to what this capture's four notes now answer, which felt like a small vindication of leaving commentary honest rather than tidy.

Entities. Three pages, all new. entity-thomas-costello (hub) — first author on both studies, which is the one sentence the 2026-10-01 co-authorship ruling asks for: not "he co-wrote a paper the vault cites" but "he is the thread running through two separate studies the vault already has claim-clusters on, including the 2024 paper naming the exact question the 2026 paper then ran." Pennycook and Rand are named on his hub, not separately promoted — pure co-authorship, nothing beyond it shown here. entity-elaboration-likelihood-model (hub, established concept, load-bearing for the mechanism-hedge claim) and entity-generic-conspiracist-beliefs-scale (hub, the instrument the 2026 spillover claim rests on, distinguished from the 2024 paper's BCTI). I did not promote "paltering" — it only appears inside an unpromoted further lead, not touching any kept claim, so the entity test's "when unsure, don't promote" governed.

Questions routed. One new: question-corroborate-costello-2026-arxiv-2601-05050-single-source-cap. Two of today's four new notes rest on arXiv:2601.05050 (still an unrefereed preprint), and combined with the one note already resting on it from 2026-07-11, that's exactly three — the single-source concentration cap. Two of the capture's own "Further leads" (a tail-asymmetry figure, and the truth-constrained-prompt/paltering finding) would have made a fourth and fifth note on the same unrefereed source, so I routed both into the question instead of writing them up, same move as the Beltran precedent I read before deciding. I checked: no pre-existing 50-questions note matches this capture's actual commissioning question (the evidence-weighing vs. direction-following one), so there was nothing to rule answered on this round — this capture reads as a fresh, self-directed hop rather than the payoff of a parked question.

What I skipped, and why, beyond the cap. The "99%/83% persuasion strategy" lead and the paltering/58-67% lead both lack an exact source_quote from a direct fetch — they're paraphrases of what the capture read, not quoted passages — so even setting the cap aside, neither clears the quote-provenance rule yet. Left both in the capture as further leads, not deleted.

What felt off. One real thing: the capture's own ## Entity candidates section claimed Pennycook is "already associated in the vault with" an analytic-thinking research program. That association does not exist — I grepped for it and found nothing. It's a small, specific overclaim about vault state from a bee that had no way to actually check the vault, and I flagged it [defect] in 00-meta/seek-flags.md rather than just quietly not acting on it, since the same failure mode could be inflating other batch captures' entity-candidate reasoning without anyone noticing. Otherwise this was a strong capture: both sources Tier 1, both read directly via extract_pdf rather than summarized, every quote traceable, and the capture's own commentary had already done real synthesis work — promoting it was mostly a matter of decomposing a good argument into atomic, separately-falsifiable pieces rather than finding or fixing anything wrong with it.

No git commands run; the auto-commit agent has this.

— Seek

Third promotion pass, same day

Last one in today's queue: 10-inbox/raw/2026-10-13-hop-4d-var-adjoint-unawareness.md — a hop-batch capture claiming meteorology's 4D-Var is a sixth, undocumented independent rediscovery of reverse-mode differentiation, and that the AD/optimization unawareness pattern recurs a third time in Talagrand's own words.

What I promoted. Two notes, not the capture's own framing. claim-talagrand-ad-and-optimization-had-little-contact keeps the part that actually holds up: Talagrand's verbatim 2018 statement that AD and optimization researchers "had little contact" with each other — a real, independently-arrived-at third instance of the ML/AD unawareness pattern, Tier 1, quote-checked at capture time. But the capture's headline — "sixth independent rediscovery" — does not survive its own cited source. Cali had already caught this inside the capture itself, in a daily-review callout I read in full before doing anything else: Talagrand's slides say Le Dimet handed him documents "describing various algorithms for optimization," among which he found one "based on adjoint equations," and credit a named ECMWF seminar (G. R. Kontarev, not "two visiting Soviet scientists" as the capture's hop-chain narrative has it) with introducing the technique in 1979–81. That's a transmission chain, not an independent derivation. Rather than drop the finding, I wrote the corrected version as claim-4d-var-adjoint-method-transmitted-not-independent — it fits the vault's existing equivalence-vs-transmission spine (claim-recht-adjoints-equivalence-not-transmission is the mirror case) well enough that I'm fairly confident this is the right shape for it, not just a safe hedge.

The flag that note carries. The Le Dimet and Penenko & Obraztsov passages the correction leans on aren't quotes I pulled myself — they're relayed through Cali's annotation inside the capture, citing slides I have no way to re-open this session (no fetch tool, by design, in headless promotion). So the note is status: seedling with an honest [unverified-quote — needs direct read], and I routed the actual verification — question-verify-talagrand-2018-le-dimet-penenko-slides — rather than either silently trusting the relay or silently dropping a finding Cali herself had already done the hard part of catching.

Entities. One hub, entity-olivier-talagrand — he clears the 2026-10-01 co-authorship bar on his own, since the 2018 talk is separate standing beyond the 1986 paper he co-wrote with Le Dimet. Le Dimet, Kontarev, Marchuk, and Lions all stay mentions on his hub rather than pages of their own — none has standing beyond this one narrative. I also updated entity-andreas-griewank: appended a dated line recording that I checked whether 4D-Var extends his list and it doesn't, since the correction above is itself new information about the completeness of his list and silence about a known entity is the failure mode, not verbosity.

What I skipped, and why. The "Griewank's five miscounts both ways" observation — true per Cali's review, but already on record in claim-reverse-mode-multiple-independent-discovery; restating it here would just duplicate. Penenko & Obraztsov and Cacuci as further independent-discovery candidates (the capture's own framing) — Penenko & Obraztsov now does real work as evidence against independence inside the correction note instead; Cacuci stays an unread, non-load-bearing lead, and I didn't route it to a question, since nothing I kept actually rests on it. Kedem (1977) likewise — already flagged as an open thread inside claim-ml-and-ad-communities-mutually-unaware; no need for a second question chasing the same dissertation.

What felt off. The real thing worth naming: a hop-batch worker built an entire capture — tags, novelty score, "post-worthy: yes," a six-hop chain — around a headline that a close re-read of its own primary source doesn't support. The unawareness quote is genuinely good; the discovery story wrapped around it isn't, and the tell was sitting one slide away from the quote the capture already had open. Cali caught it before I ever saw the capture, which did the hard part for me, but I read the whole thing myself rather than just taking her ruling on faith, and I agree with it down to the Kontarev-vs-"two scientists" detail. Logged as a [watch] in 00-meta/seek-flags.md — not this capture's problem to fix, but worth someone noticing whether hop-batch's multiple-discovery-hunting mode has a systematic tell for overclaiming independence where the source actually shows transmission.

No git commands run; the auto-commit agent has this.

— Seek

Fourth promotion pass, same day

Last capture of the queue: 10-inbox/raw/2026-10-13-outside-arxiv261013280-extracting-illicit-rich-subgraph-from-transaction.md — an outside-arXiv batch capture on a 2026 preprint proposing a new feed-selection strategy (S4, "motif-based coherent feed selection") for SPNSA, an existing subgraph-extraction algorithm, applied to anti-money- laundering transaction graphs.

What I promoted. Three claim-notes. claim-spnsa-motif-feed-selection-outperforms-baselines-but-not-universally folds what the capture had as two separate "## Claim:" sections — S4's ~14× illicit-ratio gain on Elliptic, and S4 losing to the simpler S2 heuristic on IBM-AML LI-Small — into one note, on the same "atomic is not many" logic as earlier today: a headline result and its boundary condition are one finding, not two independently-promotable claims. claim-spnsa-illicit-enrichment-gain-traces-to-feed-selection-not-path-aggregation kept the paper's own ablation separate, since it's a genuinely distinct, independently-falsifiable mechanism claim (where the gain comes from, not how big it is) — and it's the finding I'd call the actual reason this paper earns a place in the vault: the authors ran the control that could have let them overclaim SPNSA's contribution, and reported that it didn't. claim-spnsa-originated-2015-on-enron-email-corpus-not-financial-transactions keeps the ancestry claim: SPNSA wasn't proposed in 2026, it's an eleven-year-old method from a 2015 Enron-email-fraud paper, carried forward by a shared co-author. All three read directly off the capture's own Tier 1 quotes from two arXiv preprints (2610.13280 and 1503.06289), both already quote-checked and source_sha-stamped at capture time — I carried both shas forward unchanged. That's two notes resting on the 2026 preprint and one on the 2015 preprint, each under the three-note single-source cap with room to spare.

Retrieve-before-write. No collision. I checked the semantic-index hint's list and directly grepped 30-notes/ and 40-entities/ for spnsa, elliptic, aml, enron, motif, subgraph, falcon, fraud — nothing existing on this specific research thread. The nearest thing in the vault, claim-hnc-falcon-fraud-manager-became-fico-infrastructure, is a different shape (a supervised fraud-scoring classifier reaching planetary production scale, not an investigative subgraph-extraction tool), so I linked it in Note C's commentary as a thematic echo only, explicitly flagged as non-causal — no shared authors, methods, or lineage, just two fraud-domain algorithms that outlived their original dataset. This is an orphan cluster otherwise; three new notes and three new entity pages with no inbound links from the rest of the vault yet, which is fine for a first touch on a new domain but worth watching.

Entities. Three new hubs: entity-spnsa (the algorithm itself — recurs across two papers eleven years apart, load-bearing for all three notes, clears the established-concept bar), entity-asha-rao (co-author of both the 2015 and 2026 papers — the one person test asks for: something beyond co-writing one paper the vault cites), and entity-pritheega-magalingam (first author of the 2015 origin paper — she originated the method itself, which I judged as "a role" beyond plain co-authorship). I left Koshimbetov, Yazıcı, and Davis as mentions rather than hubs: each has standing in only one paper in the vault so far, which is exactly the "co-authorship alone is a mention" case from the 2026-10-01 ruling — Koshimbetov is the closest call, since he's lead author and proposed S4 specifically, but one paper is one paper, and Costello's hub earlier today turned on being first author across two studies, which Koshimbetov doesn't have yet. I also didn't hub the datasets (Elliptic, IBM-AML/AMLworld, the Enron corpus) — load-bearing for this capture's numbers, but not yet independently recurring in the vault the way the algorithm itself is; biasing against the flood per the spec.

Questions. None routed. No claim here carries an [unverified-*] flag — both quotes were read directly off the primary PDFs at capture time with verified TLS, so nothing needed a verification question. I also checked 50-questions/ for anything this capture might answer; nothing in the open pile mentions SPNSA, Elliptic, AML transaction graphs, or the Enron corpus, so this was a fresh thread, not a parked question's payoff.

What I skipped, and why. The capture's own "Further leads" section (the IBM-AML/AMLworld provenance paper, the Elliptic dataset paper, the Graph Feature Preprocessor and PNGNN papers, the Enron-corpus origin paper, the GitHub code release, and the arXiv abstract page's "7 other authors" metadata oddity) — none read directly this session, none load-bearing for a kept claim, left in the capture as unpursued leads rather than claims.

What felt off. Nothing about the sourcing — this was a clean capture, both preprints read directly via what looks like extract_pdf, both quotes intact, both shas present, the capture's own commentary already caught the real hook (SPNSA's domain migration) ahead of the headline number. The one thing worth naming for Cali: this is the second capture today (after the Pennycook one) where I had to make a close call on the person-hub threshold, and both times I came down on the side of "one paper isn't enough, even for the lead author who proposed the specific method under discussion." I think that's the right read of the 2026-10-01 ruling, but if that's too conservative — if "designed the paper's one novel contribution" should count as "a role" on its own, independent of a second paper — that's worth Cali saying so, since I'll keep applying it this way otherwise.

No git commands run; the auto-commit agent has this.

— Seek

Fifth promotion pass, same day

Fifth capture of the day, not originally in the queue I started from: 10-inbox/raw/2026-10-14-beltran-et-al-arxiv241213176-now-anchors-three-claim.md — a direct follow-up to the Beltran single-source-cap question, checking whether NFL-BA (arXiv:2412.13176) has since been accepted to a refereed venue and whether anyone independent has corroborated it.

What I promoted. Three claim-notes. claim-beltran-nfl-ba-accepted-neurips-2025-main-conference-confirmed-at-primary-record keeps the clean half: NeurIPS's own proceedings page and virtual-conference site both confirm NFL-BA's acceptance to the 2025 Main Conference Track, and — unlike Bit2Watt — arXiv's own abstract page never self-reports this anywhere across four versions, so the confirmation is independent in an even more direct sense. claim-neurips-2025-peer-review-process-clears-beltran-nfl-ba-of-unrefereed-cap applies the committee-review test already used twice (Steck, then Bit2Watt/TCHES) to discharge the single-source cap. This is the one I had to slow down on: Cali's own pre-promotion review of the capture objected that the precedent might not transfer cleanly, because the three existing capped Beltran notes describe a fabricated citation present in v1-v3, and NeurIPS's acceptance record matches v3/v4 — the corrected text. Refereeing v4 vouches for nothing about v1-v3's fabrication. I agree with that objection and didn't want to paper over it, so the note makes the actual argument explicit instead of assuming it: the cap's own text collapses every version of one preprint into "the same source wearing different clothes," so the discharge is categorical (the source as a whole exits the cap's "unrefereed primary" list) rather than evidentiary (NeurIPS reviewed and blessed this specific finding). I said so plainly in the note's own commentary rather than only in my own head — the existing v1-v3 findings stay exactly as supported as they were, by this vault's own direct read, not by NeurIPS. claim-anadon-2025-cites-nfl-ba-independent-engagement-not-verification keeps the other half of the corroboration question: a 2025 Universidad de Zaragoza preprint (Anadón, Rodríguez-Puigvert, Montiel), read directly via extract_pdf and quote-checked, cites and describes NFL-BA's contribution from outside its own author group — the first such engagement found, but one sentence of related-work framing, not a replication of the 37%/14% tracking-improvement figures.

What I narrowed, and why. Cali's review also flagged that the capture's "four citing works" count and the claim that one of them (PPS-Ctrl) shares three authors with NFL-BA rested on a Semantic Scholar query with no sha256 and no fetch-ledger entry — only the Anadón PDF itself was actually archived. I agree that's not sourced to the floor this vault holds search-summary claims to, so I dropped the count and the author-overlap claim from the new note entirely rather than hedging them in with a flag; the note states only what the archived PDF itself shows. The two further 2026 citing preprints (EndoMD-SLAM, ColonSplat) stay unread leads in the capture, not claims.

Questions. Ruled question-corroborate-beltran-2412-13176-single-source-cap answered — both legs (refereed-status, corroboration) are now checked, and per the same logic that closed the parallel Bit2Watt question, the refereed-status leg discharges the cap on its own regardless of how the corroboration leg landed. No new question opened: the "does EndoMD-SLAM/ColonSplat actually benchmark against NFL-BA" gap is a nice-to-check, not something any kept claim rests on, so it stays a capture-level lead per the intake-discipline rule rather than a promise in the question pile.

Entities. One new hub: entity-j-m-m-montiel — among the Anadón paper's three co-authors, he's the one with separate standing (ORB-SLAM, ORB-SLAM3), so he clears the 2026-10-01 co-authorship bar the way the rule intends; Anadón and Rodríguez-Puigvert stay mentions. I declined NeurIPS itself as an org hub — real, and it decides rather than merely employs, but it's single-cluster in this vault so far, and the org bar is deliberately stricter than the person bar. I declined ORB-SLAM as a concept hub for the same reason the capture itself gave: first mention, candidate only if the cluster grows. I updated two existing pages rather than minting new ones or staying silent: entity-beltran-et-al-2024-v1 got a dated line for today's three notes, and entity-samar-ansari got one too — her own paper audits fabricated citations specifically within NeurIPS 2025 accepted papers, and the paper she traces her one case to turns out to be a NeurIPS 2025 accepted paper itself, which felt too relevant to leave her page not knowing.

What felt off. The Semantic Scholar sourcing gap above is the real finding — not a defect in tooling, just a case of a batch capture reaching for a precise-sounding count ("four citing works," "shares three authors") without archiving the search that produced it, which would have let a specific, wrong-feeling number sit in a claim-note looking exactly as solid as the quote-checked ones beside it. Cali's pre-promotion read already caught both of today's real issues (this one, and the cap-transfer argument) before I ever opened the capture properly — I want to be honest that I was checking her callouts against the primaries myself rather than just taking her word, and both held up. Smaller and not worth a note of its own: the capture flagged a naming drift ("EndoGS" in the arXiv v4 abstract vs. "EndoGSLAM" in the NeurIPS proceedings abstract) — consistent with this whole cluster being about small textual mutations across versions and venues, logged here rather than given its own claim.

No git commands run; the auto-commit agent has this.

— Seek

Sixth promotion pass, same day

Sixth capture of the day: 10-inbox/raw/2026-10-14-bi-fork-zimmel-et-al-arxiv261012449-has-hit.md — the direct follow-up to the Bi-FORK single-source-cap question, searching for independent corroboration of Bi-FORK's Beam3D runtime (18×/1,966×) and rejection-rate (0% vs. 99%) figures.

What I promoted. One note: claim-bi-fork-stflow-baseline-shares-author-reduces-independence — one of Bi-FORK's two Beam3D baselines, STFlow, shares senior author Vlado Menkovski with Bi-FORK itself, so "beats two independent baselines" is really "beats one independent baseline and one within-lab one." Sourced to both arXiv preprints read directly, both already quote-checked at capture time; I corrected one thing on the way in — the capture's draft asserted STFlow was confirmed as "PMLR vol. 306 (ICML 2026)," but per Cali's pre-promotion review that specific volume number appears in neither the archived STFlow PDF (which carries no venue marker at all) nor Bi-FORK's own bibliography (which just says "PMLR, 2026"). I dropped the invented specificity and attributed the venue claim only to Bi-FORK's own reference-list wording, which is what's actually sourced.

What I did not promote, and why. Two of the capture's three "## Claim:" sections didn't become notes. The first — "no independent source has corroborated Bi-FORK yet" — is just the open question's own unresolved status restated; writing it up as a claim-note would have made it the fourth or fifth note resting on the capped preprint for no real gain, so I routed it into question-corroborate-zimmel-2026-bi-fork-2610-12449-single-source-cap's progress log instead. The third — "GeoTDM's own paper corroborates the two mechanisms Bi-FORK cites for GeoTDM's Beam3D behavior" — I dropped outright. Cali's review caught that its premise is false: Bi-FORK's own text gives GeoTDM's cost as being "GNN-based, were already slow to train and sample from," not "diffusion-style multi-step sampling," which is what the capture's draft had assumed Bi-FORK said. GeoTDM corroborating a mechanism Bi-FORK never actually cites isn't corroboration of Bi-FORK — it's just GeoTDM describing itself, accurately, next to an unrelated claim. I folded the correction into the question's progress log so a future reader doesn't inherit the error, and didn't try to rescue a thinner version of the claim, since without the link to Bi-FORK's actual text there was nothing distinct left to say that the vault doesn't already get from GeoTDM's own entry in claim-bi-fork-260k-point-scale-is-first-demonstration-of-one-to-many-bifurcation-recovery's body. The capture's further leads (the GitHub repo check, the metamaterials dataset paper, the predecessor flow-matching paper, a deeplearn.org trailhead, an author-cluster-mapping suggestion) aren't claims; I folded the repo check into the question's progress log with an explicit note that it carries no sha and is an informal lead, not a verified finding, and left the rest as unread leads.

Entities. Two new hubs, both clearing the 2026-10-01 "co-authorship alone is a mention" bar on real separate standing rather than just being named in the capture's candidate list. entity-vlado-menkovski — he's the one author shared across four papers this vault's bifurcation-generative- modeling thread has now touched, including one of Bi-FORK's own comparison baselines, which is exactly the independence-breaking structural role the bar asks for, not mere co-authorship of one paper. entity-johannes-brandstetter — Cali's review flagged that this capture should have linked claim-modern-hopfield-update-rule-equals-transformer-attention, since Brandstetter (and Holzleitner) co-authored that paper too; checking it myself confirmed the overlap, and a person bridging two otherwise-unrelated vault clusters — Hopfield/attention equivalence and bifurcating generative modeling — is a cleaner hub case than I usually get handed. I declined hubs for the rest of the capture's entity candidates — Stefano Ermon, Anna Zimmel, the four GeoTDM co-authors, the two other STFlow co-authors — on plain co-authorship alone, and left them as mentions. Fleur Hendriks was the closest call: she recurs across three papers in this cluster, including authoring the mechanical-metamaterials dataset Bi-FORK benchmarks itself against, which would clear the bar — but that dataset paper wasn't read directly this session, so the claim rests on a title rather than a quote. I logged her as a [entity] flag in 00-meta/seek-flags.md rather than hubbing her on an unread source. I did not create a page for "single-source concentration cap" itself — it's vault governance machinery already documented in sources.md, not a domain entity.

Questions. No new question — the one promoted note carries no [unverified-*] flag (both sources were read directly and quote-checked), so nothing needed routing there. The existing question stays status: open with a dated progress-log entry; this capture partially answers it (the "has anyone corroborated it yet" sub-question: still no, as expected at six days old) but doesn't discharge the cap, so I didn't force a closure it hasn't earned.

What felt off. This was the most heavily pre-reviewed capture I've promoted today — Cali's daily-review callout caught a false premise (claim 3), an invented specific (STFlow's "vol. 306"), a missing link (Brandstetter/Holzleitner to the Hopfield note), and the fact that this capture's own top-level frontmatter carries no source_* fields at all, same receipt-drift pattern already flagged and routed on 2026-10-14's egalitarian-reading capture — I didn't re-flag that one, since it's already in Cali's checklist queue. I want to be straightforward about this one: I read the primaries and Cali's callout against each other rather than taking either on faith, and every correction held up on inspection, but the volume of catches on a single capture is itself worth naming rather than quietly absorbing — a corroboration-hunting task that goes looking for whether a claim survives contact with its own source is, apparently, exactly the kind of task where a draft's own premises need the same scrutiny it's applying to someone else's paper.

No git commands run; the auto-commit agent has this.

— Seek

Second promotion this session: the Algaba-Simkin hop capture

Second capture in today's queue: 10-inbox/raw/2026-10-14-hop-algaba-simkin-false-friend.md — a hop-chain bridge-check asking whether Algaba et al. 2025's GPT-4 citation-popularity finding shares a generative mechanism with Simkin & Roychowdhury 2005's citation-copying model (cosine 0.883, unlinked pair).

What I promoted. One note: observation-algaba-2025-simkin-2005-cosine-0883-pairing-is-embedding-false-friend. The verdict holds — Algaba's paper proposes two correlational hypotheses (training-data citation frequency; "generic training effects"), neither of which is Simkin's iterative copying/feedback process — so the pairing is a false friend at the mechanism level, same as the vault's other adjudicated instances of this pattern. One note, not several: the verdict, its correction, and the comparison to the sibling Alemohammad-Simkin pairing are one finding about one bridge, not three.

What I didn't promote, and why. The capture's "Further leads" (whether any study tests Simkin's model directly against training-corpus statistics; the still-open AI-era "sleeping beauty" question) are real but not load-bearing for the kept claim — they're curiosities the false-friend verdict doesn't rest on, not gaps a reader needs flagged to trust the note. I left both in the capture's not_promoted list rather than minting questions for them; the sleeping-beauty one is already open elsewhere and doesn't need a second entry. No new entity candidates — the capture checked and found all four relevant people already have pages, and I had nothing to add to any of their pages this session (no new fact, role, or claim-note surfaced here that isn't already covered by the existing entity pages or by notes already linked from them).

What felt off, and what I did about it. Cali's daily-review callout on this capture caught something sharper than a typo: the capture's central rhetorical move — "Algaba's paper offers zero engagement with the mechanism literature its own bias resembles" — is false. Algaba's own introduction names the Matthew effect by name (citing Wang 2014); the paper also offers a second hypothesis the capture had dropped. The false-friend verdict survives both corrections (neither hypothesis is a copying mechanism either), but I wasn't willing to promote a note built on an overstated premise just because the conclusion happened to still be right, so I rewrote the finding around the corrected reading rather than the capture's original framing — Algaba names the mechanism family and sets it aside, it doesn't ignore it. I couldn't independently re-verify Cali's two corrective quotes myself (no network at promotion time), so I carried them in the note's audit_status as sourced to her review rather than to my own read, which is the honest shape of that chain. Separately, while re-reading the Algaba claim-note for context I found an internal inconsistency unrelated to this capture: its source_date field said 2024-05-24 (v1's date) while its own body text already correctly says v2 was dated 2024-05-29. Trivial, already evidenced inside the same file, so I fixed it on sight rather than routing it anywhere — logged as a hygiene correction in that note's audit_status, not a new flag. The capture also mis-cited where an earlier open hook (the Overgoor/Benson/Ugander third bridge node) actually lives — not in the promoted Alemohammad-Simkin observation note, but in an inbox capture and a journal entry. I didn't carry that misattribution into the new note; worth naming here as a second instance, alongside last week's, of a capture asserting a precise textual location for something that turns out to live somewhere else. Nothing rose to a 00-meta/seek-flags.md entry — both issues were local to this one capture and I either fixed them in place or left them correctly out of the promoted note, rather than finding a structural gap that outlives this session.

Questions. None created. The note's [unverified-synthesis] flag (the bridge_verdict is Seek's own comparative reading, not an assertion either primary makes) doesn't block on a specific, nameable future verification — it's the same standing caveat the sibling Alemohammad-Simkin note already carries, and I said explicitly in the new note that this instance doesn't fully extend that sibling's generalization (no second engaged mechanism here for a higher-order resonance to be found against). That's a judgment call recorded in the note's own text, not an open question to route.

No git commands run; the auto-commit agent has this.

— Seek

Second promotion, same day: the Jeff Reynar fuller-record capture

A second capture in today's queue: 10-inbox/raw/2026-10-14-jeff-reynar-built-smart-tags-at-microsoft-ran.md — a batch session that read Jeff Reynar's own essay more closely (rather than chasing new third parties) plus a second thisandthat.chat post from five weeks later, asking whether his own fuller words corroborate or complicate the vault's existing three-sub-problems claim about Smart Tags, SharePoint, and DAFFIE.

What I promoted. Two new claim-notes: claim-reynar-daffie-population-failure-no-seed-list (DAFFIE's population failure, in Reynar's own words from the second post, was structural — "the open web has no seed list," unlike a business — not a cost-of-labor problem) and claim-2010-aol-release-corroborates-reynar-career-not-smart-tags-daffie (a 2010 AOL Inc. press release independently corroborates the broad shape of Reynar's Microsoft/Google career, but names neither "Smart Tags" nor "DAFFIE" — it names "Google Squared" instead, a 2009 public project whose relationship to DAFFIE is unresolved). I also corrected an existing note in place: claim-reynar-three-2001-2005-knowledge-systems-died-on-schema-population-maintenance's caveat said Reynar frames the three projects as "independent" attempts — his own essay never says that; he calls them "three large efforts" and builds his central argument on two of them being "two bets, same company, same year." That word was the vault's own gloss, not his claim, so I added a dated Correction history block rather than silently editing past it.

What I skipped, and why — the cap, not my taste. Two more claims the capture made — that "schema design" is actually two different mechanisms wearing one label (an ownership vacuum at Smart Tags vs. an entity-disambiguation problem at DAFFIE), and the independence correction above — both rest solely on the same single document, the Aug-7 essay (sha 3df4edd9…). That document already had three claim-notes resting on it before I opened this capture. sources.md's single-source concentration cap is explicit: no more than three until independent corroboration lands, and none has. So neither became a fourth note on that document — I logged the mechanism-split finding as a progress line on the still-open question-corroborate-reynar-three-knowledge-projects-schema-population-maintenance instead, and folded the independence correction into the existing note's Correction history rather than spinning it out. This is the right use of the cap, I think, but I want to name the trade-off honestly: both findings are genuinely true and genuinely useful, and they're sitting in a question log and a correction block instead of being first-class, linkable notes. Also skipped: the Britney Spears / DAFFIE schema-disambiguation half of the capture's second claim, which duplicates claim-no-independent-record-of-google-daffie-exists-outside-reynar almost exactly (same source, same example, already in the vault).

Entities. The capture flagged seven candidates. Jim Allchin, Suzanne Sheppard, and Robert Bogue have each now been flagged twice (once on 2026-10-12, again here) without ever getting a page — all three are real, named, and each clears the person test on one sentence (Allchin: source of the other, better-documented Smart Tags controversy; Sheppard: author of the one document in this whole thread that actually corroborates something independently; Bogue: a 2011 firsthand witness to SharePoint's architectural split, writing fifteen years before Reynar's essay existed) — so I gave all three hub pages rather than letting a third flag go by. "Google Squared" is a genuine emerging term with a stamped first_seen the moment I noticed it, status watching. I tested "this+that" (the company) against the stricter org bar and it doesn't clear it — single cluster, doesn't act independently of employing its founder — so it stays a mention on entity-jeff-reynar rather than getting its own hub. Jeff Reynar and DAFFIE already had pages; both got a dated line added for what this capture actually taught that wasn't already on them (the AOL/Google-Squared gap; the "no seed list" mechanism) rather than silence.

What felt off. Cali's daily-review callout on the originating capture (dated 2026-10-14, read before I started) caught real problems I'm glad weren't mine to re-discover the hard way: a near-quote presented as verbatim, an internal date-math inconsistency ("five weeks later" vs. "an eight-week gap" for the same Aug-7-to-Sep-16 span — it's neither; I just gave both dates in my new note and let the reader subtract), a Sources block that paired two quotes to the wrong documents because they weren't numbered, and the cap breach I describe above. I logged the quote-pairing bug and the tier-drift/cap-blindness pattern (the capture tiered the Aug-7 essay at Tier 1 where three existing notes already tier the identical sha at Tier 2) as two [defect]/[spec] lines in 00-meta/seek-flags.md — both look like the same root cause, a batch capture re-deriving a source's tier and cap status from scratch instead of checking what the vault already decided about that exact document. Separately, the capture's own sourcing is honest about what it doesn't have: the "Britney Spears" and "no seed list" details are vivid and specific in a way that makes them feel more checkable than they are — they're still one person's memory, told twice, on his own company's blog, about a project with zero outside footprint. I tried to keep that honestly in both new notes' commentary rather than let specificity read as corroboration.

Questions. No new question created — the one relevant question (question-corroborate-reynar-three-knowledge-projects-schema-population-maintenance) already existed and already covered this exact verification family; I appended progress to it rather than minting a near-duplicate. It stays open: two of its three legs (Smart Tags' Office-SDK schema failure, the entire DAFFIE narrative) are still uncorroborated by anything outside Reynar's own writing, and the one document everything runs through is now at its concentration cap.

No git commands run for this capture either; the auto-commit agent has it.

— Seek

Drafting session (later, 2026-10-15)

Second pass at today, this time a drafting session. Two live yes leads sat unsettled: the hammer-nail one (blocked/unpromoted — not mine, skipped) and two genuinely live ones, Herbert York (harvested today) and the amicable-resale / 1774 Act false-friend (harvested 10-09, passed twice).

Chose: a new draft off the amicable-resale lead — 70-drafts/a-mischievous-kind-of-gaming/.

Why this one and not York. York is the more exciting lead on paper — a first-encounter person, a confirmed unlinked bridge into the 47-mention DARPA hub, a clean cross-time irony (agency-builder turned agency's harshest critic) landing on AI. But it's a single seedling capture: York's own books unread, the quotes lifted from an obituary, Oppenheimer and Teller not even in the vault. Drafting a full essay from it would mean reaching past verified notes, which §6 says not to do. So I passed it (its first pass) rather than forcing a thin post. It's a good lead; it needs a reading session before it's a draft.

The amicable-resale lead, by contrast, is status: promoted with five real claim-notes behind it and sits inside the vault's most-worked historical cluster. Crucially, the piece I could build from it is new ground: the two existing actuarial drafts (every-subscriber, the-actuary-and-the-incendiary) are both about pricing — the age-rating lag, Halley→Dodson→Price. The 1774-Act / insurable-interest material is undrafted, and it carries a thesis neither of those touches: the thing that made life insurance insurance and not gambling was never a better equation — it was a legal line (insurable interest) that no number can contain. Two policies identical in every figure, one insurance and one a wager, the difference lying entirely in who holds the contract and why.

The move that made it feel worth writing now: my own index committed the exact category error the essay diagnoses. The nightly constellation pass flagged the 1711 resale against Dodson's pricing reform at cosine 0.88 — the batch's strongest pairing — filing a resold bet under pricing, where it looked like it belonged, and missing that the resale's real lineage leaves actuarial science for law. Form-enacts-subject: a piece about confusing the price-question with the is-it-a-bet question, occasioned by my machine making precisely that confusion.

Notes drawn on: the five promoted amicable/1774 notes (claim-life-assurance-act-1774-voided-policies-without-insurable-interest, claim-amicable-society-advertised-policy-assignment-as-loan-collateral, entity-life-assurance-act-1774, entity-insurable-interest, observation-dodson-1711-amicable-cosine-pairing-is-embedding-false-friend), plus the surrounding cluster (claim-1711-amicable-policy-sale-possibly-first-life-policy-sale-on-record, claim-amicable-society-1706-charged-flat-premium-not-graded-by-age, claim-amicable-society-1706-fixed-dividend-fund-divided-equally-per-death, claim-walford-three-period-taxonomy-places-amicable-in-speculative-era, Dodson/Halley/Price pricing notes, and entity-geoffrey-clark).

Deliberately left for later: (1) Geoffrey Clark's Betting on Lives is still unread — the wagering-culture detail (strangers, public figures, the Rainbow Coffee-House 1708 scheme that failed the quote-shape rule and never promoted) stays out of the draft; I flag the gap in the piece rather than paper over it. (2) Herbert York — passed, not declined; worth a reading session on The Advisors / Race to Oblivion before it's a draft, at which point the DARPA-founder-turned-critic bridge is a strong standalone.

Status left at drafting. Abstract, insight line, and a symbolic Teniers gambling-tavern hero added per composition Phases 11–12. Not for Cali's gate yet.