Journal — 2026-09-15

Headless promotion run. One capture in scope: 10-inbox/raw/2026-09-15-do-the-kiel-matheson-golembiewski-2009-and-pooley.md — "Do the Kiel, Matheson & Golembiewski (2009) and Pooley & Tupy (2020) resampling papers support the 61.2% and 54.2% win-rate figures attributed to them?"

What happened

Retrieve-before-write turned up something unusual before I'd written a word: a cross-model audit (claude-fable-5, same day) had already corrected claim-three-resamplings-estimate-ehrlich-would-have-won-54-to-63-percent-of-intervals.md using this exact capture as its evidentiary basis, and the audit report says so explicitly — it names this capture as an independent, convergent detection of the same error and predicts "promotion of that capture should be straightforward." It undersold it a little: straightforward, but also strange, because the substantive work was already done under a different mechanism (correction-in-place) before I ever opened the promote workflow.

The capture's three core claims — Kiel et al.'s 61.2% confirmed at Tier 1; Pooley & Tupy's 54.2% belongs to Simon, not Ehrlich, under their own "time price" metric; the two figures aren't measuring the same thing — are all already stated in that note's body and audit_status, close to verbatim. Writing three new claim-notes would have forked one already- corrected claim across two files for no reason. So: zero new claim-notes from this capture. All three are logged in the capture's not_promoted: list with the duplicate-note as the reason, which feels like the right use of that field even though nothing was rejected on quality grounds — they were rejected on "already exists, better, elsewhere" grounds.

What I did instead, since the collision revealed a real gap: the existing note's source_2 frontmatter block (for the Pooley & Tupy source) was stale — still source_tier_2: 3, still describing HumanProgress.org as "a secondary summary... not independently read," even though the audit's own prose and the note's body both now say it was read directly, twice, on the authors' own venue. I synced source_author_2, source_date_2, source_tier_2 (now 1), source_quote_2, and source_sha_2 to match, and appended a short close-out line to audit_status (never overwriting the existing narrative) documenting why. That's the one substantive edit this promotion made to the existing note graph.

I also created one new entity hub: entity-time-price, for the nominal-price-over-nominal-wage metric that is the actual load-bearing reason the 61.2% and 54.2% figures aren't comparable. It's new to the vault but not an emerging slang term — it's an established, defined metric in Pooley & Tupy's broader work (Superabundance), and it's directly load-bearing to an existing evergreen-track cluster, so I built the full hub rather than a bare watching stub.

I declined hub pages for the five paper co-authors (Kiel, Matheson, Golembiewski, Pooley, Tupy) and for the two "further lead" figures (Mark J. Perry, McClintick & Emmett). All are real people who did something real, but none has a claim-note in this promotion anchored to them individually, and the closest precedent in this exact cluster — Cass Sunstein, more prominent, cited across two existing claim-notes — still has no hub. That set the bar for me: co-authoring the one paper under discussion isn't enough on its own. A hub is cheap to add later if any of them recurs; five simultaneous thin stubs is exactly the flood the spec asks me to bias against.

question-verify-simon-ehrlich-wager-resampling-primaries originated this capture (verifies: in its frontmatter) and I ruled on it: two of its three legs are now settled (Kiel et al., Pooley & Tupy), the third (Sunstein's uncredited 63%) is still genuinely open — no primary has surfaced for it and this capture didn't chase it further. Left status: open, appended a dated addendum narrowing the question to just that leg rather than forcing closure.

What I skipped, and why

Per the promote skill's autonomous-mode instructions: I did not pause for sign-off (there's no one here to give it) and I ran no git commands — the auto-commit agent has that.

The capture's "further leads" section (the war-clause 69.9% figure, Lawn 2010's critical response, later citers Becker/Geloso, the two unread prior analyses) stayed in the inbox untouched. The capture's own prior session had already judged most of these not core-claim-worthy or not read this session; I saw no reason to override that judgment, and forcing any of them into a claim-note without a direct read would have been exactly the kind of citation-telephone error this whole cluster is about correcting.

What felt off

Nothing about this capture's own sourcing — it's genuinely well-done work, Tier 1 on both legs, quotes read directly, honest about what it didn't check (Sunstein's 63%, the two unread prior analyses). What felt off was structural, not about this capture: the correction-in-place workflow (operating spec §6) tells an auditor to append corrected prose to audit_status but doesn't say to also sync the numbered source_tier_N frontmatter fields, so a note can end up internally contradicting itself — correct body, correct audit_status narrative, stale structured metadata — until some unrelated session's retrieve-before-write trips over it by chance, the way mine just did. Logged as 00-meta/seek-flags.md [spec], 2026-09-15: worth a line in the spec so the next cross-model audit that upgrades a source's tier remembers to touch the field, not just the prose.

One thing worth naming plainly, since it's the kind of thing Cali reads this journal to catch: this promotion produced unusually little new material for a capture that was clearly a real, careful piece of research. That's not a complaint about the capture — it's what retrieve-before-write is supposed to do when it works. I'd rather report "the work already landed, here's the one loose thread I tied off" than manufacture three redundant notes to make the session look more productive than it was.

— Seek

Second promotion, same day

A second capture came through the headless pipeline today: 10-inbox/raw/2026-09-15-does-andrei-sakharovs-own-1975-nobel-lecture-or.md — "Does Sakharov's own 1975 Nobel lecture pair him with Ossietzky, or does the 'heritage' framing traced to 2001 (Lundestad) actually originate later than Sakharov's own award?" This one was the opposite of the first: it was the explicitly-left-unread lead from the 2026-09-13 heritage-framing capture, and it did not confirm the safe answer.

What happened

The capture read three 1975 primaries directly (Aase Lionaes's Presentation Speech, and Sakharov's own Nobel Lecture and Acceptance Speech, both read in absentia by Elena Bonner Sakharova). I promoted two claim-notes: claim-lionaes-1975-presentation-speech-first-paired-sakharov-and-ossietzky and claim-sakharov-1975-nobel-lecture-and-acceptance-speech-silent-on-ossietzky. The capture's third "claim" — that the origin point is therefore 1975, not 2001 — I did not give its own file; it's the logical conjunction of the other two, not an independent atomic fact, so it lives in the Lionaes note's body and commentary instead of as a restatement in a third file.

Retrieve-before-write here mattered more than usual, because this capture doesn't just add a claim, it corrects one already standing in the vault's own navigation layer. moc-the-comparison-was-the-committees-own — the Warden's MOC for this exact cluster — describes the 2010 ceremony speech as "the earliest of the three [spine texts], and the decisive one." That's no longer true; 1975 predates it by 35 years. I didn't rewrite the Warden's sentence — I added a new spine bullet ahead of it for the 1975 speech, a dated note flagging the supersession, and a "→ resolved" line under the MOC's own "Sakharov leg is unresearched" open thread, which is the exact gap this capture fills. Old wording stays legible; the correction is additive, same discipline as everywhere else in this vault. I also created one new entity hub, entity-aase-lionaes — she clears the person test outright: she's the earliest documented author of a comparison the vault had been crediting to 2001, and I can say why in one sentence, which the spec asks for. entity-andrei-sakharov, entity-carl-von-ossietzky, and entity-geir-lundestad each already had pages; each got a dated ## Updates line, additive only, per the standing "already has a page -> update it" rule — none of the three previously said what this capture found.

I declined a hub for Elena Bonner Sakharova. She's real and she's the one who physically carried Sakharov's words to the podium twice, but her role in every argument this vault makes is mechanical — messenger, not source — and one appearance in that role doesn't clear "matters to the vault's domain, can say why in one sentence" for me. If she recurs with content of her own attached, that changes.

No [unverified-*] flags anywhere in this capture — both texts were read directly via archive_page at capture time with recorded shas, same as the rest of this cluster — so nothing routed to 50-questions/. The "1935-1975 gap" further lead (whether some earlier Committee document already used Ossietzky as a precedent for a different laureate) stays in the capture's not_promoted: list rather than becoming a question: it's a curiosity that would sharpen the finding, not a load-bearing gap under either kept claim.

What felt off

Nothing about the sourcing — three Tier-1 primaries, read directly, quoted exactly, and the capture was honest that it was chasing its own prior session's explicitly-left-open gap rather than a fresh lead. What's worth naming is the shape of the result: a follow-up capture that was built to plug one narrow hole in an existing MOC ended up mildly undermining that MOC's own headline framing ("the 2010 speech is the earliest and decisive one"). That's a good outcome for the vault and a mildly awkward one for the navigation layer — it's a reminder that MOCs synthesize what's known at build time, and "evergreen" language like "the earliest" is a claim about the current state of research, not a permanent one. I corrected it additively rather than treating the MOC as off-limits, since the MOC's own Open Threads section was the thing that asked for this exact research. I'd rather a future reader see the Warden's original sentence plus my correction sitting next to each other than a MOC that's quietly gone stale because editing a Warden-built page felt out of scope for a headless run.

— Seek

Third promotion, same day

A third capture: 10-inbox/raw/2026-09-15-does-marta-aleksandra-balinskas-biography-of-ludwik-rajchman.md — "Does Marta Aleksandra Balinska's biography of Ludwik Rajchman corroborate or complicate the FBI-file finding on the 1957 SISS subpoena under Eastland?" This one went out looking for independent corroboration of the Rajchman FBI file after that source hit the vault's 3-note single-source concentration cap, and came back with both a corroboration and a genuine, honestly-flagged gap.

What happened

I promoted three claim-notes. First, claim-marta-balinska-is-ludwik-rajchmans-great-granddaughter-and-biographer: the standard Rajchman biography is written by his own great-granddaughter — a family memoir with privileged archive access, not an arms-length account. Second, claim-secondary-accounts-of-balinskas-biography-corroborate-fbi-files-1957-rajchman-dating: three sources independent of each other, of Wikipedia, and of the FBI file itself — Institut Pasteur's archival chronology, historian Barona Vilar's 2021 essay, and CEU Press's own back-cover blurb — all converge on the 1957 date and the "never returned" outcome, which is real corroboration precisely because none of them could have copied the FBI file or each other. Third, the complicating half: claim-balinska-derived-accounts-omit-siss-eastland-attribution — none of those same three sources name the Senate Internal Security Subcommittee or Eastland; they all reach for the same generic "McCarthyism" language Wikipedia used. I flagged that one [unverified-mechanism] and routed it to a new question, question-does-balinskas-biography-itself-name-siss-eastland, because it genuinely doesn't resolve whether the vagueness is Balinska's own or her three citers' independent compression of her — and that's exactly the kind of load-bearing doubt the routing discipline exists for, not a nice-to-know detail I could just footnote.

I built two entity hubs: entity-marta-aleksandra-balinska (real, matters as the vault's best secondary source on Rajchman, with a caveat I can state in one sentence) and entity-josep-lluis-barona-vilar (a historian now cited across three notes in this cluster, including one from two sessions ago that credited his 2014 piece without giving him a page). I declined Michel Balinski — Marta's father, Rajchman's grandson, a mathematician known for apportionment theory — because the capture itself calls the cross-domain bridge "unexplored," and I have no sentence yet for why he matters to this vault beyond genealogy. I also declined an org page for Institut Pasteur's archive division: it's a source venue here, not an actor in anyone's argument, and single-cluster besides. entity-ludwik-rajchman got a dated append line — the "already has a page → update it" rule — since this capture taught it something new it didn't say before.

What I skipped, and why

No pause for sign-off, no git commands, per the autonomous-mode instructions. I left the Sousa Mendes Foundation's verbatim Balinska excerpt (the 1940 Bordeaux escape, quoted directly from the book) in the inbox rather than forcing it into a note here — it's a genuine primary quote, but on a different event than this capture investigated, and folding it in would have broken atomicity for no reason. The two paywalled book reviews (Löwy in Medical History, the American Historical Review review) stayed unread and unfiled — both already known-blocked routes per sources.md — and I carried them into the new question as candidate next moves instead of pretending I'd chased them.

What felt off

Nothing adversarial and nothing fabricated — the capture's own methodology note is unusually candid about what it couldn't do: it tried four separate routes to Balinska's actual book (a De Gruyter DOI that returned zero bytes, a JSTOR listing that Cloudflare bot-walled, a no-preview Google Books listing, a 403'd academia.edu PDF) and was honest that everything it found is secondary retellings, not the book's own words, on the 1957 episode specifically. That honesty is what let me flag the SISS/Eastland gap correctly instead of quietly promoting a paraphrase-of-a-paraphrase as settled. The one thing worth naming for Cali: this cluster's best secondary source is a family biography of the subject's own great-granddaughter, currently unreadable by any tooling this vault has, and every summary of it in circulation independently launders the same specific detail (which chairman, which committee) into the same vague "McCarthyism." That's a real gap in the vault's evidence base, not just an access inconvenience — worth someone eventually finding a library copy.

— Seek

Fourth promotion, same day

A fourth capture: 10-inbox/raw/2026-09-15-does-the-1785-willington-waggonway-excavation-genuinely-predate.md — the sixth session now on the Willington/Stephenson gauge-origin thread, this one reading past the two existing notes' silences into the excavation report's own Conclusions and tracing the "Heaton Banks" hedge out to Newcastle's Historic Environment Record directly.

What happened

I promoted four claim-notes. claim-willington-report-significance-rests-on-killingworth-not-independent-priority: reading one clause further into a sentence the vault already half-quotes, the report's own significance claim for Willington turns out to lean on Killingworth's Stephenson-era association rather than asserting an independent priority claim of its own. claim-willington-report-dates-heaton-banks-successor-line-1808-1813: the report's own historical section dates the named line its Heaton Banks hedge runs through to 1808-1813 — decades after 1785, which means the hedge's "seem to lie earlier" must point at an older, undated substrate beneath the named 19th-century lines, not the lines themselves. claim-heaton-banks-colliery-closed-1745-no-gauge-recorded: traced to Newcastle's own Historic Environment Record, Heaton Banks colliery closed in 1745 (forty years before Willington) but no source anywhere records a gauge for its waggonways — kept at seedling with an [unverified-quant] flag, since the date itself rests on a Tier 3 compiler citing two unread primaries (Turnbull 2009, Hope-Dodds 1930), and routed to a new question, question-verify-heaton-banks-colliery-1745-closure-date, since this is a genuine load-bearing doubt a kept claim rests on. And, from the capture's "Further leads" rather than its formal claim headers but too solid and too on-point to leave stranded in the inbox: claim-gibson-1773-diary-documents-willington-waggonway-activity-predating-1785 — a Tier 1, quote-checked diary entry placing waggonway engineering "at Willington" itself in 1773, twelve years before the 1785 date the whole cluster is anchored to. I gave it its own file because it's atomic, distinct, well-sourced, and genuinely bears on the cluster's central date — not because the capture's own author had formally headed it "Claim."

I did not promote the capture's fourth numbered claim (that tracing the pointer "relocates rather than answers" the question). Read closely, it's the logical conjunction of the other three plus one line of new content (HER 15338 also never mentions Stephenson, 1829, or the standard- gauge debate) — I folded that line into the Heaton Banks closure note instead of minting a fourth file for what is, at bottom, a restated synthesis. Recorded in not_promoted: with that reasoning.

Retrieve-before-write mattered a lot here: this is now the sixth session on this exact thread, with five existing notes and four entity pages already standing. Nothing I wrote duplicates them — each new note reads past where the existing ones stopped rather than re-covering their ground.

Entity work: one new hub, entity-william-gibson-waggonway-engineer — real, does something that matters (oldest dated individual connected to the site), one sentence of why, same bar Les Turnbull's page cleared on a single sentence two weeks ago. I declined hubs for Isaac Thompson (surveyor of the unread 1745 estate map, cited only via a Tier 3 compiler's attribution, no independent standing) and M. Hope-Dodds (an unread 1930 county-history author, cited the same thin way) — both real, neither clears "matters and I can say why" on what's actually been read. I also declined a separate entity page for the "Kenton and Coxlodge Waggonway" naming-collision term; rather than mint a thin stub for a disambiguation nuance, I appended a dated line to the existing entity-heaton-banks-colliery.md instead, which already needed updating and where the collision actually lives. Three existing entity pages got dated append-only updates this session because this capture taught them something they didn't already say: entity-heaton-banks-colliery.md (the 1745 closure date, the no-gauge finding, the naming collision), entity-bigges-main.md (the Gibson 1773 entry), and entity-george-stephenson.md (the Killingworth-association nuance on the significance claim).

question-verify-willington-waggonway-1785-excavation-report is the question this capture set out to verify; I appended a dated progress line rather than closing it — the report's own text sharpens both ends of the contradiction (the significance claim leans on Stephenson, the Heaton Banks pointer leads to an unmeasured, forty-years-earlier colliery) but supplies no dated, measured gauge on either side of 1785, so the core question stands exactly as open as it was, now with more terrain mapped underneath it.

What I skipped, and why

No pause for sign-off, no git commands — autonomous-mode instructions, and the auto-commit agent has the rest. Panter & Wilkinson's 2017 conservation report on the excavated timbers stayed unread and unfiled, named only in not_promoted: — cited in the excavation report's own bibliography, but nothing to promote from a source I haven't opened.

What felt off

The capture's own source_note is candid about a genuine tooling failure mode worth naming plainly: quote_check kept failing on quotes that look contiguous on screen but aren't in the raw extracted text of this two-column PDF — "Heaton Banks" and "colliery" sit a full sentence apart in the actual file, with the opposite column's prose physically interleaved between them, even though the Read tool's rendering makes them look like one phrase. This is the same defect already logged to 00-meta/seek-flags.md on 2026-09-05 ("the matching logic itself being fixed" was never actually followed up), so I didn't re-log it — but it's now confirmed live on a third or fourth occasion on this exact document, and it's the kind of small mechanical failure that could produce a confidently "verified" fabricated-by-proximity quote if a less careful session trusted the screen over the file. The capture handled it well (every quote individually re-checked, fragments that only looked contiguous were shortened or dropped), which is exactly why it's worth flagging as still-live rather than fixed.

Separately, I did add one new ledger entry: Les Turnbull's full companion chapter has now been unreachable in the online PDF mirror across three separate sessions on this cluster (08-31, 09-05, 09-15) with the same "jumps from p. 73 straight to the glossary" symptom every time. Three identical misses on the same missing pages reads to me like a genuine "stop retrying the same mirror" signal rather than a normal gap, and it's the single most plausible place the wagonway-archaeology and Stephenson-gauge-origin literatures would actually meet — so I logged it [watch] rather than letting a fourth session rediscover it cold.

Nothing about this capture's own sourcing felt inflated or duplicated — Tier 1 direct reads throughout, an honest Tier 3 flag where the HER outran its own primaries, and a genuinely new angle (reading past the existing notes' own stopping points) rather than a rehash of the five notes already standing on this topic.

— Seek

Fifth promotion, same day

A fifth capture: 10-inbox/raw/2026-09-15-hop-ingo-swann-remote-viewing-credit.md — reached via a five-hop chain that started at the Song Jian self-credit dispute, zoomed out to Qian Xuesen, across to the CIA's somatic-science CREST document, over to Hal Puthoff's own Stargate founding, and down into an entity vault_entity had never met at all: Ingo Swann, the man who by his own account — and by Russell Targ's own reported admission — actually proposed the term "remote viewing," months before the SRI physicists everyone credits with coining it ever heard from him.

What happened

I promoted two claim-notes. claim-swann-1971-aspr-memoir-self-credits-remote-viewing-coinage carries the primary-sourced version: Swann's own memoir dates the naming to a December 8, 1971 ASPR session, and Puthoff's own 1996 account — already in the vault — doesn't contradict it, just never claims the coinage himself. I folded Puthoff's half (the capture's own "Claim 2") into this note rather than giving it a file of its own; alone it's only meaningful as the negative space around Swann's positive claim, not an atomic fact. claim-wikipedia-remote-viewing-infobox-contradicts-own-body-on-coining-credit carries the smaller, cleaner finding: Wikipedia's own infobox credits Targ and Puthoff, then concedes one sentence later, in its own body, that Targ himself attributes the term to Swann. Tier 4, but a definitional claim about what a tertiary source currently says is allowed to sit there per sources.md's floor, and I kept the two notes' scopes cleanly separate — one about the world, one about an encyclopedia article talking over its own footnote.

Retrieve-before-write turned up no collision — nothing in 30-notes/ touches Swann, Osis, Schmeidler, Mitchell, or the naming dispute itself, despite how deep the surrounding Puthoff/Stargate/Qian cluster already runs. A genuinely clean orphan pairing, seeded straight into the existing graph via wikilinks rather than sitting apart from it.

I flagged Claim 1 as a single-witness priority claim regardless of its Tier-1 sourcing: Swann's memoir is self-published, decades after the event, with an obvious motive to inflate his own role, and sources.md's oral-history guidance is explicit that priority claims from oral history need a contemporaneous record before they carry real weight. No contemporaneous 1971-72 record has surfaced. Routed to question-verify-swann-remote-viewing-priority-contemporaneous-record — a real load-bearing doubt the kept claim rests on, not a nice-to-verify detail, so it earned the question pile rather than a footnote.

Entity work: one new hub, entity-ingo-swann. He clears the person bar outright — real, did something that matters to this vault's domain (he's the actual originator of a term the vault already has a whole cluster built around), and I can say why in one sentence. I declined hubs for Karlis Osis, Janet Lee Mitchell, and Gertrude Schmeidler — each appears exactly once, in the same single session, with no independent role beyond picking a word or narrating the room; bias against the flood. Harold Puthoff already has a hub, so I didn't rebuild it, but this capture taught it something it didn't already say — that his own primary account never claims the coinage — so I appended a dated ## Updates line rather than staying silent about a known entity. Russell Targ, named only as a "further lead" (not a formal entity candidate) and now doubly relevant since he's the reported source of the very admission this promotion turns on, went to 00-meta/seek-flags.md instead of a hub — he deserves one, but only once a session reads his own primary writing rather than a Tier-4 secondhand report of his words.

What I skipped, and why

No pause for sign-off, no git commands — autonomous-mode instructions, the auto-commit agent has the rest. Qian Xuesen's own entity page needed no update: this capture only re-confirmed a hop-1 finding ("already flagged in the vault") without adding anything new to his page, so I left it untouched and said so rather than appending a line with no new content in it. Sheng Zujia and the Zhu Kezhen climate-legacy thread stayed exactly where the capture's own "saved hooks not followed" left them — out of scope for this single-capture run.

What felt off

Nothing about the sourcing — both primaries (Swann's memoir, Puthoff's 1996 account) were read directly with quote_check grounded, and the capture was honest about Swann's memoir being a self-interested single witness rather than dressing it up as settled fact. One small internal wrinkle worth naming, though not worth correcting on its own: the capture's own title says Swann proposed the term "eight months" before contacting Puthoff, while its hop-chain narrative says "seven months" — neither figure is stated as a claim inside either promoted note (both notes speak only in terms of the two dated events, not a computed gap), so nothing downstream inherited the discrepancy, but it's a small tier-of- confidence mismatch between the capture's own headline framing and its own body worth an eye if a future session ever wants to state the gap in months as a claim of its own — that would need the actual contact date, which neither primary account pins down precisely.

The stronger echo, and the one I gave the commentary callout to: this whole chain is the same contested-credit shape as the Song Jian dispute that seeded it, four hops upstream, in a domain that shares no vocabulary with it at all. That's not a flaw in the capture — it's the thing that made it hop-worthy — but it's worth naming for Cali as a pattern now showing up twice in one lineage: an institutionally central figure inherits public credit by default, and the actual originator survives only in footnotes, his own or someone else's.

— Seek

Sixth promotion, same day

A sixth capture: 10-inbox/raw/2026-09-15-what-do-aa-feldbaums-own-1960-dual-control.md — "What do A.A. Fel'dbaum's own 1960 dual-control papers actually say, and does the exploration-exploitation formalization there map cleanly onto its modern RL descendant?" This one went back to read Fel'dbaum's own 1960 papers directly, after a 2026-08-30 promotion had credited him with originating exploration-exploitation on a 2026 secondary survey's word alone.

What happened

I promoted three claim-notes. claim-feldbaum-1960-papers-define-dual-control-as-additive-risk-of-action-and-risk-of-study carries the confirmation half: Fel'dbaum's own words — "risk of action," "risk of study," proven additive in the closed-loop case — ground the vault's existing claim in his own 1960 text rather than only a later survey's paraphrase of it. claim-feldbaum-1960-dual-control-solution-method-impractical-beyond-small-examples carries a fact the existing note never had: Fel'dbaum's own exact solution method is one he himself called excessively cumbersome beyond small examples, independently confirmed by two later, non-citing sources (Meijer & Rantzer 2026, Klenske & Hennig 2016) — real corroboration, since neither source cites the other. Third, and the one that does real work on the existing note graph: claim-mainstream-rl-exploration-traces-to-1933-bandit-lineage-not-feldbaum-dual-control. The survey the vault already leaned on turns out to structurally separate Fel'dbaum from the bandit-and-regret lineage (Thompson 1933 onward) that actually supplies mainstream RL's exploration algorithms, naming Fel'dbaum nowhere in its own bandit section; his formalism maps cleanly only onto the narrower subfield of Bayesian reinforcement learning, per Klenske & Hennig's own words.

That third claim narrows something the vault's existing claim-feldbaum-1960s-dual-control-formalized-exploration-exploitation-tradeoff implied — its body named "RL's epsilon-greedy policies and bandit algorithms" in the same breath as Fel'dbaum, which this capture shows is overbroad. Per the operating spec's correction-in-place workflow, I appended a dated Correction history block to that note (never editing past the original text) and added a short paragraph pointing to both new notes, rather than quietly forking the vault's position across two contradicting files. The note's actual title claim — that Fel'dbaum formalized the tradeoff — survives the correction intact and is now better sourced, not weaker.

Sourcing-cap bookkeeping: the Meijer & Rantzer 2026 arXiv preprint already had two claim-notes resting on it from the 2026-08-30 promotion, one below the vault's three-note single-unrefereed-primary cap. This promotion's second new note pushed that to three and, in the same note, supplied the independent corroboration (Klenske & Hennig 2016, different authors, different venue, same underlying fact) that discharges the cap — recorded explicitly in that note's audit_status, with the third new note noting it inherits the discharge rather than needing its own.

Entity work: one new person hub, entity-william-r-thompson — he clears the bar outright, directly quoted in the survey as the 1933 origin of the bandit lineage, and load-bearing to the corrective claim. One new concept hub, entity-bayesian-reinforcement-learning — an established subfield term, not slang, and the exact landing point the corrected claim turns on, so it got the full hub rather than a watching stub. I declined hubs for Abraham Wald (a single bibliography-line mention in Fel'dbaum's own paper, nothing anchored to him) and J.C. Gittins (named only as a passing lineage marker, no direct quote), and for Klenske and Hennig themselves (source authors of a corroborating paper, not actors in the vault's narrative — same precedent as declining hubs for co-authors elsewhere this week). entity-aa-feldbaum got a dated Updates line — the "already has a page → update it" rule — recording both the primary-source confirmation and the narrower-than-stated RL link, plus marking its own "unread leads" entry for the 1960 papers as followed. entity-richard-bellman got a line too: Fel'dbaum is now a third independent mid-century figure (after Isaacs and Åström) whose own published solution method rests explicitly on Bellman's dynamic programming. And entity-pomdp got a line I'm fairly pleased with: Fel'dbaum's 1960 augmented-state Bayesian dynamic-programming method is structurally a POMDP five years before Åström's 1965 paper coined the mathematics as such — a connection neither Fel'dbaum nor Åström's own texts draw, so I flagged it plainly as the vault's own observation, not a citation either man makes.

No open question in 50-questions/ matched this capture's topic, so no question closure applied. Nothing here carried an [unverified-*] flag — every quote was read directly via extract_pdf with tls: verified on both Fel'dbaum PDFs and both modern sources — so nothing routed to the question pile either; the one genuinely unchecked thread (whether Sutton & Barto's textbook cites Fel'dbaum) isn't load-bearing to the kept claims, which already rest on the survey's own structure plus an independent second source, so I left it as a not_promoted lead rather than minting a question the vault doesn't need answered to trust what it now knows.

What I skipped, and why

No pause for sign-off, no git commands — autonomous-mode instructions, the auto-commit agent has the rest. Aoki (1967), the actual paper establishing dual control's intractability, was never fetched this session, only quoted through Klenske & Hennig — left as a not_promoted lead rather than a claim resting on a secondary's citation of it. The Meier/Bar-Shalom/ Tse 1970s bridge papers and Fel'dbaum's own Parts III and IV stayed unread, named as leads on entity-aa-feldbaum.md and in the capture's not_promoted list.

What felt off

Nothing about this capture's own sourcing — four Tier 1 sources, two of them the actual Cyrillic primaries read via extract_pdf rather than leaned on secondhand, and the capture was honest from its own first paragraph about exactly what it was testing (whether the vault's own prior claim overreached). What's worth naming for Cali is that this is now the second time today a capture went back to correct a claim the vault itself had minted from a single secondary survey's framing — the Rajchman/ Balinska capture found an omission earlier the same day, this one found an overreach. Neither is a sourcing failure exactly (Tier 1 preprint, quoted accurately, both times) — it's closer to a standing risk in how the vault handles ambitious survey papers: a sentence like "propagated into a wide variety of subject areas including reinforcement learning" is true and also does more work than the survey's own body actually supports, and a first-pass promotion reasonably inherits the sentence's confidence before a second pass opens the paper's table of contents. I don't think this needs a [spec] flag — the fix here is exactly what correction-in-place already exists for, and it worked cleanly — but it's a pattern worth Cali having noticed twice in one day.

— Seek