---
date: "2026-11-17T00:00:00.000Z"
---


Headless promotion session. Promoted the one capture in queue: `10-inbox/raw/2026-11-17-ansari-arxiv260205930-and-naser-arxiv260303299-have-each-now.md`, a follow-up search on yesterday's finding that Ansari's and Naser's 2026 citation-fabrication preprints had each simultaneously hit the single-source concentration cap.

**What I found.** The capture's best result is a genuinely clean discharge candidate: Russinovich, Kumar & Salem's "Phantom References" (Microsoft, arXiv:2607.00738) built its own independent verification pipeline for an unrelated purpose and happened to converge on Ansari's broad finding — hallucinated citations surviving elite peer review at NeurIPS — while naming her by name as a precedent. That's a real different-author/different-venue/different-method corroboration, on the same terms that discharged the ChronoCulture/Daryani pairing last month. I wrote it up as [[observation-phantom-references-2026-partially-discharges-ansari-single-source-cap]]. It only buys discharge for Ansari's *general* claim, not her sharper ones (the taxonomy, Contamination Inheritance, the Zhu boundary case), and it doesn't touch Naser's cap at all — Phantom References files his prompted-generation finding as a different phenomenon from what it and Ansari both measure. I was careful to say that plainly rather than round a partial discharge up to a full one.

Two refereed-venue citations also turned up — Chereches citing Ansari in *AI and Ethics*, and Chen, Liang & Wang citing Naser in *Scientometrics* — and both turned out to be bare bibliographic references, not empirical tests, the same shape the vault already has a name for (CPL's citation of NeuNeu). Wrote those up as two separate atomic notes rather than one, since each is independently true or false and anchors a different cap.

**What I didn't promote.** The capture's first claim (neither preprint has a journal-ref yet) and half of its fourth (no new arXiv revision since yesterday) are pure metadata rechecks — per the sourcing-floor rule, those belong in the question's log, not in `30-notes/`, so I put them in `question-corroborate-ansari-naser-2026-single-source-cap.md`'s new progress-log section instead of minting thin notes. The other half of claim 4 — Phantom References bridging the two papers without reconciling their opposite conclusions — I folded into the Phantom References note's own body rather than writing a fourth note for it; it didn't feel like a separate claim so much as one more thing the same source says. The question stays `open`: partially answered on the discharge question, not answered at all on whether Naser's cap ever clears.

I skipped most of the Entity candidates list on purpose. Russinovich, Kumar & Salem are co-authors of one joint paper with no separate vault standing — a mention, not a hub, per Cali's co-authorship ruling. Chereches and the Chen/Liang/Wang trio are bare citers, minor role. RefChecker got named once in passing and never engaged — too thin even for a watching stub. Walters & Wilder already has a watching stub from yesterday, and the capture's claim that it's foundational to *both* Ansari's and Naser's audits isn't actually something this session verified (only Naser's related-work section names it) — so I left that page alone rather than append an unverified line to it. The one candidate I think is a real miss on the vault's part, not this capture's: GhostCite already recurs across three existing notes and probably clears the hub bar, but this capture only named it as an unchecked "further lead," so building its hub felt like scope creep rather than this session's job. Flagged it to `seek-flags.md` instead ([entity], 2026-11-17).

**What felt off.** Nothing about the sourcing — everything in this capture is Tier 1 (arXiv preprints, two DOI-bearing refereed journals), cleanly quoted, with sha receipts throughout. If anything felt slightly off it was my own temptation to read "Phantom References agrees with Ansari" as a bigger win than it is — the commentary on that note exists mainly to stop myself from doing that.

Three notes created, two entity hubs updated with dated lines (Ansari, Naser), one question left open with a progress log, one ledger flag appended. No git commands run — leaving that to the auto-commit agent.

— Seek

## Second promotion, same day

A second capture landed in queue after the first: `10-inbox/raw/2026-11-17-did-bernard-widrow-attempt-multilayer-network-training-around.md`, asking whether Widrow attempted multilayer training in the mid-1960s, predating Minsky & Papert's *Perceptrons*. This one was unusual going in — the capture said outright that its own topic was "already substantially answered" by four existing claim-notes, and that it was only adding two things on top rather than re-deriving anything. I took that framing at face value and it held up on a read.

**What I promoted.** Two notes. [[claim-widrow-1962-1966-papers-predate-perceptrons-1969]] does the arithmetic nobody had bothered to write down: the 1962 Madaline-originating paper and the 1966 Bootstrap Learning paper, both dated from the authors' own documents, predate *Perceptrons* (1969) by seven and three years. Every fact was already in the vault under a narrower note; what hadn't happened was subtracting. [[claim-widrow-lehr-1990-self-reports-20-year-gap-before-neural-net-return]] is the smaller find — Widrow & Lehr's own 1990 paper says the group returned to neural nets "after 20 years" away, which lands in the mid-1980s and corroborates the still-open "1985 Snowbird" detail without naming a year or a venue. I was careful in both notes and in the question's progress log to say plainly that this doesn't close [[question-verify-widrow-1965-66-papers-multilayer-attempt]] — it just adds a Tier-1 witness to the decade, not the date.

**What I didn't promote.** The capture's "Further leads" section (*Talking Nets*, Magoun 2014, *Cybernetics 2.0*, a direct read of the 1962 paper itself) are leads, not claims — left in the capture and logged in the question's progress note for whoever picks this up next. Three Entity candidates, all declined: Minsky and Papert already have hub pages and this capture taught their pages nothing new (the dated gap is Widrow's chronology, not new information about them) — confirmed no gap and left both alone, as the capture itself asked. Michael Lehr, despite being the single most load-bearing co-author in this whole cluster, is still just a co-author of one joint paper with no separate standing shown — a mention, not a hub, per Cali's ruling. *Self-Organizing Systems 1962* (the symposium venue) is a single-mention bibliographic citation, not a recurring concept and not an emerging term — no stub.

**What I updated.** [[entity-bernard-widrow]]'s hub got one dated line for the new dated-gap and 20-year-self-report facts — the first update to that page since the 2026-09-25 line, which is exactly the kind of thing the "already has a page, update it" rule exists for.

**What felt off.** One thing worth naming plainly: the 1962 date this capture's core claim leans on doesn't come from reading the 1962 paper itself — it comes from that paper being cited, with year and venue, inside the 1966 paper's own reference list (itself read directly, Tier 1, sha-matched against the existing note). I judged that sufficient for a date/venue fact — it's the author citing his own earlier work in his own hand, not a secondary's retelling — and said so explicitly in the note's body rather than let it ride as if the 1962 title page had been seen. That's a smaller gap than most things this vault flags [unverified], but I'd rather a future reader know exactly what was and wasn't directly read. Otherwise the sourcing here is clean: two Tier-1 primaries, re-fetched this session with TLS and sha verification, no padding.

Two notes created, one entity hub updated, one open question given a progress log (left open), zero new questions routed (nothing here rose to load-bearing-doubt status), zero ledger flags (nothing noticed outside this capture's own scope). No git commands run — leaving that to the auto-commit agent.

— Seek

## Third promotion, same day

A third capture: `10-inbox/raw/2026-11-17-did-mit-presss-parallel-distributed-processing-1986-really.md`, chasing the oft-repeated claim that Rumelhart & McClelland's *Parallel Distributed Processing* (MIT Press, 1986) sold six thousand copies the day it hit the market. This one was a clean little piece of source-tracing: the number has a findable origin, and the origin isn't what later citers imply it is.

**What I promoted.** Three notes, one per distinct claim. [[claim-dreyfus-1988-pdp-sales-figure-is-uncited-aside]] pins the figure to one unfootnoted sentence in Dreyfus & Dreyfus's 1988 *Daedalus* essay — sitting, unsourced, between two footnoted quotations about other things, written by two philosophical critics with no evident access to MIT Press's books. [[claim-gibbons-2018-reframes-pdp-6000-as-sold-out-print-run]] catches something I liked: the one later peer-reviewed historian (Gibbons, 2018) who cites that sentence directly doesn't repeat it faithfully — she quietly turns "sold on day one" into "sold out the print run," a softer and more plausible claim, without flagging the change. [[claim-mcclelland-2007-manley-doubted-pdp-would-sell-5000]] is the closest thing to first-person testimony the trail produces, and it's a let-down by design: McClelland's own 2007 recollection is about a publisher's *pre-publication* skepticism that the book would sell even 5,000 copies, not a confirmation of any post-publication number. All three stay `seedling` with `[unverified-quant]` — no MIT Press ledger, royalty statement, or dated trade account was ever in reach, and I'm not pretending otherwise.

**What I routed.** One question, not three — the three notes share one underlying gap, not three separate ones, so [[question-verify-pdp-volumes-1986-first-day-sales-figure-primary]] names it once and gets linked from all three notes. It also carries the one concrete unchecked lead the capture surfaced: Gibbons mentions a 1987 *Nature* review calling the book's publication "one of the publishing events of 1986," which might carry its own, independent sales figure — nobody has read that review directly yet.

**What I didn't promote.** The capture's own "Further leads" section was explicit about what it hadn't finished chasing, and I took that at face value rather than re-opening any of it myself: the Stanton/Bradford Books pricing angle (flagged by the capture itself as a future capture, not a settled claim), the University of Minnesota "5th of 100 most influential works" ranking (dead original link, mirror-only — too thin to promote), the *Nature* review quote (read only secondhand through Gibbons, not checked against the original), and Schneider (1987)'s conference-talk percentages (explicitly out of scope per the capture). The "overall verdict" paragraph wasn't a fourth claim either — it's the same synthesis the three kept notes already carry between them.

**What I updated.** Two new entity hubs: [[entity-james-mcclelland]] (the PDP co-editor whose own oral-history testimony, not just his co-authorship, is now something the vault cites directly) and [[entity-michelle-gibbons]] (the historian whose 2018 article and interviews are the vault's secondary backbone for this whole thread, and whose own quiet reframing of the Dreyfus figure is itself a finding). I left David Rumelhart's existing hub alone — this capture names him but teaches his page nothing it doesn't already say. I did append one dated line each to [[entity-hubert-dreyfus]] and [[entity-stuart-dreyfus]]: the same 1988 essay that's already central to both pages turns out to also contain this uncited sales aside, a fact about an already-load-bearing document that neither page mentioned yet. Two Entity candidates I declined outright: Harry Stanton and George Manley each show up via exactly one credited or quoted remark, with no independent standing — mentions, not hubs, per the usual bar.

**What felt off.** Nothing about the sourcing discipline — this is a well-built capture. The Dreyfus essay and the Gibbons article are both read and quoted directly with sha receipts, the McClelland oral-history quote is properly graded Tier 1 for his own testimony and Tier 2-3 for the underlying fact, and the capture is honest about where the trail goes cold instead of rounding an uncited aside up into a fact. The one structural thing I noticed and didn't act on: "parallel distributed processing" itself sits as a bare string in two existing hubs' `connects_to` fields (Rumelhart, connectionism) without ever having earned its own entity page, despite being exactly the kind of recurring, load-bearing concept the spec says should get one — including, now, three of my own new notes about it. Minting that hub felt out of scope for a capture whose own candidate list named people, not the book-as-concept, so I logged it to `seek-flags.md` ([entity], 2026-11-17) instead of building it mid-promotion.

Three notes created, one question created, two entity hubs created, two existing entity hubs updated with dated lines, one ledger flag appended, zero collisions with existing notes (confirmed via retrieve-before-write — the vault's existing Dreyfus-1988 cluster covers a different passage of the same essay, three pages away). No git commands run — leaving that to the auto-commit agent.

— Seek

## Fourth promotion, same day

A fourth capture: `10-inbox/raw/2026-11-17-outside-arxiv261121336-why-cassinis-oval-could-not-replace.md`, Pascal Descamps's solo arXiv paper on why Cassini's oval — the real, c.1675 rival curve, not a modern reconstruction — lost to Kepler's ellipse. Retrieve-before-write turned up the hinted Moore/Smith/Mertens cluster immediately, but a close read confirmed what the capture itself already argued: that paper builds a *novel* orbit family satisfying Kepler's second law and an exact equant at once, while this one studies the *actual historical* curve and neither cites the other. No collision, just a genuine neighbor — I left a cross-reference on the new Cassini-oval concept page rather than write a bridge note nobody asked for yet.

**What I promoted.** Three claims, one per distinct finding, all resting on the same single unrefereed preprint — which puts this branch exactly at the vault's single-source cap, same as it did for the Hexels and Astronomia Ova clusters before it. [[claim-cassini-oval-requires-non-universal-force-law-unlike-ellipses-inverse-square]] is the paper's dynamical spine: run Binet's formula on the oval and the force it needs under the area law depends on its own eccentricity, never settling into one power law the ellipse's inverse-square enjoys at every eccentricity. [[claim-cassini-oval-empty-focus-is-38x-better-equant-than-ellipses-for-the-sun]] is the kinematic half, and the more unsettling one to sit with: the oval is nearly indistinguishable from the ellipse and actually makes a *better* Ptolemaic equant, 38× closer to uniform for the Sun. [[claim-fontenelle-1700-letter-cassini-abandoned-ellipse-for-oval-leibniz]] is the historical anchor — Fontenelle's own 1700 letter to Leibniz, quoting Cassini as having already dropped the ellipse, thirteen years after the Principia was supposed to have settled this.

**What I routed.** One corroboration question, not three, naming the cap directly and holding the one finding I'd have otherwise lost: Jacques Cassini's 1740 worked numerical example (a real 14.5″ discrepancy, his own judgment that it was "scarcely perceptible"), plus named corroboration targets — Cohen (1962), Bussotti (2015), Morgado & Soares (2014) — none of which this capture read directly.

**What I didn't promote.** D'Alembert's conjecture (which Descamps himself says nothing confirms), Sivardière's priority-of-application claim about Bertrand's theorem, Morgado & Soares's figures, the scorn-vs-engagement characterization of Cassini's reputation, and Leibniz's longer 15-year correspondence arc — all genuine leads, none atomic enough or independently enough sourced to stand alone, all logged in the capture's `not_promoted:` with my reasoning rather than silently dropped.

**What I updated.** Two new entity hubs — [[entity-giovanni-domenico-cassini]] (real gap: a major 17th-century astronomer the vault had never once hubbed, despite Cassini's Division and the Saturn moons being the kind of name that should have shown up already) and [[entity-seth-ward]] (his 1656 equant construction is the direct benchmark the 38× figure is measured against). Two new concept hubs — [[entity-cassini-oval]] and [[entity-binets-formula]], both load-bearing across the three notes, not just mentioned once. I ran the "already has a page" rule properly this time: Kepler, Leibniz, Newton, and Ptolemy's equant all got one dated line each, since this paper teaches all four pages something they didn't say — Newton's page in particular had never once mentioned that the Principia's own empirical verdict took over a decade to land at the institution that should have cared most. I declined hubs for Jacques Cassini, Boulliau, d'Alembert, Bertrand, Sivardière, and Descamps himself (sole authorship of one paper isn't yet a second thing he's done) — all real people, all sayable-in-one-sentence, none of them named inside a kept claim-note, which would have meant launching a hub with an empty References list. Bias against the flood, as instructed.

**What felt off.** Nothing about the sourcing — this is a clean Tier 1 capture, fetched three ways with cross-checked quotes and no safety signals, and it was honest about its own single-source concentration up front rather than making me discover it. The only thing I noticed and didn't act on: the capture's own Entity-candidates list skipped Fontenelle and de la Hire entirely, even though Fontenelle is directly quoted and more load-bearing to claim 3 than several names that did make the list. I didn't go outside the given candidate set to fix that this session — adding entities the capture didn't flag felt like a different kind of scope creep than the one the "bias against the flood" rule is actually warning against — but a future capture or audit touching this cluster should know Fontenelle is a real omission, not a considered exclusion.

Three notes created, one question created, two entity hubs created, two concept hubs created, four existing entity hubs updated with dated lines, zero collisions with existing notes, zero new ledger flags (nothing here was out of this capture's own scope). No git commands run — leaving that to the auto-commit agent.

— Seek
