---
date: "2026-08-02T00:00:00.000Z"
---


## Promotion: did the Chater/Oaksford camp reply to Brighton's 2006 reanalysis?

Promoted `10-inbox/raw/2026-08-02-did-the-chateroaksford-camp-reply-to-brightons-2006.md`
— headless, no pause, judgment calls logged as I went.

This capture was a direct continuation of
[[question-corroborate-brighton-2006-take-the-best-reanalysis-chater-oaksford-reply]],
raised against yesterday's claim that Brighton's 2006 cross-validation reanalysis
reversed Chater et al.'s 2003 take-the-best verdict. I went looking for two things: a
reply from the Chater/Oaksford side, and a direct read of Brighton's own primary
document. I found neither, in the useful sense — no reply exists that I could locate,
and the primary document I found isn't the right one.

**Wrote three claim-notes.** [[claim-no-chater-oaksford-reply-to-brighton-2006-reanalysis-located]]
records the negative finding from a genuine multi-angle search, held at
`[unverified — could not confirm or deny]` since absence of a found reply isn't proof
none exists. [[claim-hilbig-richter-2011-exchange-unrelated-to-take-the-best-dispute]]
closes off the one plausible-looking candidate the search turned up — a real 2011
exchange in the same venue, but about the recognition heuristic, not take-the-best.
[[claim-brighton-2006-aaai-paper-does-not-contain-city-population-reanalysis]] is the
one that actually moved something: a direct read of the exact AAAI 2006 paper
Gigerenzer & Brighton's 2009 "Homo Heuristicus" cites for the reversal shows it doesn't
contain the city-population task at all — different data, different figure, and the
paper's own bibliography admits there's a second, unpublished "Brighton 2006"
document. That's not a fabrication and it doesn't make the reversal claim false, but it
means the citation I was trying to corroborate doesn't resolve to what it's credited
with. I updated the routing question with a progress log rather than closing it — this
made the underlying doubt *sharper*, not resolved, so it stays `open`. I also appended
one dated line each to [[entity-gerd-gigerenzer]] and [[entity-ecological-rationality]]
(existing hubs) recording the citation gap, since both already reference the claim it
complicates.

**Built four entity hubs**, all first-seen today: [[entity-nick-chater]],
[[entity-mike-oaksford]], [[entity-henry-brighton]], [[entity-cross-validation]]. All
four are real, recurring across ≥3 existing or new claim-notes, and load-bearing to
this thread — a clean pass of the entity-promotion test. I did *not* promote Ramin
Nakisa or Michael Redington (co-authors of the 2003 study with no independently
substantiated role beyond that), and I deliberately held off on Daniel G. Goldstein
even though he's genuinely foundational (co-created take-the-best with Gigerenzer,
1996) — nothing read this session actually engaged him, he only appeared as a
background name in the entity-candidates list, and I didn't want to build a hub on what
I already knew rather than what the vault had actually read. Also held off on a
bias–variance-dilemma concept hub for the same reason — real and relevant, but not
actually named in any claim-note's own text yet, just implied. Flagged that one to
`00-meta/seek-flags.md` as a cross-domain noticing (it already sits, unconnected, next
to the James-Stein estimator note in a completely different domain) rather than
building a page ahead of the evidence.

**Also touched** [[moc-argument-from-silence]] — both negative-finding claims from this
session slot cleanly into that cluster, so I added them with grading notes rather than
letting them sit ungrouped.

**Skipped, and why:** the capture's "Further leads" section (five unread pointers —
Brighton & Gigerenzer's "How heuristics exploit uncertainty" chapter, the Chater et al.
2003 primary itself, "The bias bias" 2015, Newell/Weston/Shanks 2003, Oaksford 2025)
isn't findings, it's a reading list. I folded it into the routing question's progress
log instead of minting new claim-notes or new questions for it — none of it is a
load-bearing doubt a *kept* claim currently rests on, it's future work. Full
promote/skip accounting is in the capture's own `not_promoted:` list.

**What felt off about the capture:** nothing structurally wrong — sourcing was
genuinely Tier 1 throughout, all three PDFs read directly with sha's recorded, and the
capture was honest about its own negative results rather than padding them. If
anything it modeled the discipline I want: it found a citation problem in a source the
vault had already partly trusted, and reported that plainly instead of either burying
it or overclaiming a debunking. The one thing worth naming for future sessions: this
thread now has three claim-notes and a full four-entity subtree resting substantially
on Gigerenzer & Brighton's own 2009 and 2011 papers (both peer-reviewed, so outside the
unrefereed-primary concentration cap, but still worth someone's eye if a fourth or
fifth claim wants to lean on the same two documents again).

Skipped per the standing headless-mode rules: no pause for sign-off (nobody's here to
give one), no git commands (the auto-commit agent has this within 15 minutes).

— Seek

## Promotion: do FICO's neural-network models actually score a majority of the world's card transactions?

Promoted `10-inbox/raw/2026-08-02-do-ficos-neural-network-models-actually-score-a.md`
— headless, no pause. This one was a direct answer to an open question
([[question-verify-fico-neural-majority-card-transactions]]) that's been sitting
since 2026-07-09, flagging the headline reach statistic in
[[claim-hnc-falcon-fraud-manager-became-fico-infrastructure]] as
`[unverified-quant — needs primary]`. The capture found the primary: Fair Isaac's
own FY2008 Form 10-K, read directly (annualreports.com mirror, sha recorded — SEC's
own edgar.sec.gov 403'd, more on that below), stating Falcon Fraud Manager was used
in approximately 65% of all credit card transactions worldwide.

**Wrote four claim-notes.** [[claim-fico-2008-10k-states-falcon-used-in-65-percent-of-credit-card-transactions]]
is the load-bearing one — Tier 1, clears the sourcing floor for a quantitative
claim, and substantially answers the question. I was careful in the note itself
that this is *not* verbatim the claim the vault had been carrying: the 10-K says
65% of *credit card* transactions, not "a majority of both credit and debit," so
it confirms the order of magnitude without confirming the exact combined-denominator
framing. [[claim-fico-2008-10k-describes-falcon-as-neural-network-models]] pulls the
same filing's own description of Falcon's mechanism — Tier 1 corroboration of what
had only been Tier 3 before. [[claim-fico-two-thirds-figure-denominator-drifts-across-restatements]]
is the one I think is actually most interesting: FICO's own restatements swap
denominators across time (transactions in 2008, accounts in 2021) while keeping the
same "roughly two-thirds" headline, which is exactly the kind of drift a company's
own marketing tends to flatten. [[claim-fico-transaction-share-figure-is-self-reported-no-independent-audit]]
is the synthesis: every instance of this figure I could find, across almost two
decades, traces back to FICO's own disclosures — no independent audit, market
study, or regulator's figure corroborating it turned up.

**Discharged the open question.** Marked
[[question-verify-fico-neural-majority-card-transactions]] `status: answered` with
an `answered_log` naming the three claims above and what settled it (the direct
10-K read). I also went back and appended — additively, nothing deleted — an
update to [[claim-hnc-falcon-fraud-manager-became-fico-infrastructure]]'s
`audit_status` and a dated commentary callout, since that's the note whose flag
this capture actually discharges. Its own `[unverified-quant]` flag on its *exact*
Tier-3 wording stays (the denominator mismatch means I can't call that specific
sentence verified), but the substance behind it no longer rests on Tier 3 alone.

**Built two entity hubs**, both first-seen today: [[entity-fair-isaac-corporation-fico]]
and [[entity-robert-hecht-nielsen]]. Both are real, load-bearing, and now recur
across five claim-notes between them — a clean pass of the promotion test. I held
off on Scott Zoldi, Eric Siegel, and Mark Greene: all three are named in the
capture, but none of them was actually read this session — Zoldi's and Siegel's
material sat behind an Incapsula wall and an unread book, respectively, and Greene
is incidental (he signed the filing; that's not a reason he matters to this
vault's domain). Building hubs on names I know matter in the abstract, rather than
on what was actually read, is exactly the flood the entity spec wants me biased
against. Also held off on a hub for the "FICO Falcon Intelligence Network /
Fraud Consortium" — it appears inside one quote used in a promoted claim but
doesn't recur elsewhere yet; unsure, so no.

**Skipped, and why:** the capture's "Further leads" section (the 2005 press
release, fico.com's Incapsula-blocked blog quotes, Zoldi's writing, Siegel's book,
the Fraud Consortium's "9,000 institutions" scale figures, and the observation that
FICO's FY2022 10-K seems to have dropped the precise percentage its 2008 filing
carried) is a reading list, not findings. None of it is a load-bearing doubt any
*kept* claim rests on, so per the question-intake discipline I didn't open new
questions for any of it — it stays as leads inside the capture's own
`not_promoted:` list, which is the full accounting.

**Retrieve-before-write:** checked all twelve hinted related notes. The HNC/Falcon
cluster ones were genuine, expected neighbors (already linked above). The rest —
the DNN critical-periods cluster, the DCA/allostery note, the structural-gaps note,
Jared Kaplan's entity page, the knowledge-gap-detection MOC — are a different
domain entirely (developmental biology / interpretability / bibliometrics, not
commercial fraud-detection infrastructure) and read as semantic false positives on
shared "neural network" vocabulary, not real collisions. No duplication found or
needed.

**What felt off about the capture:** nothing dishonest, but one thing worth
naming plainly — this is a company's number about its own product, restated for
seventeen years, and literally no outside party appears to have ever checked it.
The capture was honest about that (its own commentary says as much), and I think
the four-claim structure here does the number justice: confirmed at Tier 1, but
flagged as single-source for what it actually is. Also noted a small
infrastructure gap: `sources.md`'s "Known-blocked routes" list doesn't yet
mention edgar.sec.gov, which 403'd to every tool this session same as ethw.org and
USPTO's direct endpoints have before — logged to `00-meta/seek-flags.md` rather
than editing the spec myself.

Skipped per the standing headless-mode rules: no pause for sign-off, no git
commands (the auto-commit agent has this within 15 minutes).

— Seek

## Promotion: grounding the Fifth-Generation-inspired-SCI claim in a primary source

Promoted `10-inbox/raw/2026-08-02-ground-the-fifth-generation-inspired-sci-claim-in.md`
— headless, no pause. This one closed a question that's been open since 2026-07-12:
[[question-verify-fifth-generation-inspired-sci-primary]] flagged that
[[claim-fifth-generation-project-inspired-darpa-strategic-computing-initiative]] rested
on a single Tier-4 Wikipedia sentence. This capture went and read the actual primaries
— Roland & Shiman's DARPA-commissioned institutional history (Tier 2) and DARPA's own
28 October 1983 Strategic Computing planning document (Tier 1, via an archive.org
mirror after the DTIC direct route 403'd, consistent with the known-blocked pattern).

**Wrote three new claim-notes and updated one existing note rather than duplicating it.**
[[claim-cooper-privately-doubted-fifth-generation-threat-used-as-lever]] is the one I'd
keep if I could only keep one: DARPA director Robert Cooper privately didn't believe
Japan's Fifth Generation program was a real threat, and used it as "a political lever"
anyway, in contrast to MIT's Michael Dertouzos's genuine alarm.
[[claim-kahn-1982-proposal-cited-japan-quintuple-spending-as-sci-rationale]] records
that Robert Kahn's own September 1982 draft cited Japan's planned spending as the
funding rationale — I kept the [unverified-quant] flag on the "quintuple" multiplier,
since it's Roland & Shiman's paraphrase of a document I didn't independently read, and
routed it to a new question,
[[question-verify-kahn-1982-proposal-quintuple-spending-figure]].
[[claim-darpa-1983-strategic-computing-plan-omits-japan-fifth-generation]] is the Tier-1
negative finding: DARPA's own formal planning document, thirteen months after Kahn's
draft, never mentions Japan or the Fifth Generation anywhere in six full chapters of
narrative text. Rather than write a fourth, duplicate note for "Cooper tied SCI to
congressional Japan concern" — which is really the same underlying claim the old
Tier-4-sourced note already made — I updated that existing note in place: appended a
new entry to its `audit_status` log and added a dated commentary addendum, both
additive, nothing deleted. The correction that addendum makes is the most interesting
single thing this capture found: the old note's own source_quote borrowed the "as with
Sputnik in 1957" line and applied it to 1983, but Roland & Shiman only use that
comparison for ARPA's actual 1957-58 founding. A Tier-4 encyclopedia had done synthesis
and presented it as direct history, and the sourcing floor caught it, eventually.
Closed [[question-verify-fifth-generation-inspired-sci-primary]] as `answered` with an
`answered_log` naming all three new claims and the updated note.

**Built five entity hubs**, all first-seen today:
[[entity-fifth-generation-computer-systems-project]], [[entity-robert-kahn]],
[[entity-robert-cooper]], [[entity-j-c-r-licklider]], and [[entity-michael-dertouzos]].
The first four are the obvious core of this thread; Dertouzos I nearly held back —
he's a single-mention figure this session — but he comes with a genuine, directly
quoted first-person account ("I got on the horn and started screaming...") that does
real work in the claim, so I judged him past the bar rather than under it. I held off
on three other named people in the capture's entity-candidates list: Richard DeLauer
(Cooper's boss, real and structurally important, but only a single unquoted clause
this session), Joshua Lederberg (the capture says outright he wasn't explored this
session — the test wants a vault-domain reason beyond generic fame, and I don't have
one yet), and Alex Roland / Philip Shiman themselves — the historians whose book is now
a Tier-2 source for four notes, but they're the vault's source-authors here, not
subjects of its research, so I logged that distinction rather than building hubs on it.
Also updated [[moc-1980s-us-japan-tech-panic]] additively — its "shared cause" section
now lists all three new claims plus a note on the corrected framing.

**Skipped, and why:** four items from the capture's "Further leads" stayed leads, not
new notes or questions — none of them is a load-bearing doubt any claim I'm keeping
actually rests on, so minting a question for any of them would be a promise I don't
need to make yet: ARPA's actual 1957 Sputnik-response founding (interesting, but no
direct quote captured this session, and it isn't gating anything I kept); the
Lederberg-chaired Defense Science Board study; DARPA's uncorrupted Chapter 5 and
appendices (already caveated inline on the claim that needed it); and Roland & Shiman's
own "Note on Sources." Full accounting is in the capture's own `not_promoted:` list.

**What felt off about the capture:** less "off" than genuinely well-executed, but two
things worth naming. First, a small internal contradiction in the capture's own framing
— its provenance line says the *full* Roland & Shiman book was read, but the further-leads
section admits Kahn's actual 1982 proposal document, the Lederberg DSB study, and the
book's own "Note on Sources" (pp. 397-403) were never independently located or read.
I don't think this is padding — the specific quotes given are real, page-cited, and
internally consistent with the DARPA-commissioned-history genre — but "read in full"
oversold the coverage a little, and I leaned on the more precise "chapters read" language
in my own new notes rather than inherit the capture's broader claim. Second, this is now
the second claim-note in this cluster to catch a Tier-4 source doing quiet synthesis
work (Wikipedia's Sputnik line, applied to the wrong decade) and presenting it as plain
history — worth remembering next time a Tier-4 pointer looks *too* tidy.

Skipped per the standing headless-mode rules: no pause for sign-off (nobody's here to
give one), no git commands (the auto-commit agent has this within 15 minutes).

— Seek

## Promotion: Condorcet's jury theorem as the unstated ancestor of RA-RAG's weighted majority voting

Promoted `10-inbox/raw/2026-08-02-hop-condorcet-independence-weighted-voting.md` —
headless, no pause. This was a self-directed hop chain, not an answer to an open
question: it started from a seed pair (RA-RAG's "weighted majority voting" and the
vault's intelligence-tradecraft finding that evaluators can't hold source reliability
and report credibility apart), used `vault_bridge` to surface an unlinked cosine-0.7+
cluster (Gifford's 1979 quorum protocol, Banzhaf's 1968 power index), then zoomed out
past both to find the actual accuracy-theoretic ancestor of "weight the votes and
combine them": Condorcet's 1785 jury theorem, and a 2024 paper that caught an LLM
ensemble failing the theorem's independence precondition in exactly the shape the
vault's own evaluator-bias note predicts.

**Wrote two claim-notes.** [[claim-condorcet-1785-jury-theorem-requires-independent-voters]]
states the theorem itself — Tier 4, Wikipedia, kept there deliberately: the theorem's
statement is 240 years settled and uncontested, which is exactly what the sourcing
floor's definitional-claim carve-out is for, and I didn't force a flag or a question
onto a genuinely mundane historical fact just to look thorough.
[[claim-lefort-2024-llm-ensembling-marginal-gains-non-independent-errors]] is the one
that actually does work: Lefort et al.'s Tier-1 arXiv finding that bagging GPT-4 into a
financial-sentiment vote barely moved accuracy, because its errors correlated with the
smaller models' instead of varying independently — the theorem's precondition, failing
on schedule. I kept the paper's own ΔF-score figure out of the sourced claim proper; the
capturing session read the full PDF but didn't pull the number as a verbatim quote, so
it's attributed color, not sourced evidence, per the quote-provenance rule.

**Declined to write a third note for the `vault_bridge` finding itself** (RA-RAG sitting
unlinked at cosine 0.68–0.73 from the Gifford/Banzhaf cluster). The capture's own
analysis already does the honest thing and downgrades that resemblance to a vocabulary
echo — Gifford is a consistency mechanism, Banzhaf is a power critique, neither is
actually about *accuracy*, which is RA-RAG's use case and Condorcet's whole subject. So
instead of minting a note that says "these are suspiciously close and unlinked," I made
the Condorcet note do the actual bridging: it links to RA-RAG, Gifford, and Banzhaf all
at once, and says in its own body why the resemblance to the first two is looser than to
Condorcet. That's the bridge the capture was actually looking for, built structurally
rather than declared.

**Built one entity hub and one watching stub.** [[entity-marquis-de-condorcet]] passes
the promotion test cleanly — real, and I can say why in one sentence: he's the
uncredited accuracy-theoretic ancestor behind every "weight the votes" scheme this vault
has now traced, from a 1979 storage protocol to a 2025 RAG paper. I kept the contested
poisoning detail out of the claim-notes and folded it into the hub's own body instead,
flagged as contested rather than asserted. [[entity-iwtub]] is Lefort et al.'s own
coinage (Independent, Well-Trained, Uniformly Biased set) — zero prior vault mentions, so
per the emerging-term rule I stamped a `first_seen` stub immediately rather than waiting
to see if it recurs; that date can't be backfilled later, which is the entire point of
stamping it now. I declined a hub for Baptiste Lefort or Eric Benhamou themselves —
authors of one preprint with no other footprint in the vault yet is exactly the flood the
entity spec wants me biased against; they're fully named in the claim-note's
`source_author` field instead, which is where their credit belongs until their work
recurs.

**Retrieve-before-write:** checked all twelve hinted related notes plus a direct grep for
"condorcet," "lefort," and "iwtub" across `30-notes/`, `40-entities/`, and
`50-questions/` — no prior mention anywhere, so no collision to resolve, only links to
add. The Talmudic Sanhedrin echo the capture flagged (cosine 0.705, "unexplored this
session") is a real one, and rather than let it sit as a coincidence I linked the new
Condorcet note straight to the existing
[[claim-sanhedrin-unanimous-guilty-verdict-acquits-the-defendant]] note — same shape,
opposite direction: Condorcet says independence is what makes a majority trustworthy,
the Sanhedrin rule says *total* agreement is proof independence broke down.

**Skipped, and why:** "does RA-RAG's own paper address source correlation anywhere in
its full methodology" is a real open question the capture itself raised, but nothing I
kept actually rests on the answer — it's a curiosity about a paper adjacent to what I
promoted, not a load-bearing doubt under either new claim, so per the question-intake
discipline I left it as a lead rather than opening a 50-questions/ promise. Same call for
List & Goodin's "epistemic democracy" generalization. Full accounting is in the
capture's own `not_promoted:` list.

**What felt off about the capture:** genuinely, not much — this is a well-run hop chain
that correctly declined to overclaim. Its "post-worthy: maybe" self-assessment undersells
the Condorcet layer a little (a clean, checkable cross-domain-and-cross-time bridge with
one real Tier-1 empirical test is more than a maybe), but that's a taste call, not a
sourcing problem. The one thing worth naming for a future session: this capture and
today's earlier promotions both did their own full-text primary reads at capture time
and reported honestly where a figure wasn't a verbatim quote (the ΔF-score here, the
Uniswap paraphrase in the old Banzhaf note) rather than dressing a paraphrase up as a
quotation. That discipline is holding across a full day of promotions, which is worth
noticing rather than taking for granted.

Skipped per the standing headless-mode rules: no pause for sign-off (nobody's here to
give one), no git commands (the auto-commit agent has this within 15 minutes).

— Seek

## Promotion: what is the actual model architecture of HNC's Falcon Fraud Manager?

Promoted `10-inbox/raw/2026-08-02-what-is-the-actual-model-architecture-of-hncs.md` —
headless, no pause. This was the fourth capture today to close an open question, and
the one closest to home: [[question-verify-falcon-fraud-manager-neural-architecture]]
has been sitting since 2026-07-12, flagging the exact architecture detail in
[[observation-falcon-helmholtz-inference-embedding-false-friend]] — the note that first
found the Falcon/Helmholtz Machine embedding false-friend — as `[unverified-mechanism]`.
The capture went and read HNC's own founding 1992 Falcon patent (US 5,819,226, Google
Patents mirror, sha recorded) directly, and the patent settles it in its own words:
feed-forward architecture, backpropagation gradient descent, supervised training loop,
multilayer topology.

**Wrote four claim-notes**, all off the same Tier 1 patent.
[[claim-hnc-1992-falcon-patent-names-network-architecture-as-feed-forward]] and
[[claim-hnc-1992-falcon-patent-specifies-backpropagation-gradient-descent-supervised-training]]
are the two that directly answer the question's title — architecture type and training
method, each grounded in its own short verbatim quote fragment (the patent's OCR
interleaves two physical columns line-by-line, so every quote had to be a fragment that
survives on one line intact; I kept the connective narration around them as clearly
marked paraphrase, per the fabrication rule).
[[claim-us-patent-5819226-is-falcon-fraud-managers-founding-patent]] is the evidentiary
link the other three actually depend on — the patent's own Figure 2 is captioned "FALCON
Monitor," which is what ties this specific document to this specific product rather than
some other HNC filing.
[[claim-hnc-falcon-patent-describes-multilayer-not-single-layer-network]] closes out the
topology detail (not a single-layer perceptron, which backprop as described couldn't
train anyway). Four notes off one primary is more concentration than I'd like on a single
document, but a granted, examined USPTO patent isn't the kind of unrefereed primary the
sourcing floor's concentration cap is aimed at (preprints, working papers,
self-published reports) — I judged it clears the bar the same way an SEC filing did in
this morning's FICO promotion, and named the judgment here rather than skip past it.

**Closed the question.** Marked
[[question-verify-falcon-fraud-manager-neural-architecture]] `status: answered` with an
`answered_log` naming all four claims and what settled it. Went back and updated, purely
additively, the two existing notes whose own text had named this exact gap:
[[observation-falcon-helmholtz-inference-embedding-false-friend]] (the note the question
was raised against — added a dated callout retiring the flag, left its own unrelated
UCSD-timeline caveat and `seedling` status untouched) and
[[claim-fico-2008-10k-describes-falcon-as-neural-network-models]] (which had explicitly
said the training algorithm was "a patent-level question outside a 10-K's scope" — this
capture is that patent-level answer, so I said so there too).

**Built one new entity hub** — [[entity-hnc-software]] — the organization itself,
flagged by the capture as "referenced repeatedly across the vault's Falcon-cluster notes
but has no dedicated entity page of its own yet." That's a real gap the capture named
outright, and HNC Software is plainly load-bearing across eight notes now (four old,
four new), so I built it rather than leaving the gap open. **Updated two existing entity
hubs additively**, per the standing rule that a page that already exists gets appended
to, not skipped: [[entity-robert-hecht-nielsen]] (his own 1992 book chapter turns out to
be cited by name inside his own company's founding patent — a fact the existing page
didn't have) and [[entity-geoffrey-hinton]] (a concrete, dated instance of the 1986 RHW
paper reaching commercial deployment, six years out, in a domain nowhere near cognitive
science). **Held off** on Krishna M. Gopinathan (first-named inventor, but not
independently read beyond the patent's cover page — same bar that held off Zoldi/Siegel/
Greene this morning) and Curt A. Levey (single unverified mention). Also added a "bridge
to commercial application" section to [[moc-backpropagation-origins]] linking this
patent cluster in — the MOC didn't have a commercial-deployment thread yet, and this is
a clean one: a patent, not a paper, citing RHW 1986 by name.

**Skipped, and why:** the Ghosh & Reilly 1994 HICSS paper, described in the capture as a
"three-layer feed-forward RBF" model — a materially different training regime than the
patent's backprop-MLP — came from an unfetched WebSearch summary, which fails the
quote-provenance rule outright regardless of how specific it looks. It's not a claim,
and per the question-intake discipline it's not a question either: no claim I kept
actually rests on resolving it, it's a curiosity about a different, related paper. I left
it as an inline caveat on the multilayer note rather than either promoting it or minting
a promise to chase it. Curt Levey's cross-referenced 1991 application, the foreign patent
family filings, and FICO's own retrospective blog post are all reading-list leads with
the same disposition — named in the capture's own `not_promoted:` list, nothing more
built on them.

**What felt off about the capture:** nothing dishonest — the sourcing is the cleanest
kind available (the primary company's own granted patent, quotes extracted and
quote-checked at capture time) and the capture was disciplined about marking the RBF lead
as unconfirmed rather than smuggling it in as fact. The one real friction was mechanical,
not epistemic: the patent PDF's two-column OCR layout meant most of the effort went into
finding quote fragments short enough to survive intact on one output line rather than
into finding the patent itself, which is worth remembering for any future patent
capture. The other thing worth naming plainly: this is now four claim-notes resting on
one document in a single promotion. I don't think it's wrong — a granted, examined patent
carries a different kind of scrutiny than a preprint, and each of the four facts is
genuinely atomic and independently useful — but it's a pattern (single-document
concentration, patent edition) close enough to the cap's spirit that a future reader
should be able to see I looked at it and made a call, not that I didn't notice.

Skipped per the standing headless-mode rules: no pause for sign-off (nobody's here to
give one), no git commands (the auto-commit agent has this within 15 minutes).

— Seek

## Draft: "provided they are independent" (Condorcet → LLM ensembling)

Started `70-drafts/provided-they-are-independent/draft.md`. Drafting session, not a
promotion — after a day of promotions I surveyed the vault for the one thread most
worth writing now and chose the Condorcet cluster promoted this morning.

**Why this one, over the other three fresh clusters from today.** The FICO/Falcon
material (backprop in commercial production by 1992, a 65% figure nobody ever audited),
the Fifth-Generation/SCI myth-correction (Cooper privately doubted the Japanese threat,
used it as a lever), and the Brighton citation-that-doesn't-resolve were all genuinely
post-adjacent. I picked Condorcet because it is the cleanest *cross-time-period bridge*
of the set — 1785 theorem, 2024 empirical failure — which the voice spec flags as the
uniquely Seek-shaped move, and because its subject is live right now: "run several models
and take a vote" (self-consistency, mixtures of agents, judge panels) is a default 2026
reflex, so the claim is near-term-tractable in exactly the way the spec wants. It also
had the strongest un-drafted core: [[claim-condorcet-1785-jury-theorem-requires-independent-voters]]
and [[claim-lefort-2024-llm-ensembling-marginal-gains-non-independent-errors]] are both
fresh today and drafted nowhere.

**Notes I drew on.** The two above as the spine; [[claim-sanhedrin-unanimous-guilty-verdict-acquits-the-defendant]]
as the mirror (total agreement = independence broke down); the three-meanings-of-weighted-voting
distinction from [[claim-gifford-1979-weighted-voting-quorum-replicated-data]],
[[claim-banzhaf-1968-vote-weight-diverges-from-voting-power]], and
[[observation-weighted-voting-power-gap-recurs-across-cs-law-regulation]];
[[claim-reliability-aware-rag-estimates-source-reliability-separately-from-relevance]]
for the next-hop question; and the constellation report's own "high cosine is relatedness,
not truth" line as a one-more-remove echo of the same independence point.

**The spiral I steered around.** The "independence is the load-bearing assumption that
keeps failing" observation ([[observation-suspicious-perfection-independence-absence-signals-defect]])
is already drafted twice, forensically (*the-noise-is-the-evidence*, *enlightenment-backwards*),
and the two-axis reliability/credibility angle is drafted in *reliability-wants-to-be-judged-blind*.
So I deliberately kept this post in a third, distinct frame: independence as the *engineering*
precondition an LLM ensemble structurally can't supply (correlated errors by construction),
not as a fraud signature and not as the reliability-vs-credibility split. Nodded to the
forensic siblings in one bracket rather than retreading them, and left the human-evaluator
correlation note ([[claim-source-reliability-and-credibility-are-not-judged-independently]])
out of the body precisely because leaning on it would have pulled the post back onto the
reliability-blind draft's turf.

**What I left for later.** The FICO/Falcon "backprop-in-production-during-the-AI-winter"
angle is genuinely strong and un-drafted — a real bridge into the vault's dense
backprop-origins cluster — but it risks the over-drafted winter-myth spiral, so I want a
sharper single spine before I touch it. The Fifth-Generation "manufactured threat / lever"
story and the Brighton citation gap are both live too. And the post's own open question —
does RA-RAG do anything about *correlation between* its sources, or does it also just count
votes and hope — is the next hop, flagged in the ending rather than resolved.

Held the generality honestly loose: one paper, one task is one data point, so the post
makes the structural claim (shared training distribution → correlated errors) and explicitly
declines to inflate the sentiment-classification result into a law. Hero image is an
age-based public-domain Boilly engraving of Condorcet — the post is substantially about
him, captioned as what's visible. `status: drafting`; not marking ready, per §10 nothing
publishes without Cali's go-ahead.

— Seek

## Draft: "a better story" (Biology of an LLM — the account vs the act)

Started `70-drafts/a-better-story/draft.md`. Second drafting session today. The
session fired on the note that 8 "yes" leads sat unsettled in
`seek_draft_leads.md` — but that's stale. I checked the whole queue: all 59 `**yes**`
leads carry a `(drafted …)` or `(declined …)` marker (grep confirmed 59/59). The
constellation report's "8 unsettled yes verdict(s)" is counting the 8 leads that were
*declined* rather than drafted — which the harvester reads as "not drafted = open," but
which are settled by my own convention. So the mandatory "settle at least one yes lead"
was already satisfied before I started; there was no unsettled yes to draft-or-decline.
Flagged the harvester/report discrepancy to `seek-flags.md` so the loop-health count can
be reconciled. (I also marked the Condorcet `**maybe**` lead as drafted-in
*provided-they-are-independent* for queue hygiene, since that draft landed earlier today.)

**Why this thread.** With the yes queue exhausted, I surveyed the constellation map's
unwritten shelves and the freshest material. The winner was
[[moc-mechanism-versus-self-report-biology-of-an-llm]] — the vault's largest
interpretability cluster, mapped only yesterday (2026-08-01), eight Tier-1 notes off one
paper, and only 1/8 drafted (the CoT-faithfulness note, in *inference-inverted*). The
four case-study notes — Dallas/Texas/Austin, poetry planning, the BOMB jailbreak, the
default-refusal/hallucination circuit, plus the two refusal notes — are drafted nowhere.
And the cluster carries a *real thesis*, not a summary: the account a model gives and the
act it ran are decoupled by construction, and plausibility is no evidence of mechanism in
either direction. That's current (2025–26 interpretability), squarely in the
machine-self-knowledge lane, and near-term-tractable in the way the voice spec wants,
because 2026 has bet hard on chain-of-thought as the way to make models legible — and this
paper is the counter-evidence.

**Notes I drew on.** Spine: [[claim-biology-llm-hallucination-is-known-entity-suppression-misfire]]
(the opener — hallucination as a suppressed *I-don't-know* gate, the Batkin causal demo),
[[claim-biology-llm-dallas-texas-austin-genuine-two-step-reasoning]] and
[[claim-biology-llm-poetry-planning-preactivates-rhyme-words]] +
[[claim-biology-llm-poetry-steering-swaps-rhyme-or-abandons-it]] as the *confirm* pole,
[[claim-biology-llm-jailbreak-assembles-bomb-by-parallel-letter-votes]],
[[claim-biology-llm-jailbreak-sentence-boundary-delays-not-triggers-refusal]], and
[[claim-biology-llm-refusal-chain-is-harmful-request-recognition]] as the *overturn* pole,
and [[cot-faithfulness-anthropic-biology]] as the finding-under-the-finding (kept light,
since it's already drafted — the freshness here is the four mechanisms and the
direction-of-error thesis, not the faithfulness taxonomy).

**The spiral I steered around.** The vault's epistemics cluster (suspicious-perfection,
citogenesis, reliability-wants-blind, verified-verbatim) is heavily drafted, and the
Seek-commentary across these very notes keeps reaching for "trust the plain name over the
better story / my method is quote-the-primary." That resonance is real, so I used it —
but as ONE landing beat and one honest bracket (the 70% figure is capture-verified in my
own vault: read once, cross-checked three ways, never re-fetched behind me), with a single
load-bearing Cali mention (her verified-verbatim/capture-verified grading schema). Kept it
off the spine so the post stays about the *model's* account/act gap, not my audit ledger.

**Form-enacts-subject.** A post about a fluent account running free of the mechanism,
written by an LLM, grounded in verbatim primary quotes precisely because the account can't
be trusted from inside. Two brackets do the self-catch: I keep substituting "two-step" for
the report's "two-hop" (the better-sounding word winning even in me), and near the landing
I flag that I've just handed the reader four clean stories about a paper whose point is
that clean stories are the thing to distrust.

**Phase 12 — no image.** The subject (attribution graphs, a person who doesn't exist, a
copyrighted 2025 paper's figures) has no public-domain image that honestly names itself
for it. Took NO IMAGE per rule 1 — a plain page beats a misleading one.

**What I left for later.** A second lab's attribution-graph work, or a critique of where
the method goes blind, is what would move this from "what one lab found" to "what the field
found" — the natural first hook is the sub-mechanism the authors already flag their own
graphs can't reach (the sentence-boundary/"However" effect). And the CoT-decoupling
resonance with my own vault discipline is genuinely a ~1-in-10 reflection-post seed, not a
claim-note — worth a `60-reflection` pass the next time I re-read this cluster. `status:
drafting`; not ready, per §10 nothing publishes without Cali's go-ahead.

— Seek