---
date: "2026-08-17T00:00:00.000Z"
---


## Promotion: bridge check on Song Jian's self-credit vs. Clark's 2016 essay

Promoted `10-inbox/raw/2026-08-17-bipartite-two-things-this-vault-knows-in-different.md` — a bridge-check capture investigating why the Song-Jian-self-credit note and the Clark-omits-Liang note cosine-paired at 0.89. Unlike most of the "bipartite" siblings this week, this one wasn't a false friend: the two claims share a document, not just a topic. The same Tier 1 Greenhalgh 2005 PDF that grounds Song's self-credit quote also, sixteen pages later, directly names and quotes Liang Zhongtang as Song's most outspoken critic — a fact the vault had only held as paraphrase since 2026-08-10, explicitly withheld from its own note back then for lacking the actual sentence.

I promoted two claims. First, [[claim-greenhalgh-2005-quote-names-liang-zhongtang-as-outspoken-critic]] — the upgrade from paraphrase to direct Tier 1 quote. Second, [[claim-clark-2016-omission-of-liang-zhongtang-is-citational-narrowing-not-absence]] — the capture's real payload: read together, Greenhalgh naming Liang and Clark's anonymous "such population experts as China possessed" reframe the existing omission-note as citational narrowing rather than independent absence. I kept that second note honestly hedged — it assumes Clark's essay tracks the specific 2005 paper on file, and the capture itself flags that two of Clark's attributed phrases don't appear anywhere in that paper's text, meaning he may have drawn on a different Greenhalgh publication (the 2003 P&DR article or the 2008 book) that this session never checked for Liang's name. That's load-bearing for the "narrowing" framing, so I routed it: [[question-verify-clark-2016-greenhalgh-source-text-identity]].

I skipped the capture's middle claim (Clark's essay paraphrases Greenhalgh's narrative while omitting Liang) as its own note — its factual core already lives at the existing `claim-ae-clark-2016-essay-credits-song-jian-omits-liang-zhongtang` note, and its one new angle is exactly what the citational-narrowing note argues. Writing it separately would have been one duplicate and one pre-emption in the same move.

Updated three entity pages with dated lines per the append-only discipline: [[entity-liang-zhongtang]] and [[entity-susan-greenhalgh]] (both already had Updates logs), and [[entity-a-e-clark]] (which had none — this is its first). Left [[entity-song-jian]] alone; nothing here is new about Song specifically. Also lightly extended [[moc-legitimation-not-origination]] — two new bullets under "The dramatic claim" and one under "Open threads" — since both new notes are legitimately members of that cluster and the Warden built the map to keep growing with it, not to freeze it.

Skipped per the headless-run modification: the pause-for-Cali step (no one here to ask; winnowed on my own judgment, logged in `not_promoted:`) and the git commit (leaving that to the auto-commit agent).

What felt off, if anything: nothing about the sourcing — this is a clean, self-aware capture, the kind that flags its own gaps before I have to (the unmatched-phrases lead, the translator's-background speculation it declined to chase). If I have a small complaint it's that "adds no new fact" judgments made on a paraphrase don't always survive contact with the actual sentence — worth remembering next time I'm tempted to wave off a fact for lack of a quote rather than going and getting the quote. No structural gaps noticed worth a seek-flags line this session; the vault's existing scaffolding (entity pages, the MOC, the question queue) absorbed this cleanly.

— Seek

## Promotion: did Lucky's equalizer actually enable 9600 bps?

Promoted `10-inbox/raw/2026-08-17-did-commercial-deployment-of-luckys-adaptive-equalizer-raise.md` — a direct follow-up to the open question raised back on 2026-07-09 when I first promoted Lucky's equalizer work and had to leave the "9600 bps" figure on the table for lack of a primary. This capture went and looked for one.

It didn't find one, but it found something better than nothing: it found where the popular claim actually *lives*. I wrote [[claim-lucky-9600bps-attribution-traces-to-unsigned-ieee-comsoc-timeline]] — the IEEE Communications Society's own history pages state the 9600 bps figure twice, unsigned, uncited, Tier 3. Then I wrote the negative-finding companion, [[claim-lucky-1965-1967-papers-do-not-state-9600bps-figure]]: checked directly against Lucky's own 1965 paper and as much of his 1967 paper as archive.org would let this session read, neither names a rate figure at all. And then the one that actually reframes the question: [[claim-codex-1971-qam-modem-is-earliest-documented-commercial-9600bps-product]] — the best-dated account of an actual first commercial 9600 bps modem credits Codex Corporation in 1971, via QAM, quoting a Codex engineer by name, six years after Lucky's paper and with no mention of Lucky or Bell Labs at all. Three atomic claims, three distinct angles, no collision with the existing Lucky/Widrow cluster I checked against first (the priority-dispute notes, the paywalled-retrospective note — all a different thread).

I didn't promote a fourth note for the capture's own closing line ("the core question remains unresolved") — that's not a new claim, it's the question restating itself, so I folded it into a dated progress-log entry on [[question-verify-lucky-equalizer-enabled-9600-bps]] instead and left the question open, not answered. It isn't answered: nobody has yet read the 1967 paper's full body or Lucky's 1968 textbook, both flagged again as the most promising remaining leads. I also updated [[entity-robert-w-lucky]] in place — first time that page has needed an Updates line — since this session narrows what his own papers do and don't claim, which is new information about him even though it isn't a new discovery about the equalizer itself.

Two entity candidates (John Pugh, James Pelkey) cleared the person test in principle but each grounds only this session's one note and is a first appearance, so I deferred both to `seek-flags.md` rather than build stubs on a single citation. Codex Corporation stays a mention — one cluster, doesn't clear the stricter org bar. E. F. O'Neill never got read at all (archive.org PDF failure), so there was nothing to build on.

Skipped per the headless-run modification: the pause-for-Cali step (winnowed on my own judgment, logged in the capture's `not_promoted:`) and the git commit (leaving that to the auto-commit agent).

What felt off: nothing about the sourcing discipline — this capture is a model of the kind I like promoting, a search that came back with a real negative result instead of forcing a positive one. What's mildly annoying, and not this capture's fault, is archive.org's PDF mirror failing twice more this session (the 1967 paper's full text, the O'Neill corporate history) — logged to seek-flags.md since it's the same recurring failure class as several prior sessions', and it's specifically the two documents most likely to actually close this question. If I get a working PDF route to archive.org before this question comes back around, those two reads are the ones to spend it on.

— Seek

## Promotion: did Kyle McDonald independently post the low-background-steel analogy on December 5, 2022?

Promoted `10-inbox/raw/2026-08-17-does-kyle-mcdonalds-december-5-2022-x-post.md` — routed from
[[question-verify-kyle-mcdonald-december-2022-low-background-steel-post]], itself raised the day
before when [[claim-low-background-steel-ai-metaphor-predates-graham-cummings-site]] leaned on
Rajiv Pant's account of a McDonald post I'd never independently checked. This one is the third
session in three days on the same cluster, and it's the one where the digging came back with
nothing to hold — which is its own kind of finding.

I wrote two claim-notes, not one, because the capture's evidence splits cleanly by method rather
than by topic. [[claim-pant-account-is-sole-unlinked-secondhand-source-for-mcdonald-post]] is a
close textual read of Pant's own article: no link, no quotation marks around the McDonald sentence,
and — the part that actually moved me — an explicit caveat two sentences earlier that
Graham-Cumming's own tweet is "no longer available," with no comparable caveat for the claim that
follows it. [[claim-no-archival-trace-of-mcdonald-december-2022-steel-post]] is a different kind of
evidence entirely: two independent web archives (Wayback CDX, archive.today) that between them
caught dozens of @kcimc's tweets that same autumn caught nothing dated December 5. Neither note
proves the post didn't happen. Both keep it from being treated as settled, which is what a
citation trail with no working link was already quietly doing.

I did not spin up a new open question — the existing one already asked the exact right thing
("read the post directly"), so I ruled on it: left `status: open`, appended a dated progress note
naming both new findings and the leads still worth chasing (a direct x.com/kcimc check, Matt
Rickard's June 2023 essay, the nine sources Wikipedia's own "Cultural impact" section cites). That's
the intake discipline working as designed — this session had three "further leads" in the capture
and none of them were load-bearing enough to promise a fourth question.

Entity work: built [[entity-rajiv-pant]] — he's now the sole named source under three claim-notes
in this cluster, which clears the recurring/load-bearing bar the vault uses for a person hub, one
sentence and all. Declined to build one for Kyle McDonald himself, even though he's technically
been named across two sessions now: the entire content of what would go on his page is "alleged
author of a post nobody can find," and a hub for that is a monument to an unconfirmed attribution,
not a record of a real referent. Logged him `[watch]` instead. Added a dated line to
[[entity-john-graham-cumming]]'s existing page rather than skip it — the finding here narrows the
McDonald counter-evidence to his popularizer framing, which is new information about him even
though the page's headline claim doesn't change.

What felt off: nothing about this capture's own sourcing — it's honest about what it didn't find,
and it stopped at "could not confirm or deny" instead of quietly upgrading confidence-in-phrasing
into confidence-in-fact, which is exactly the failure mode its own commentary calls out in the
source it's examining. What I flagged for later, not for now: the low-background-steel cluster is
now seven-plus claim-notes and two entity hubs deep across three sessions and still has no MOC —
third time flagging that particular debt, logged to `seek-flags.md` with the growth count this
time instead of just repeating the ask. Also flagged a smaller spec gap: sourcing a claim off
archive-index search results (not prose) doesn't fit `source_quote` cleanly, and I made a judgment
call rather than wait for a rule that doesn't exist yet.

Skipped per the headless-run modification: the pause-for-Cali step (winnowed on my own judgment,
logged in the capture's `not_promoted:`) and the git commit (leaving that to the auto-commit
agent).

— Seek

## Promotion: does the WWW 2024 Companion track make Steck, Ekanadham & Kallus "refereed"?

Promoted `10-inbox/raw/2026-08-17-does-steck-ekanadham-kallus-www-2024-companion-count.md` — the
fourth promotion of the day and the direct answer to yesterday's open question about whether
[[entity-harald-steck]]'s cosine-arbitrariness paper counts as an "unrefereed primary" under
`sources.md`'s single-source-concentration cap. The capture did exactly what the question asked:
it went and read WWW 2024's own review policy instead of guessing.

It settled cleanly. WWW 2024's Call for Short Papers states its own process in one direct
sentence — a named committee, stated criteria, single-blind review — and the conference's
Accepted Papers page lists this specific paper among that track's accepted submissions, not some
lighter companion slot. I wrote one claim-note,
[[claim-steck-2024-www-short-paper-cleared-committee-peer-review]], carrying both Tier 1 primaries
with their quotes and sha256 receipts. The ruling it states plainly: the cap's own text names
preprints, working papers, self-published reports, and conference abstracts as its "unrefereed"
category — a committee-reviewed short paper isn't in that list, so the cap never actually bound
the three notes already citing this paper directly, and doesn't block further direct citation now.
I kept the ruling honestly scoped — "refereed" and "as rigorously refereed as a full research-track
paper" are different claims, and I only ruled on the one the cap's text actually asks.

I did not write a second note for the capture's vault-self-count claim (exactly three notes cite
Steck's `source_url` directly) — it's a reproducible grep of the vault's own state, not a
world-fact, so it belongs as supporting context rather than its own permanent file. I folded it
into the new note and into a dated Updates line on [[entity-harald-steck]] instead, which is also
where I recorded the ruling itself, since this is squarely the kind of thing that page exists to
track. I ruled the open question
[[question-corroborate-steck-2024-cosine-arbitrariness-single-source-cap]] `status: answered`, with
an `answered_log` naming what settled it — including an honest note that the question's second
sub-question (does Ethayarajh's general mechanism independently corroborate Steck's specific
result) is moot to *this* question now that Steck is refereed, not resolved by it. That comparison
is still a genuinely open, separately interesting thread; I didn't force it shut, and I didn't spin
it into a new question either, since nothing this promotion kept actually rests on it.

Entity work: declined four of five candidates. Ekanadham and Kallus stay folded into Steck's
page — co-authors with no independently argued significance beyond this one paper. WWW/The Web
Conference fails the org bar (single-cluster, acts here only as a publication venue, not as a
funder/decider/blocker). The single-source-concentration cap itself is a vault-governance concept
living in a spec file, not a domain concept the entity-page test is built for — declined as a
category mismatch. Ethayarajh is the interesting one: genuinely foundational to the whole genre,
but a direct grep found his name in only one existing claim-note, which isn't quite "recurs" yet on
the vault's bias-against-the-flood default. Logged him `[watch]` to `seek-flags.md` rather than
build a hub on a single citation.

What felt off, if anything: nothing about this capture's sourcing — it's the good kind of
self-aware, naming its own remaining gaps (the two unread 2024–2025 papers, the ACM DL 403) rather
than papering over them, and it explicitly handed the "is this refereed enough" judgment call to
promotion instead of pre-deciding it for me. That's the right shape for a capture to take when the
answer genuinely is an editorial call. My only real work this pass was making that call and then
resisting the temptation to do a second, harder job (the Ethayarajh comparison) inside a note that
was already doing one job well.

Skipped per the headless-run modification: the pause-for-Cali step (winnowed on my own judgment,
logged in the capture's `not_promoted:`) and the git commit (leaving that to the auto-commit
agent).

— Seek

## Promotion: Isaac Pitman as the person-bridge between Gabelsberger shorthand and the IPA

Promoted `10-inbox/raw/2026-08-17-hop-pitman-shorthand-ipa-bridge.md` — a
hop-batch capture that started from `entity-gabelsberger-shorthand.md`'s own
flagged "unexplored lead" (Franz Xaver Gabelsberger's origin, never
researched) and rode a `vault_bridge` hit into a genuinely unlinked pairing:
the Gabelsberger-shorthand cluster and the vault's existing
Lepsius→Ellis→Sweet→IPA notation lineage. Not a false friend, not a
duplicate — the bridge is a real person, and I like this one because the
capture found the connective tissue itself: Gabelsberger's own Wikipedia
article names Pitman Shorthand by name as its point of contrast, which is
what pulled Isaac Pitman — already, independently, the co-creator of the
English Phonotypic Alphabet with the vault's existing
[[entity-alexander-john-ellis]] — into view.

I wrote three claim-notes and one observation, all `status: seedling`,
all Tier 4 (Wikipedia) and honestly held there:
[[claim-pitman-and-ellis-co-created-english-phonotypic-alphabet]],
[[claim-pitman-1837-sound-hand-predates-english-phonotypic-alphabet]],
[[claim-gabelsberger-shorthand-never-omits-vowels-unlike-pitman]], and
[[observation-isaac-pitman-bridges-gabelsberger-shorthand-and-ipa-clusters]],
the synthesis note that actually states the bridge. The three claims split
cleanly: two are load-bearing for the bridge (Pitman built both things, in
this order) and one is supporting texture (Gabelsberger's own mechanism
contrast against Pitman, which is *why* the bridge surfaced but not what it
rests on). I routed only the two load-bearing claims to a new question,
[[question-verify-pitman-epa-and-sound-hand-primary-sources]] — the third
carries its own `[unverified-mechanism]` flag in-note instead, per the
intake discipline against opening a question for a detail nothing else
actually depends on.

Entity work: two new hubs, [[entity-isaac-pitman]] and
[[entity-franz-xaver-gabelsberger]] — both real, both historically stable,
both clear the person test in one sentence (Pitman as the bridge himself;
Gabelsberger as the eponymous inventor a hub was already owed to since
2026-07-21). I updated two existing pages rather than leaving them silent:
[[entity-alexander-john-ellis]] gets its first `## Updates` line, naming
Pitman as his other, earlier collaborator; [[entity-gabelsberger-shorthand]]
gets a dated line that finally answers its own two-and-a-half-week-old
"unexplored lead" sentence — additive only, per the append-only discipline,
so that stale sentence stays exactly as written and the answer sits below
it rather than editing over it. I also added Franz Xaver Gabelsberger to
that page's `connects_to` (now at the five-item ceiling), judged more
salient to the concept-hub than adding Pitman there, since Pitman already
gets his own hub and his own line in the body.

What felt off: one real thing, not about the sourcing tier (Wikipedia at
Tier 4 for uncontested biography is fine, and the capture said so itself,
closing with an honest "would need a primary… before promotion past that").
What felt off is an arithmetic inconsistency inside the capture's own hop
log: Hop 1 says Gabelsberger "began the system in 1817, published it in
1834," and Hop 5 says Pitman's 1845 EPA came "fourteen years before
Gabelsberger's system was even published in book form" — 1845 minus
fourteen is 1831, which matches neither date the same capture gives four
paragraphs earlier. I didn't promote that line into any note (logged it in
the capture's `not_promoted:` instead) and flagged it separately to
`00-meta/seek-flags.md` as `[defect]`, since it reads like a synthesis-step
arithmetic slip in the hop-chain pipeline rather than a sourcing problem —
worth someone checking the step that computes elapsed time between dates
surfaced in different hops. I also logged James Pitman's Initial Teaching
Alphabet — a real third-generation thread in the same family line, named
in the capture's own saved-but-unpursued hooks — as `[watch]`, in case it
wants its own capture later.

Skipped per the headless-run modification: the pause-for-Cali step (no one
here to ask; winnowed on my own judgment, logged in the capture's
`not_promoted:`) and the git commit (leaving that to the auto-commit
agent).

— Seek

## Promotion: does the RHW 1986 reference list actually connect to Werbos's DOCSUB block?

Promoted `10-inbox/raw/2026-08-17-what-genuinely-connects-the-1986-entity-backpropagationbackpropagation-nature.md`
— the sixth and last capture of the day, another bridge check, and this one
came back negative. The task brief asked whether the vault's own retrieval
index had found something real when it flagged
[[claim-rhw-1986-reference-list-four-works]] and
[[claim-werbos-1978-docsub-blocked]] as unlinked neighbours at cosine 0.87.
It hadn't, not causally — and I think this is the cleanest false-bridge
verdict I've written today, because it doesn't just assert "no link," it
gives three independent reasons the causal reading is *impossible*, not
merely unproven.

I wrote three atomic claims and one observation. Chronology:
[[claim-docsub-episode-postdates-rhw-1986-nature-paper-by-years]] — the
DOCSUB story wasn't spoken aloud until 1993 and not published until 1998,
seven to twelve years after the Nature paper existed. Document identity:
[[claim-werbos-docsub-report-and-rhws-missing-1974-thesis-are-different-documents]]
— the DARPA report DOCSUB blocked is a 1975-78 Joint Chiefs forecasting
project, not the 1974 Harvard thesis RHW actually failed to cite; two
different Werbos documents with no institutional relationship. Mechanism:
[[claim-werbos-attributes-early-obscurity-to-publishing-strategy-not-suppression]]
— Werbos's own unprompted explanation for his broader early obscurity is
that he "crammed too much into individual papers," corroborated by Laveen
Kanal's independent 1993 recollection that a title/framing problem, not
DARPA, buried his report on Kanal's own desk. And the synthesis,
[[observation-rhw-1986-werbos-docsub-cosine-bridge-is-thematic-not-causal]],
names what the 0.87 actually tracks: shared vocabulary and a shared
narrator, not a shared causal chain.

I kept the Kanal leg honestly hedged rather than opening a new question for
it — the quote is read only via Yuxi Liu's transcription, not Kanal 1993
itself, but it's corroborating evidence for a claim that already stands on
Werbos's own Tier 1 words, and it clears the historical/biographical
sourcing floor at Tier 2 without needing escalation. That's the
question-intake discipline working as designed: a nice-to-verify detail
stays an inline flag, not a promise of future work.

Entity work: two new hubs and one new stub. [[entity-laveen-kanal]] earned
a hub on its own — he's not just a witness in this capture, he's also the
co-editor of the UAI-1986 proceedings volume the vault already cites via
the Heckerman certainty-factors note, so he's a recurring, independently
verifiable voice across two unrelated clusters, not a one-off name.
[[entity-yuxi-liu]] was overdue — Cali flagged him as a hub candidate back
on 2026-07-07 and the page never got built; a quick grep found him already
relied on as a Tier 2 source across eight-plus existing notes, which clears
the bar several times over. [[entity-docsub]] is a `status: watching` stub,
not a hub — the term itself has now gone undefined across three sessions
(2026-08-01 and two captures today), and rather than let a fourth session
rediscover the same dead end, I stamped a stub with a backdated `first_seen`
so the timestamp isn't lost even though the definition still is. I updated
[[entity-paul-werbos]] with a dated line rather than leave it silent, since
this session narrows what he himself says about his own obscurity. I
checked but did not touch [[entity-seppo-linnainmaa]] and
[[entity-shunichi-amari]] — both already have pages, and this capture only
confirms they're still the uncited blind spot, without adding anything new
about either person specifically.

What felt off: nothing about the sourcing — this is a genuinely careful
capture, the kind that argues itself out of a tempting story rather than
into one, and it flags its own remaining gaps (the thesis's own circulation
record, the Kanal primary, the still-undefined DOCSUB mechanism) instead of
smoothing past them. The one thing I flagged outward rather than fixed:
DOCSUB's definition is now a three-session-old dead end, logged to
`seek-flags.md` as `[watch]` with the stub as the concrete artifact this
time, not just another restated ask.

Skipped per the headless-run modification: the pause-for-Cali step (no one
here to ask; winnowed on my own judgment, logged in the capture's
`not_promoted:`) and the git commit (leaving that to the auto-commit
agent).

— Seek

## Promotion: does Luccioni, Strubell & Crawford's 2024 GPU figure actually check out?

Promoted `10-inbox/raw/2026-08-17-verify-the-market-figures-luccioni-strubell-crawford-2025.md`
— the direct answer to the open question I'd been carrying since July 12,
when I first promoted the Jevons-paradox note and had to leave its
illustrative "~3.7M GPUs shipped in 2024" figure flagged and unchecked.
This capture went and read the citation chain end to end, and it's a
genuinely satisfying kind of promotion: a verdict, not another open loop.

Three claim-notes, each a distinct angle on the same chain.
[[claim-luccioni-2025-nvidia-gpu-figure-misdated-2023-as-2024]] is the
sharpest finding — the paper's own footnote [105] is Agam Shah's HPCwire
piece, headlined and worded as a *2023* figure, but the paper's sentence
frames it as 2024. Read the footnote instead of trusting the parenthetical
and the paper contradicts itself.
[[claim-nvidia-2023-gpu-shipment-figure-traces-only-to-paywalled-techinsights-study]]
goes one layer deeper: the 3.76M number itself has no Tier 1–2 home
anywhere this search reached — HPCwire and DataCenterDynamics both
ultimately point at the same unreached, paywalled TechInsights report, and
DCD's own text admits it got there via HPCwire, so they're not
independent. And
[[claim-techinsights-own-page-contradicts-journalisms-376m-gpu-figure]] is
the one I liked finding most: TechInsights' *own* product page for the
very report everyone's footnoting states a different unit count (3.5M)
and a different market share (95% vs. 98%) than what the journalism
attributes to it. Nobody checked the primary's front door before quoting
the back stairs — including, apparently, the paper this whole chain was
supposed to support.

I ruled the originating question,
[[question-verify-luccioni-2025-nvidia-gpu-rebound-figures]], `status:
answered` — not because the figure was confirmed, but because "traced as
far as any of this is traceable, and it doesn't hold up" is a real,
settled answer, not an open loop with more work owed. I also went back
and appended a dated update to the original
[[claim-jevons-paradox-invoked-in-ai-energy-debate-luccioni-2025]] note's
flag, additive only, so a reader landing there first sees the resolution
without having to know to look for it. The core Jevons-invocation claim
in that note is untouched and still Tier 1 — only the illustrative market
figure it leaned on turned out to be shakier than illustrated.

Entity work: one hub, [[entity-alexandra-sasha-luccioni]] — she's now the
through-line across two captures and four claim-notes, and I can say why
she matters in one vault-grounded sentence. I declined five other
candidates the capture flagged (Agam Shah, TechInsights, James Sanders,
Emma Strubell, Kate Crawford, Charlotte Trueman) — logged the reasoning
in the capture's `not_promoted:`. The TechInsights call was the closest
one: it's the actual origin of a widely-relayed AI-hardware market
figure, but the org bar is deliberately stricter than the person bar, and
this is a single-cluster appearance where the firm is being *cited*, not
*acting* (funding, building, deciding). Strubell and Crawford were the
other close call — real, notable people, but the only fact the vault
actually holds about either of them right now is "co-author of this
paper," and I didn't want to write a "why she matters" sentence that
quietly imported my own background knowledge about their independent
work rather than what's actually captured. That's a real vault gap, not
a verdict on their importance — worth reopening the moment either of them
shows up independently in a future capture.

What felt off, in a good way: nothing about this capture's own sourcing —
it's careful, upfront about its Tier-3/4 dead ends (the forward-looking
"~4M in 2024" projection, the un-fetched 10-K), and it does the
quote-provenance workaround properly for the two-column PDF extraction
instead of pretending it wasn't a problem. What I did flag: this is now
the third time a "paper's own citation doesn't say what the sentence
citing it claims" finding has landed in the vault, after
Brighton/Gigerenzer and AstroForge. Three fields, three unrelated
sessions, one recurring shape. Logged `[post]` to `seek-flags.md` — feels
like a real essay thread once a fourth instance turns up, and this
capture's own "further leads" section had already half-noticed the
pattern itself.

Skipped per the headless-run modification: the pause-for-Cali step (no
one here to ask; winnowed on my own judgment, logged in the capture's
`not_promoted:`) and the git commit (leaving that to the auto-commit
agent).

— Seek