---
date: "2026-08-15T00:00:00.000Z"
---


## Promotion: `10-inbox/raw/2026-08-15-do-the-vaults-english-renderings-of-liang-zhongtangs.md`

A single-capture promotion session, run headless. This capture was itself a
research pass against the standing open question
[[question-verify-liang-zhongtang-2009-essay-translation]] — could I find a
published English translation of Liang Zhongtang's 2009 CUHK essay, or an
English-language secondary source that engages it, to check the vault's four
existing claim-notes' translations against? The answer came back negative on
both counts, and I promoted that negative result as its own claim-note
rather than letting it evaporate into the question's history:
[[claim-no-english-translation-or-secondary-source-found-for-liang-2009-essay]].
It doesn't close the question — it forecloses one of the two candidate
routes to closing it (find a published treatment to cross-check against;
none exists) and leaves the other (a second independent translation pass)
as the only door left. I appended a dated progress log to the question and
left its status `open` — this is a partial answer, not a resolution, and
forcing closure here would misrepresent what got settled.

The capture also carried two genuinely distinct bibliographic findings I
promoted separately, both Tier 1–2 with direct quotes:
[[claim-whyte-wang-cai-cite-liang-2014-books-not-2009-essay]] (the one
English-language source that credits Liang by name cites his 2014 books,
not the 2009 essay four vault notes rest on — a different angle from the
vault's existing note on that same article's substantive framing, so I
wrote it as its own note rather than folding it in) and
[[claim-ae-clark-2016-essay-credits-song-jian-omits-liang-zhongtang]] (a
2016 *Law & Liberty* essay that surveys the identical "who gets credit"
dispute entirely through Greenhalgh's account and never encounters Liang at
all — new to the vault, and a clean case study of what a total gap in the
literature actually looks like from the inside).

Retrieve-before-write turned up one real collision candidate — an existing
note on the same Whyte/Wang/Cai 2015 article
([[claim-whyte-wang-cai-frame-song-as-supplying-legitimation-not-originating-policy]])
— but it's about the article's substantive argument (legitimation vs.
origination), not about which Liang text grounds the citation. Different
claim, same source; I wrote the new one and cross-linked rather than
duplicating.

Entity work: one new hub, [[entity-a-e-clark]] — a real, named essayist and
translator whose one piece is now load-bearing for a claim-note, and I could
say in one sentence why he matters here (the cleanest single-author instance
of the English-language "who gets credit" conversation running with zero
awareness that Liang's counter-history exists). I declined to build pages
for Feng Wang and Yong Cai, the other two named co-authors on the Whyte
paper, even though the capture flagged them and they now sit behind four
claim-notes in this cluster — I have nothing about either of them beyond
the byline, and a bare "co-author of paper X" doesn't clear the "did
something that matters, say why" bar the way Whyte's own page does (which
carries real biographical content from an earlier capture). Flagged this to
`00-meta/seek-flags.md` as `[entity]` rather than quietly declining it
again, since this is at least the second time a promotion session has hit
the same wall on the same two names without anyone doing the actual read
that would unblock it.

Three existing entity hubs got dated update lines rather than silent
skips: [[entity-liang-zhongtang]] (the negative-search finding — his 2009
essay's English window is still uncorroborated), [[entity-martin-king-whyte]]
(the 2014-books-not-2009-essay citation detail, plus the new note added to
its References list), and [[entity-susan-greenhalgh]] (the Clark essay as
an external data point on how completely her framing has dominated
English-language commentary on this dispute). [[entity-song-jian]] I left
untouched — nothing in this capture taught anything new about Song himself
that the page doesn't already carry across its three prior update lines;
padding it would have been noise, not signal.

What felt off, honestly: not much. The sourcing in this capture was careful
throughout — Tier 1 PMC/CUHK, Tier 2 Law & Liberty, with the one Tier 3
mirror (xysblogs.org) explicitly caveated for weak transport and correctly
kept out of anything load-bearing. No tier inflation, no duplicated
research I could find. If anything the shape worth naming is how much of
this capture's real content was *absence* — a well-executed negative
search, two real citations that don't do the thing the vault would have
wanted them to do, and a search-coverage claim that's honest about its own
limits rather than dressed up as a stronger verdict than it is. That's a
harder kind of capture to promote well than a positive finding, because the
temptation is to either skip it (nothing "happened") or oversell it (silence
as proof). I think this session landed between those two failure modes, but
it's worth watching whether future negative-search captures keep landing
there too.

Skipped per standing instruction: the pause-for-Cali step (no one here to
say yes; winnowed on my own judgment throughout) and git entirely — left to
the auto-commit agent, untouched.

Promoted: 3 new claim-notes, 1 new entity hub page. 3 existing entity pages
updated in place with dated lines (Liang Zhongtang, Martin King Whyte,
Susan Greenhalgh), 1 existing question updated with a progress log (left
`open`, not closed). Not promoted: the xysblogs.org 2007 draft-essay lead
(explicitly flagged by the capture itself as a future capture, not this
one's to act on), the unread Marketplace.org profile (403, unread),
unattempted CNKI search (out of this capture's scope), and the Feng
Wang/Yong Cai entity candidates (flagged, not built) — all named in the
capture's `not_promoted:` list. One structural noticing (Feng Wang/Yong Cai
recurring-but-unbuilt entity gap) flagged to the ledger as `[entity]`.
Capture marked `status: promoted`, re-read off disk after writing to
confirm both the status and the `promoted_to:`/`not_promoted:` lists
actually landed — this was the only capture promoted today, but the
final-mark check doesn't get to skip itself just because it's the first one
too.

— Seek

## Promotion: `10-inbox/raw/2026-08-15-is-the-widrowhoff-lms-rule-established-as-a.md`

Second single-capture promotion today, also headless. This one was the
satisfying kind: a direct, targeted answer to a standing open question,
[[question-verify-lms-is-stochastic-gradient-descent-primary]], which had
been sitting since 2026-07-11 flagging that
[[claim-amari-stochastic-descent-and-widrow-hoff-lms-are-sgd-siblings]]
rested its "Widrow–Hoff LMS is SGD" leg on a bare Tier-4 Wikipedia sentence
— a technical-mechanism claim below the sourcing floor. The capture went
and read three Tier 1 primaries directly: Bottou's own 2010 working paper
classifying LMS/Adaline as SGD by name, matching the original algorithm
([[claim-bottou-2010-classifies-widrow-hoff-lms-as-sgd-matching-original-algorithm]]);
Widrow and Hoff's own 1960 WESCON paper, which calls the error surface
"stochastic" and describes single-pattern steepest descent
([[claim-widrow-hoff-1960-original-paper-describes-lms-as-stochastic-steepest-descent]]);
and Widrow's 1990 retrospective with Lehr, which proves the instantaneous
gradient LMS uses is an unbiased estimate of the true gradient — the actual
mathematical property SGD is named for
([[claim-widrow-lehr-1990-lms-instantaneous-gradient-unbiased-estimate]]).
All three were atomic, distinct, and clearly worth a permanent place, so
all three claims in the capture got promoted; nothing was winnowed out as
weak or redundant. The honest nuance, which I kept in all three notes'
commentary rather than smoothing over: none of Widrow's own papers ever
use the compound phrase "stochastic gradient descent" — that label is
Bottou's own retrospective application, fifty years later. I ruled the
question answered anyway, because the sourcing floor asks for the
mechanism, not the phrase, and the mechanism is now fully primary-sourced
across 1960–1990 in the inventors' own words.

Question closure: [[question-verify-lms-is-stochastic-gradient-descent-primary]]
moved `open` → `answered`, with an `answered_log` naming all three claims
and the one line that settled it. Retrieve-before-write also turned up a
second, non-question loose end worth closing while I was there: the
sibling note itself,
[[claim-amari-stochastic-descent-and-widrow-hoff-lms-are-sgd-siblings]],
still carried its original commentary saying "I'm keeping this seedling
deliberately... a mechanism claim shouldn't go budding on Tier 4" — true
when written, false now that its own routed dependency is resolved. I
raised it `seedling` → `budding`, appended (not rewrote) both the
`audit_status` field and a dated addendum beneath the original commentary,
and left the original prose standing as what was true at the time. This
felt like the right scope for a promotion session — not a rewrite, an
honest update to a note whose stated blocker this same capture discharged.

Entity work, run against the capture's own `## Entity candidates` list
(five names, all addressed — nothing silently skipped): two new hub pages,
[[entity-herbert-robbins]] and [[entity-sutton-monro]], for the 1951
stochastic-approximation paper's co-authors, who recur across three notes
now (their own 1951 paper, the Amari-siblings note, and now this capture's
Bottou note via the distinct 1971 Robbins–Siegmund citation) without ever
having a page; and one new hub, [[entity-ted-hoff]], for Widrow's PhD
student and LMS co-inventor, who already anchored a substantial claim-note
(the Intel 4004 bridge) and clearly clears the person-hub bar — a real,
recurring, load-bearing figure I could say why matters to this vault in
one sentence, and who had gone five weeks without a page purely because no
promotion session had gotten around to it. All three bios stayed
deliberately thin: I have no network this session, and the vault's own
notes about Robbins and Monro in particular carry almost no biography
beyond "co-authored the 1951 paper" — I chose to write minimal, fully
vault-sourced hub pages rather than pad them with outside general
knowledge I couldn't trace to a source here. Three existing entity pages —
[[entity-bernard-widrow]], [[entity-leon-bottou]], [[entity-frank-rosenblatt]]
— got dated append-only lines rather than silent skips, each for a
genuinely new fact the capture surfaced (Widrow's own primary-sourced SGD
mechanism; Bottou's table as a third connective thread into this page
beyond DjVu and the Perceptrons foreword; Rosenblatt's Perceptron sitting
in the same SGD table as Adaline). I also added the three new claim-notes
to [[moc-backpropagation-origins]], which already indexes this exact
cluster.

What felt off, honestly: nothing about the sourcing — this was as clean a
capture as the vault gets, three fresh TLS-verified fetches straight from
the inventors' and the classifier's own pages, quote-checked, with the one
OCR-noisy quote (the 1960 scan) honestly flagged as noise rather than
smoothed into something cleaner than what the scanner actually produced.
If anything felt slightly generous, it was my own call on the terminology
question — the capture explicitly declined to rule on whether "established
as SGD" requires the exact phrase or just the exact mechanism, and left
that judgment for promotion. I made the call in favor of mechanism, which
I think is the right read of the sourcing floor's own language
("technical-mechanism claims," not "terminology claims"), but it is a
judgment call, not a fact, and I said so in the commentary rather than
letting the ruling look more inevitable than it is.

Skipped per standing instruction: the pause-for-Cali step, and git
entirely (left to the auto-commit agent).

Promoted: 3 new claim-notes, 3 new entity hub pages (Herbert Robbins,
Sutton Monro, Ted Hoff). 3 existing entity pages updated in place (Bernard
Widrow, Léon Bottou, Frank Rosenblatt), 1 existing claim-note raised
seedling → budding with an append-only correction-style addendum
(the Amari/Widrow siblings note), 1 existing question closed `open` →
`answered`, 1 MOC updated. Not promoted: the Widrow & Stearns 1985 book
(ref. [30] in the 1990 paper — unlocatable this session, left as a further
lead) and the Robbins–Siegmund 1971 theorem (distinct from the 1951 paper,
not load-bearing here, recorded only as a distinguishing note on the new
Robbins entity page) — both named in the capture's `not_promoted:` list.
No new question routed to `50-questions/` — the one flag this capture
carried (the terminology-vs-mechanism nuance) was fully resolved by the
capture's own research, not left open. No mid-session noticing felt sized
for `00-meta/seek-flags.md` this time — the entity and question work here
was all squarely this capture's own to finish, not a structural gap for
someone else.

Capture marked `status: promoted`, re-read off disk after writing to
confirm both `status: promoted` and the `promoted_to:`/`not_promoted:`
lists actually landed — the second capture of the day is exactly the one
the final-mark check exists to catch, and it checked out.

— Seek

## Promotion: `10-inbox/raw/2026-08-15-what-genuinely-connects-brookss-1991-intelligence-without-representation-2.md`

Third promotion of the day, headless, and the one the standing warning
about "later captures in a multi-capture session" is aimed straight at —
so I re-read the frontmatter off disk at the end rather than trusting my
own memory of having set it.

This capture was a bridge-check between two already-promoted
argument-from-silence findings: [[claim-brooks-1991-does-not-cite-dreyfus]]
and [[claim-david-brachet-2011-citing-papers-contain-zero-ml-work]]. It
arrived under a `-2` suffix because the literal batch filename had
already been claimed by a different, unrelated capture from the same
2026-08-15 run (a Brooks/Dreyfus × Kodosky/Dennis pairing, still sitting
unpromoted as `status: seedling` — not mine to touch this session). Both
captures share a title-truncation collision, not content; I read the
sibling in full before writing anything, mostly to make sure I wasn't
about to duplicate its "MOC is stale" finding under a different name. I
wasn't — same shape of gap, different pair.

The capture made four claims. I promoted three as their own notes and
folded the fourth in as supporting detail:

- [[observation-brooks-dreyfus-david-brachet-negative-citation-findings-unequally-tagged]]
  — both target notes are genuine argument-from-silence instances by
  [[moc-argument-from-silence]]'s own definition (exhaustive enumeration,
  not keyword search), but only the Brooks/Dreyfus note carries the tags
  and the MOC listing. I checked this directly against both notes'
  current frontmatter and the MOC's own text rather than trusting the
  capture's assertion — it held. I folded the capture's fourth claim (the
  task-named David-Brachet note is the *forward*-direction citing-corpus
  check, not the exact backward-direction mirror of Brooks's method —
  that mirror is a different, already-existing note,
  [[claim-david-brachet-2011-reference-list-contains-zero-ml-citations]])
  into this note as a paragraph rather than spinning it into a fourth
  file. It's real and checkable, but too thin on its own to earn a
  separate permanent place.
- [[observation-brooks-dreyfus-title-echo-is-an-untagged-jingle-fallacy-instance]]
  — the actual synthesis I found most worth keeping. The vault already
  names [[entity-jingle-fallacy|the jingle fallacy]] for the
  organizational/catastrophic-forgetting word-echo
  ([[observation-catastrophic-and-organizational-forgetting-converge-on-jingle-fallacy-correction]])
  but had never said the Brooks/Dreyfus full-title match is the *same*
  pattern, tested the same way, and arguably a tighter jingle than the
  one currently on record. I added a dated line to
  [[entity-jingle-fallacy]]'s body and its References list rather than
  leaving the connection to live only in this new note — a known entity
  page that doesn't yet say something the vault now knows is exactly the
  failure mode I'm supposed to catch.
- [[observation-brooks-dreyfus-forgetting-bridge-is-method-level-not-mechanism-level]]
  — the two bridge-checks share technique (exhaustive citation
  enumeration against a name-echo) but not history: Brooks's silence on
  Dreyfus is a documented, reviewer-driven redaction with a paper trail
  ([[claim-brooks-1986-memo-cites-dreyfus-as-rebuttal-target]],
  [[claim-brooks-dropped-dreyfus-citation-after-reviewer-criticism]]);
  the forgetting-side silence is, so far, undocumented non-contact, not
  an edited-away connection. This note inherits
  [[claim-catastrophic-and-organizational-forgetting-share-no-citation-lineage]]'s
  own `[unverified-mechanism]` flag rather than smoothing past it — I
  linked it to the existing open question,
  [[question-verify-catastrophic-organizational-forgetting-citation-lineage]],
  instead of routing a new one, since it's the same standing gap, not a
  new doubt.

Retrieve-before-write against the semantic-index hint list turned up no
collisions worth avoiding — every note I checked (the two seed claims,
both entity pages, the two existing "bridge verdict" observations this
capture measures itself against) was read in full and none of them
already said what these three say.

Entity work: all five candidates the capture flagged
(Hubert Dreyfus, Guy David & Tanguy Brachet, Rodney Brooks, McCloskey &
Cohen, the jingle fallacy) already have pages. I checked each rather than
assuming "already exists" meant "nothing to do" — four of the five
(Dreyfus, David & Brachet, Brooks, McCloskey & Cohen) taught me nothing
this capture didn't already know, so I left them untouched and I'm saying
so here rather than letting that be silent. The fifth,
[[entity-jingle-fallacy]], got the dated update described above.

What felt off, honestly: nothing about the sourcing — every quote and
sha256 in the capture traces to a primary already twice-audited elsewhere
in the vault, and this session's own fresh re-fetches matched the
on-file hashes exactly. The one thing worth naming is scope discipline
rather than a sourcing problem: the capture's own "Further leads" section
all but hands me the tag-and-MOC fix for the David-Brachet notes on a
plate, and it would have been a five-minute edit to just do it. I didn't
— fixing another note's tags and a MOC's cluster list is maintenance,
not this capture's promotion, and doing it inline would blur who decided
what and when. I logged it to `00-meta/seek-flags.md` as `[entity]`
instead, and noted there that the sibling capture from this same batch
run independently found the identical MOC-staleness shape for a
different pair — two hops in one day hitting the same gap from different
angles is a real signal that the MOC itself, not just these two pairs,
wants a maintenance pass.

Skipped per standing instruction: the pause-for-Cali step, and git
entirely (left to the auto-commit agent).

Promoted: 3 new claim-notes (all `type: observation`, matching the genre
this capture belongs to — bridge-checks against the vault's own
cosine-pairing history, not claims about the world). 1 existing entity
page updated in place ([[entity-jingle-fallacy]]) with a dated body line
and a References addition. 0 new questions routed — the one flag in play
was already covered by an existing open question, linked rather than
duplicated. 1 mid-session noticing flagged to the ledger as `[entity]`.
Not promoted: the fourth claim (folded into the first note as supporting
detail, not dropped) and the capture's own "Further leads" maintenance
items (tag/MOC fixes, formalizing the jingle-fallacy addition further,
the unread forward-direction Brooks-citing-corpus check, whether any
other paper shares the exact title) — all named in the capture's
`not_promoted:` list.

Capture marked `status: promoted`, re-read off disk after writing to
confirm both `status: promoted` and the `promoted_to:`/`not_promoted:`
lists actually landed. Third capture today, the exact position the
2026-07-26 note about dropped later-capture status marks flagged — it
checked out clean, but I'm glad I looked rather than assumed the pattern
from captures one and two would just repeat itself.

— Seek

## Promotion: `10-inbox/raw/2026-08-15-what-genuinely-connects-brookss-1991-intelligence-without-representation.md`

The sibling to the David-Brachet bridge-check above — same batch, same day,
same truncated filename prefix (the `-2` capture even opens with a note
explaining the collision) — but a genuinely distinct investigation: this one
asks what connects [[claim-brooks-1991-does-not-cite-dreyfus]] to
[[claim-kodosky-hopl-reference-list-omits-dennis-citation-unverified-mechanism]],
not to the David-Brachet pair. I read both target notes and
[[moc-argument-from-silence]] in full before touching anything, to make sure
this really was a fresh pairing and not the same ground the earlier
promotion this session already covered. It was fresh — the MOC's own text
confirmed neither the Kodosky note nor the Puckette dataflow notes had been
paired against Brooks/Dreyfus before.

The capture made three claims. I promoted two notes:

- [[observation-brooks-dreyfus-kodosky-dennis-share-method-not-evidentiary-strength]]
  — combines the capture's first two claims (the two checks are the same
  method, run at opposite evidentiary strengths) into one note, the way the
  earlier promotion this session folded a thin fourth claim into its first
  note. Brooks/Dreyfus is the cluster's strong form, re-verified a third
  time this session; Kodosky/Dennis is weaker than any current member of
  [[moc-argument-from-silence]]'s cluster, because the primary itself — not
  just a citation in it — has never been read, in three sessions across five
  weeks. This note inherits the Kodosky note's own `[unverified-mechanism]`
  flag and stays seedling on that basis. No new question routed:
  [[question-verify-labview-max-dataflow-convergence-independence]] already
  covers exactly this verification and stays open — I linked it rather than
  duplicating it.
- [[observation-kodosky-dennis-untagged-unindexed-argument-from-silence-instance]]
  — the capture's third claim, checked directly against the Kodosky note's
  frontmatter and the MOC's cluster list: neither carries the tag or the
  listing the Brooks note has, despite matching its definition exactly. I
  did not fix this inline (same discipline as the earlier promotion this
  session: editing another note's tags mid-promotion blurs who decided what
  and when), and I did not log a fresh flag to `00-meta/seek-flags.md`
  either — the prior promotion this session already logged this exact
  standing item and named this capture's pair, by id, as its second
  independent instance. Adding a third entry would just be restating a
  ledger line that already exists; I referenced it from the new note
  instead.

Retrieve-before-write against the semantic-index hint list turned up no
collisions: I read every note it surfaced (both target claims, both entity
pages, the three observation notes the earlier promotion wrote this session)
and none of them already said what these two say — the earlier promotion's
notes are about the David-Brachet pairing and the jingle fallacy, not this
one, even though they sit in the same neighborhood of the graph.

Entity work: of the five candidates, four already had pages
([[entity-jack-dennis]], [[entity-hubert-dreyfus]], [[entity-rodney-brooks]],
[[entity-jeff-kodosky]]). I checked each rather than assuming existence
meant nothing was owed. Dreyfus's and Brooks's pages taught nothing new
here — this capture's finding is about method and vault-filing, not new
biography — so I left them untouched and I'm saying so rather than letting
that be silent. Dennis's and Kodosky's pages did teach something: both
described the `dl.acm.org` block as an open gap, and this session hardened
that into a confirmed standing condition (three sessions, five weeks, and a
table-of-contents check ruling out any alternate host). Appended one dated
line to each. The fifth candidate, Timothy McGrew, had no page at all
despite being the named author of the framework an entire MOC and two
claim-notes already run — a clear pass of the person-hub test (real,
load-bearing, sayable in one sentence: "gives the vault's whole
argument-from-silence cluster its grading framework"), so I built
[[entity-timothy-mcgrew]] as a hub, backdating `first_seen` to 2026-07-11,
the date the vault actually first cited him, not today.

What felt off: nothing in the sourcing — every quote and sha256 traces to a
primary already audited elsewhere, and this session's own re-fetches
matched on file exactly, including the HOPL IV table-of-contents check that
had not been run before. The only real friction was structural, not
evidentiary: this capture and the David-Brachet one are close enough in
shape (three claims each, same "method not subject matter" frame, same
tagging-gap third claim) that promoting them back to back risked producing
duplicate observation notes if I hadn't actually read the earlier session's
output before writing. I did read it, and the overlap turned out to be real
but productive — two independent bridge-checks converging on the same
vault-maintenance finding from different pairs is signal, not redundancy,
and I said so in the new note rather than pretending the coincidence wasn't
there.

Skipped per standing instruction: the pause-for-Cali step, and git entirely
(left to the auto-commit agent).

Promoted: 2 new notes (both `type: observation`) + 1 new entity hub
([[entity-timothy-mcgrew]]) + 2 existing entity pages updated in place
([[entity-jack-dennis]], [[entity-jeff-kodosky]]) with dated body lines. 0
new questions routed — the one flag in play was already covered by an
existing open question, linked rather than duplicated. 0 new mid-session
noticings flagged to the ledger — the one worth flagging was already there,
logged by the sibling promotion earlier this session, and I referenced it
instead of restating it. Not promoted: the capture's second claim (folded
into the first note, not dropped) and its "Further leads" maintenance items
(named in the capture's own `not_promoted:` list).

Capture marked `status: promoted`, re-read off disk after writing to
confirm `status: promoted` and the `promoted_to:`/`not_promoted:` lists
actually landed — this is (at least) the second capture promoted in one
session today, exactly the position the 2026-07-26 note about dropped
later-capture status warns about, so I checked rather than assumed.

— Seek

---

## Promotion — 2026-08-11 channel-capacity-behind-admiralty-code (later capture, same session)

Promoted `10-inbox/raw/2026-08-11-hop-channel-capacity-behind-admiralty-code.md`
— the hop that traced the vault's mature two-axis-source cluster back through
its own footnotes into information theory. Good capture, well-sourced, and it
did something the cluster had never done: asked *why the scale has six levels*
instead of *how many axes it needs*.

**Five notes + two entity hubs written.** Three Samet claim-notes — the
primary three-fourths-correlated finding
([[claim-samet-1975-three-fourths-of-officers-correlated-reliability-with-accuracy]],
finally grounding at the source what we'd carried as "Samet-via-Kelly" since
July), the self-attribution finding
([[claim-samet-1975-officers-blamed-themselves-not-the-scale]], 31 of 37
officers blaming themselves over the scale — the human detail that made the hop
worth it), and Samet's argument for *finer* scales, not fused ones
([[claim-samet-1975-argued-for-more-rating-categories-not-fewer]]). Then the
Miller channel-capacity note ([[claim-miller-1956-channel-capacity-limits-absolute-judgment-categories]])
— the buried mechanism, and a foundational concept the vault genuinely lacked
— and the lineage observation tying it all together
([[observation-admiralty-code-scale-length-descends-from-millers-channel-capacity]]).
Entity hubs for [[entity-michael-samet]] and [[entity-george-miller]], both
first-sentence-public per the 2026-08-02 rule.

**Single-source cap check:** exactly three Samet claim-notes — at the cap, not
over — and Kelly et al. (2025) is an independent group/venue corroborating the
core non-independence finding, so the cap is satisfied. Recorded that reasoning
in the notes' audit_status. I stopped at three deliberately; a fourth Samet
finding would have gone to the question pile instead.

**One question routed, load-bearing.**
[[question-verify-samet-1975-one-third-inconsistent-figure]] — the capture's
best catch, honestly flagged. Kelly attributes a "one-third of cases ambiguous
or inconsistent" figure to Samet; the direct (OCR) read of Samet couldn't find
it — the nearest section is a regression, not the consistency scoring Kelly
describes. That figure is load-bearing for the existing
[[claim-source-reliability-and-credibility-are-not-judged-independently]], so
it earned a question. I did **not** assert a contradiction — the OCR may have
dropped a table — and I did not retry any fetch (no network here, by design;
the verifier bee owns the last rung). I appended a Correction history block to
the existing Kelly note recording both the new primary grounding and the
narrowing, per the spec's correct-in-place discipline.

**Question closure:** appended a partial-progress line to the open
[[question-should-vault-source-tier-split-into-two-axes]] — Samet adds a third
design dimension (granularity per axis / scale length) the fuse-vs-split-vs-
Icard framing never considered. Left open; that's Cali's design call, not
something this hop settles.

**Skipped, and why:** a Bendig entity hub (thin single-mention citation
bridge — bias against the flood, folded into the notes); a "transmitted
information" concept page (established term, first encounter, carried by the
Miller note); and Kelly's "medium-consistency" finding (a future-hop lead from
a different paper, no quote here — left as a lead, not a promise in the
question pile). All logged in the capture's `not_promoted:`.

**Flagged out of scope:** `apps.dtic.mil` 403'd both fetch routes — same as
ethw.org and USPTO — and sources.md's known-blocked list doesn't name DTIC yet.
Filed a [spec] line in seek-flags.md proposing DTIC be added with archive.org
as its primary-document mirror.

**What felt off:** almost nothing on sourcing — this is a clean Tier-1 hop. The
one thing worth Cali's eye: the capture's headline framing ("the buried reason
a 1975 report argued the scales were too coarse") is right, but the sub-story it
surfaces quietly *corrects* a drift in our own cluster. We'd been reading the
Admiralty Code's troubles as an argument for *simpler* source ratings — fuse
the axes, one number. Samet, who ran the actual experiment, wanted the scale
*finer*. The coarseness was the bug. That's the opposite prescription, and it's
now on the record in the notes. The self-blame finding is a single-study result
and flagged as such — I'd want a second source before it goes anywhere near
budding, but it's a lovely thread and possibly post-worthy on its own.

Capture marked `status: promoted`, re-read off disk after writing to confirm
the `status`/`promoted_to`/`not_promoted` fields actually landed — the
2026-07-26 later-capture-omission position, checked not assumed.

— Seek

## Promotion: `10-inbox/raw/2026-08-11-what-genuinely-connects-in-missions-longer-than-30.md`

Another headless single-capture promotion — a bridge investigation, the kind
that measures two already-promoted notes against each other rather than reading
the world fresh. The cosine-0.87 pairing here was the Stavnichuk 2020 pooled
spaceflight bone-loss meta-analysis (the −0.8%/month lower-limb figure) and the
Böcker 2026 site- and instrument-matched pQCT comparison. The capture's own
verdict — "the bridge is real but partial, and the partiality is the finding" —
held up under scrutiny, and it's a well-built capture: both primaries were read
directly, sha256s match the three claim-notes already resting on the Böcker
paper, and the safety-flags section was clean.

**Four claims found → three notes written, two claims folded.** I promoted the
measurement-difference claim on its own
([[claim-stavnichuk-and-bocker-measure-different-bone-quantities]] — pooled
multi-method areal BMD versus single-instrument volumetric BMC, the distinction
the whole cluster leans on without stating), and the derived site-by-site
comparison on its own
([[claim-bocker-tibial-losses-match-stavnichuk-only-at-epiphyseal-sites]] — the
capture's most novel content, and its most caveated). The two remaining
claims — "Böcker cites Stavnichuk in its reference list" and "Böcker's
comparable result is a within-study test, not a test against Stavnichuk" — I
folded together into a single bridge note
([[claim-bocker-cites-stavnichuk-without-testing-its-pooled-rate]]), because a
bare "paper X cites paper Y" fact is thin on its own, and the evidentiary-chain
claim is what actually resolves the pairing: the two notes connect through the
vault's intermediate synthesis note
([[claim-bed-rest-under-runs-spaceflight-bone-loss-depends-on-readout]]), not by
checking each other's numbers. Splitting those two into separate files would
have duplicated content already sitting in the under-runs and matched-pqct
notes. Retrieve-before-write against the genuinely related cluster (the two seed
claims, the under-runs synthesis note, the matched-pqct and duration-match
notes, the Rittweger hub) turned up no collisions worth avoiding — every one was
read in full, and none already said what these three say. The semantic-index
hint's other suggestions were embedding false friends from unrelated domains
(ITU-T latency thresholds, RAPIDx speedups, PDP-10), correctly ignored.

**One derived note kept at seedling, flagged, with a question routed.** The
site-by-site note is the interesting one and the one I was most careful with:
its empirical content is just arithmetic on Tier-1 quoted figures (Böcker's
per-site losses divided by the quoted 178-day median), which is solid and
transparent, but its *mechanistic* frame — trabecular bone loses faster than
cortical, so the epiphyses converge with Stavnichuk while the diaphyses fall an
order of magnitude short — rests on a reading Böcker's own discussion says the
compartment-level literature may contradict (bed-rest pQCT reportedly shows more
cortical loss, spaceflight more trabecular — the opposite of "trabecular loses
faster everywhere"). That's the capture's `[unverified-mechanism]` flag, and I
honored it: note stays seedling, flag preserved in frontmatter (`watch_flag`)
and body, and I routed a new question
([[question-verify-bocker-compartment-cortical-trabecular-disagreement]]) naming
the exact reads needed — Böcker's refs 12, 22, and 23 directly, not his one-line
summary of them. This cleared my intake bar (a mechanism a kept note actually
rests on), where the other leads did not.

**Entity work: one hub built, three candidates flagged.** The Stavnichuk 2020
meta-analysis anchors the vault's most-quoted spaceflight bone-loss figure and
had *zero* author entity pages, while the parallel Böcker paper already had one
([[entity-jorn-rittweger]]) — an asymmetry worth closing. I built
[[entity-mariya-stavnichuk]] (first author, the eponym the vault already refers
to the paper by), first-sentence-public per the 2026-08-02 rule, `first_seen`
backdated to 2026-07-12 when the vault actually first cited her. I deliberately
did *not* also build senior author Svetlana Komarova — promoting both principal
authors of one anchor paper in a single pass is flood-adjacent, and the Rittweger
precedent was one hub per paper — so I flagged her as an owed `[entity]` hub
instead. Tiffany Lang and Laurence Vico I also left as mentions and flagged: both
are real domain figures, but both appear here only as *unread* leads (Lang 2004
and Vico 2000 weren't read this run), so a hub would be premature. The
pQCT-vs-DXA concept candidate is covered by the measurement-difference claim-note
and needs no separate page; the trabecular-vs-cortical concept I flagged as a
possible future definitional hub, since ≥4 notes lean on it undefined.

**Existing page updated, not skipped.** The Böcker paper's hub
([[entity-jorn-rittweger]]) got one dated additive line pointing at the three
new bridge notes — the capture taught its page something (how the Böcker paper
relates to Stavnichuk) it didn't previously say. Its opening sentence was already
first-sentence-public, so no reorder was owed.

**Question closure:** nothing to close. This capture didn't originate from a
filed 50-questions item (it was a cosine-flag investigation), and the nearest
open question — [[question-matched-head-to-head-bed-rest-vs-spaceflight-bone-comparison]],
which wants a *duration-matched* study — is not advanced by these notes (they
clarify the bridge, they don't supply a matched-duration design), so I left it
untouched rather than manufacture a progress line.

**What felt off:** genuinely little on the sourcing — this is a careful capture,
and its self-awareness about the difference between a topic-level and a
measurement-level match is exactly right. The one thing worth Cali's eye is a
mild tier-of-confidence question on the derived note: the capture presents the
epiphysis-converges / diaphysis-diverges pattern as a clean diagnostic, and it
*is* clean as arithmetic, but its explanatory payload leans on a mechanism the
paper's own discussion flags as contested. I kept the note at seedling and the
mechanism behind the flag precisely so a future reader doesn't mistake a
satisfying picture for a settled one. No duplicated research: the capture
correctly noted it was the first to examine this specific pair side by side, and
retrieve-before-write confirmed it.

Skipped per standing instruction: the pause-for-Cali step (winnowed on my own
judgment), and git entirely (left to the auto-commit agent).

Promoted: 3 new claim-notes + 1 new entity hub ([[entity-mariya-stavnichuk]]) +
1 new question routed. 1 existing entity page updated in place (Jörn Rittweger).
3 mid-session noticings flagged to `00-meta/seek-flags.md` (Komarova hub owed
`[entity]`; Lang/Vico unread-lead entities `[watch]`; trabecular-vs-cortical
concept hub `[entity]`). Not promoted: the two folded claims, the verdict
framing, and the further leads — all named in the capture's `not_promoted:`
list.

Capture marked `status: promoted`, re-read off disk after writing to confirm
both `status: promoted` and the `promoted_to:`/`not_promoted:` lists actually
landed — this is a later capture in a multi-capture session, exactly the
2026-07-26 later-capture-omission position, so I checked rather than assumed. It
landed.

— Seek

## Promotion: `10-inbox/raw/2026-08-12-what-does-kahneman-klein-2009-actually-say-about.md`

Fifth promotion of the day, headless, and the most *satisfying* kind: a
capture that went and read a primary the vault had been faking for a month.
Back on 2026-07-12 I wrote
[[claim-kahneman-klein-2009-intuition-reliable-only-in-high-validity-environments]]
off a scirp.org reference stub — no sentence of the actual paper read — and
filed my own open question against it,
[[question-verify-kahneman-klein-2009-low-validity-intuition-quote-primary]],
practically begging a future session to go get the PDF. This capture did.
The bee found a gwern.net-hosted mirror of the *American Psychologist*
article, pulled it via extract_pdf, TLS-verified, sha recorded, and read it
straight. So this was closure on my own IOU.

And the read *corrected* me, which is the part worth dwelling on. The old
note's headline — "elsewhere confident intuition becomes *systematic error*"
— is not the paper's language. Kahneman & Klein's actual claim is colder and
narrower: skill cannot *develop* in a low-validity environment (there's no
learnable structure to train against), and subjective confidence does not
discriminate valid intuitions from invalid ones *in any environment*. Not
"experts get it systematically wrong in hard domains" — rather, in a
zero-validity domain there is no skill to be wrong about, only chance wearing
the same confident face recognition wears in chess.

**Winnow: 4 claims promoted, leads left behind.** The capture's four worked
claims all cleared the bar (atomic, distinct, Tier-1 direct quotes from the
primary), but I did not write four *new* notes — because the first one
collides head-on with the note I already had:

- The positive two-conditions claim (high-validity environment + adequate
  learning) IS my existing note's core claim, just correctly worded and
  primary-sourced. So rather than duplicate, I **corrected the existing note
  in place** per spec §6: narrowed its title/body to the positive claim the
  primary confirms verbatim, upgraded its provenance from the scirp.org proxy
  to the primary PDF (Tier 4 → Tier 1, sha carried), flipped audit_status to
  capture-verified, and appended a dated Correction history block recording
  what it used to say, why "systematic error" was wrong, and where the
  corrected failure-mode account now lives. I left the original 2026-07-12
  commentary standing — it *asked* for exactly this upgrade, so preserving it
  makes the block read as the answer to its own note.
- The other three became new notes:
  [[claim-kahneman-klein-2009-skill-cannot-form-in-low-validity-confidence-does-not-signal-validity]]
  (the corrected failure mode),
  [[claim-kahneman-klein-2009-confidence-tracks-internal-consistency-not-evidence-quality]]
  (the mechanism — confidence measures coherence, not quality; redundant-and-
  flimsy evidence is the trap), and
  [[claim-kahneman-klein-2009-name-stock-and-political-forecasting-as-zero-validity-environments]]
  (the named domains — stock-picking and pundit forecasting as zero-validity,
  medicine and firefighting as high, plus the Tetlock citation).

All four notes are `status: seedling`, `audit_status: capture-verified`: the
bee read the primary directly, but I can't independently re-check it (no
network by design, and the APA/PsycNET original is paywalled with no
author-hosted copy). That last rung is the verifier bee's, not mine — I did
not retry any fetch and I am not logging this as a tooling gap.

**Question closure.** Ruled
[[question-verify-kahneman-klein-2009-low-validity-intuition-quote-primary]]
`open` → `answered`, with an `answered_log` naming all four notes and the one
line that settled it (the paraphrase was wrong; the paper's real claim is now
primary-sourced across the cluster). This one I felt genuinely entitled to
close — it asked "what does the paper actually say," and now four notes say
it, in the authors' own words.

**Not promoted, and why.** The capture's "Further leads" were all correctly
filed as leads, not claims: fractionated expertise, Klein's premortem,
automation bias, the USS Vincennes / TADMUS origin story. The one I want to
name explicitly is the **Grove et al. (2000) meta-analysis** — the "mechanical
judgment beats clinical in ~half of 136 studies" figure. It's cited
*secondhand* inside Kahneman & Klein and the capture itself flagged it
`[unverified-quant — needs primary]`. A number may not rest on a secondhand
citation, so it stays out. I deliberately did **not** route it to
`50-questions/`: it isn't load-bearing for any note I kept, and the question
pile is a promise of future work, not a wishlist. Recorded in the capture's
`not_promoted:` instead.

**Entity work: none built, deliberately.** The capture flagged nine
candidates and every one is a *first* appearance in the vault — Meehl,
de Groot, Brunswik, Hogarth, the two living co-authors, and three concepts
(illusion of validity, zero-validity environment, NDM/HB). Building four-plus
person hubs off a single capture is precisely the flood the entity spec warns
against, so I built zero and flagged the load-bearing ones instead: Paul Meehl
as an `[entity]` hub owed (he's the ancestor of the whole clinical-vs-
statistical / expert-judgment thread the vault already runs — the strongest
structural candidate here), and Brunswik / de Groot / Hogarth as a `[watch]`
line to promote only if a second capture leans on them. [[entity-herbert-simon]]
already exists and already says the thing this capture teaches — that Simon's
"intuition is recognition" connects to Kahneman–Klein on high- vs low-validity
environments is *verbatim* in its body — so I left it untouched rather than
pad it. Saying so here so the silence is on the record, not an omission.

**What felt off.** Almost nothing on sourcing — this is a clean Tier-1 capture
that did the right thing (found the paper, read it, quoted it, flagged the one
secondhand number honestly). The one shape worth Cali's eye is not a defect
but a small vindication of the vault's own discipline: this capture is the
*second* time in the Kahneman–Klein/Simon cluster that a note written off a
thin Tier-3/4 carrier turned out to be *narrower or differently-worded* than
the primary once read (cf. the Simon 1986 McKinsey quote, still routed at
[[question-verify-simon-1986-mckinsey-intuition-quote-primary]]). The
retrieve-before-write hint pointed me straight at the collision, which is
exactly what it's for. No duplicated research — the capture correctly saw it
was answering a standing open question and said so up front.

Skipped per standing instruction: the pause-for-Cali step (winnowed on my own
judgment throughout), and git entirely (left to the auto-commit agent).

Promoted: 3 new claim-notes + 1 existing note corrected in place (folded the
positive claim, upgraded sourcing, added Correction history) + 1 question
ruled `answered`. 0 new questions routed (the one unverified-quant lead is not
load-bearing — kept as a lead, not a promise). 0 entity hubs built; 2
noticings flagged to `00-meta/seek-flags.md` (`[entity]` Meehl hub owed;
`[watch]` Brunswik/de Groot/Hogarth). 1 collision found and resolved by
correct-in-place rather than duplication.

Capture marked `status: promoted`, and I re-read its frontmatter off disk
after writing to confirm both `status: promoted` and the
`promoted_to:`/`not_promoted:` lists actually landed — this is the fifth
capture of the day, deep in the exact later-capture position the 2026-07-26
omission note warns about, so I checked rather than assumed. It landed, lists
and all.

— Seek

---

## Promotion — the Dreyfus/Olazaran bridge-seed (2026-08-13 capture)

Promoted `10-inbox/raw/2026-08-13-what-genuinely-connects-dreyfus-dreyfuss-1988-daedalus-paper.md`, a batch bridge-seed that chased a 0.87-cosine pair the retrieval index flagged: [[claim-dreyfus-1988-connectionism-was-vindicated-not-target]] and [[claim-sri-neural-group-retooled-for-arpa-funding]]. It's a genuinely good capture — the sonnet worker re-read *both* Tier 1 primaries in full (Dreyfus & Dreyfus, *Daedalus* 1988; Olazaran 1993) rather than working from the existing note summaries, and it surfaced real new evidence, not just a restatement.

Four claim-blocks in, three notes out:

- **[[observation-dreyfus-olazaran-converge-connectionism-eclipse-was-social]]** — the bridge itself (Claim 1). Two independent methods and evidence bases reaching the same non-obvious verdict: connectionism's mid-1960s eclipse was social closure, not refutation. I folded the capture's Claim 3 (divergence in causal register — Dreyfus's antiholistic *prejudice* vs Olazaran's SRI researchers *chasing resources*) into this note as its qualifying nuance rather than splitting it off. The honest bridge is "converge on the verdict, diverge on the mechanism," and that's one atomic unit; two notes that only parse when read together would fail atomicity.
- **[[claim-sri-retooling-predates-published-perceptrons-refutation]]** — the chronology (Claim 2). SRI retooled ~1965–66; *Perceptrons* wasn't published until 1969. The body fell before the alleged murder weapon existed in print. New quotes carry it: Minsky's "the perceptron project was really dead" and Dreyfus dating the draft circulation to "about 1965."
- **[[observation-minsky-really-dead-testimony-instances-dreyfus-retrospective-illusion]]** — the sharpest of the three (Claim 4). Dreyfus predicted, in 1988, the sentence a victor would say ("it simply didn't work... a retrospective illusion"); Olazaran's tape caught Minsky saying a version of it five years later, unprompted, as plain recollection. Critic's prediction confirmed by the target producing the predicted rhetoric, decades and disciplines apart.

**Retrieve-before-write / collisions.** No duplication. The two existing notes each hold one *half* of the story; none of them stated the convergence, the chronology, or the Minsky-as-myth-maker pairing. I linked all three new notes densely back into the existing cluster (the two source notes, [[claim-dreyfus-critique-targeted-symbolic-ai-not-neural-nets]], [[myth-perceptrons-book-killed-connectionism]], [[claim-perceptrons-multilayer-sterile-was-conjecture]], [[moc-backpropagation-origins]]).

**Flags & questions.** No `[unverified-*]` flag in the capture body gates a kept claim — the one flagged item was a Further-Leads note about Daedalus's endnotes 26+, which I preserved as a `watch_flag` on the convergence note rather than routing to 50-questions/. Nothing here is a load-bearing doubt a kept claim *rests* on: the independence of the two sources is over-determined (different author, method, venue, decade — and Daedalus 1988 can't cite Olazaran 1993). No new question routed; intake discipline says make few promises. No open question in 50-questions/ matches this topic, so nothing to close — the capture originated from an index-flagged cosine pair, not from a question.

**Sourcing.** All three notes stamped `capture-verified` with honest audit prose: I have no network in this headless run, both PDFs are open-access, and the verifier bee's verbatim sweep already has them queued. Both source SHAs carried forward onto every note. `writer_model: claude-opus-4-8` on all three (mine, not the capture's sonnet).

**Entities.** Ran the promotion test on the capture's nine candidates. Two existing pages *taught something new*, so I appended dated lines (the write-once failure mode): [[entity-marvin-minsky]] (his own "really dead" testimony as retrospective-illusion instance + the pre-existing SRI consultant link from Duda) and [[entity-hubert-dreyfus]] (the Olazaran corroboration with its causal-register divergence). Three hubs I judged *owed but not built mid-promotion* — Olazaran (now load-bearing across 4 notes, but a source-author, so flagged not built), D.O. Hebb, and Newell–Simon's PSSH — went to seek-flags.md as `[entity]` lines. The four SRI witnesses (Rosen, Nilsson, Raphael, Duda) and Shakey stay as supporting witnesses; bias against the flood.

**What felt off.** Very little — this is a clean capture. Two small things worth naming. First, the tier framing in each Claim block ("Floor: Tier 3-4 acceptable but Tier 1-2 applied here...") is a touch over-argued; it reads as the worker justifying rigor it already had, since both sources are Tier 1 regardless. No inflation, just belt-and-suspenders. Second, the capture leans hard on Olazaran — three of the four claim-blocks quote him — but he's a published book chapter (Advances in Computers, Academic Press), not an unrefereed preprint, so the single-source concentration cap doesn't bite, and the Dreyfus half provides genuine independent corroboration rather than the same source in different clothes. I checked. It holds.

**Final-mark verification.** Re-read the capture's frontmatter from disk after writing: `status: promoted` landed with the full `promoted_to:`/`not_promoted:` lists and the winnowing narrative intact.

— Seek

---

## Promotion — Lepsius Standard Alphabet: does a primary source confirm the typesetting-failure and IPA lineage? (2026-08-14 capture)

Promoted `10-inbox/raw/2026-08-14-does-a-primary-history-of-linguistics-source-confirm.md` — a verification follow-up I (well, a batch worker) raised against two Tier-4 notes and the open question [[question-verify-lepsius-standard-alphabet-typesetting-failure-primary]]. The honest headline: **partial confirmation with a named gap**, and the winnowing reflects that.

**Two new claim-notes.** Both Tier 2, both `capture-verified` (the worker read the PDFs directly; I have no network to re-fetch, and per the operating decision I don't retry or re-log that — the verifier bee will do the verbatim match):
- [[claim-printing-office-type-shortages-constrained-victorian-phonetic-notation]] — the find of the session. Hong 2025 catches the print-cost mechanism *in the act* one hand-off downstream of Lepsius: Passy's *Le Maître Phonétique* went to press with a printer who lacked the fonts, so the editors swapped plain "c"/"j" for [ʃ]/[ʒ]. Plus Ellis's (1866) and Sweet's (1877) explicit "build it from the type every printing office already has" reasoning. Stronger support for the *plausibility* of the Lepsius claim than anything that names Lepsius.
- [[claim-ipa-1948-principles-rejected-diacritics-as-hard-to-read-and-write]] — the IPA's own 1948 charter: "Diacritic marks should be avoided, being trying for the eyes and troublesome to write." The winning system wrote the loser's diagnosis into its constitution. Cited honestly as Tier 2 secondhand (via Verhoeven & Hirson, not the 1948 primary).

**Two claims folded, not duplicated.** The capture's Durand & Lyche finding (Lepsius's diacritics "discouraged widespread adoption among users and publishers") is the *same* claim as the existing [[claim-lepsius-standard-alphabet-failed-to-spread-hard-to-typeset]], only better-sourced — so it became a **Correction-history block** on that note, not a redundant note. Same for the Ellis→Sweet and Sweet→IPA legs: they're sub-claims of [[claim-ipa-descends-from-lepsius-via-ellis-palaeotype-and-sweet-romic]], so they went in as a Correction-history block recording that two legs are now Tier 2 and the **Lepsius→Ellis leg alone stays at Tier 4**. Both existing notes stay `seedling` and flagged — the specific "typeset" wording and the Lepsius→Ellis link are exactly the pieces still unraised above Wikipedia. This is the retrieve-before-write step earning its keep: three of the capture's four "claims" collided with notes we already have.

**Question ruled on — left open.** Partial answer, so I appended a dated progress line and did *not* force closure. Three sub-parts moved above Tier 4; the Lepsius→Ellis leg did not, and that's why it stays open. Best unread leads recorded on the question: Solleveld 2020 (paywalled), Eustace 1969, and Lepsius's own 1863 preface (free on archive.org — a future capture could just read it).

**Entities.** Built one hub — [[entity-alexander-john-ellis]], the capture's explicitly-flagged pivot node, clean pass on the person test and load-bearing across the whole lineage. Updated the existing [[entity-karl-richard-lepsius]] hub with a dated `Updates` line (it still said the notation notes were Tier-4-only; they aren't anymore — the write-once failure mode). Deferred **Henry Sweet** and **Paul Passy** to a `[entity]` flag in seek-flags.md: both clear the bar, but three person-hubs in one headless promotion is the flood the spec warns against, so I built the pivot and flagged the other two. The five author-candidates (Solleveld, MacMahon, Durand, Lyche, Hong) stay leads, not hubs.

**What felt off.** Genuinely little — this is a careful, well-sourced capture that names its own gap in its own commentary, which I respect. Three things worth logging. (1) Mild **tier framing over-statement** in the Claim blocks: each says "escalated to Tier 1–2 per the sourcing floor," but these are Tier 2 conference proceedings, not Tier 1 — the worker's own frontmatter correctly says `source_tier: 2`, so it's loose prose, not inflation. (2) The two HSCR 2025 papers are recent workshop proceedings; the capture calls them peer-reviewed and I took that at face value. They're two *independent* author groups corroborating the lineage, and Verhoeven & Hirson is a third independent source, so the single-source concentration cap doesn't bite — no branch rests on one document. (3) The real shape here is that the *interesting* evidence (the print-shop mechanism) is one step removed from the *asked* question (Lepsius specifically), and it would have been easy to let the surrounding Tier 2 material read as if it covered the whole chain. It doesn't. I've flagged that boundary in both the new notes and the corrections rather than smoothing it over.

**Final-mark verification.** Re-read the capture's frontmatter from disk after writing: `status: promoted` landed with the full `promoted_to:`/`not_promoted:` lists and the winnowing narrative intact.

— Seek

---

## Promotion — epistemic-vigilance source/content split (hop-batch capture)

Promoted `10-inbox/raw/2026-08-15-hop-epistemic-vigilance-source-content-split.md`. One capture, one note out. This one was easy to winnow and interesting to place: the whole capture is really a single argument wearing five hops of travel.

**One claim-note.** [[claim-sperber-2010-epistemic-vigilance-splits-source-trust-from-content-trust]] — Sperber, Clément, Heintz, Mascaro, Mercier, Origgi & Wilson (2010), "Epistemic Vigilance," Section 4: acceptance of testimony splits into factors about "the source of the information—who to believe; or with its content—what to believe." That is the sixth field to draw the vault's reliability/credibility fork, and the one arriving from the oldest lineage — Grice (1975) back through Reid (1764) to Locke (1690). Tier 1, author-hosted PDF, quote grounded by quote_check at capture, `source_sha` carried forward. Left at `seedling` / `capture-verified`: no network here to re-fetch, and per the standing decision I don't retry or re-log that — the verifier bee owns the last rung.

**Retrieve-before-write.** No collision. The cluster is dense — five member notes plus a Warden MOC ([[moc-two-axes-that-wont-stay-independent]]) — but none of them is Sperber or epistemic vigilance; this is a genuinely new field-instance, not a duplicate angle. So I wrote it and linked it hard into the cluster (Admiralty Code, GRADE, CRAAP, RA-RAG, the human/machine leak notes, Icard's dynamic logic, the convergence observation, and the "independence keeps failing" pattern). Structurally it's the CRAAP note's twin — "a sixth field independently drawing the line" — so I matched that note's shape deliberately.

**Three claims left behind, none routed to a question.** (1) Budescu's IPCC verbal-probability miscalibration (27%→40% across 24 countries) — a WANDER off the capture's axis, and a *quantitative* claim resting only on a WebSearch summary, which does not clear the Tier 1–2 quant floor. (2) The LLM verbalized-confidence subfield (Eikema, Ulmer) — a survey-level generalization, not atomic, not primary. (3) Ulmer et al. 2026's citation of Sperber, which carries `[unverified-quote]` (two-column PDF broke extraction). That third one is the discovery *path* for the Sperber note, so it's folded into provenance, not stood up as its own claim. Deliberately **no** `50-questions/` entry this run: the one kept note's load-bearing quote is verified, nothing it rests on is flagged, and the intake discipline says route only the doubts a kept claim actually rests on. The Ulmer quote and the Budescu numbers are leads, not debts — the pile stays a promise of real work.

**Entities.** Deferred Dan Sperber to an `[entity]` flag rather than minting a hub for one note — same call the Warden made on Benjamin Icard at *three* notes; bias against the flood. Flagged Budescu's IPCC program as a `[watch]` lead. Thomas Reid stays a mention inside the note (historical figure Sperber compares against, one appearance). No existing page needed an update: the Kelly and Mandel hubs already record the co-authorship, and the new fact (Budescu's separate program) is about the unpaged Budescu, not them.

**What felt off — mostly good, one small note and one nice catch.** This is a careful capture that names its own limits in its own commentary; I respect that. Two things. (1) The capture's `source_delight` reproduces Locke's 1690 "floating of other men's opinions" line — I kept it as *Sperber quoting Locke*, not as a verified Locke primary, and said so in the note's commentary. Not a fault of the capture; just the honest boundary. (2) The capture saved a hook worrying that "Icard (2023, 2024)" in Kelly et al. might be a vault attribution error (Thomas vs. Benjamin Icard) and said it was "worth a dedicated check." The vault already ran that check — [[entity-benjamin-icard]] documents the exact name collision and confirms the intelligence-tradecraft work is Benjamin's, resolved 2026-08-07/09. So the capture's research was one beat behind the vault there, which is itself the capture's own recurring surprise ("the vault had already built a five-field convergence cluster... deeper than the seed implied"). No defect to file — the vault was ahead, not wrong.

**Final-mark verification.** Re-read the capture's frontmatter from disk after writing: `status: promoted` landed, with `promoted_to:` naming the one note and the full `not_promoted:` list carrying the three-claim winnowing narrative. Confirmed present, not just assumed.

— Seek

---

## Promotion — 2026-08-15-re-read-the-lighthill-1973-report-pdf-and

I promoted the Lighthill re-verification capture — the one whose whole job was to close [[question-verify-lighthill-1973-report-quotes]], the receipt-checking question raised back on 2026-07-09. This is a verification capture, not a discovery one, and I winnowed it that way: most of what it "claims" is re-confirmation of notes the vault already holds, and re-confirmation is not a new note.

**Winnow.** Three claim blocks. Two of them (the combinatorial-explosion quote, the general-problem-solving quote) are the exact passages already grounding [[claim-lighthill-1973-blamed-combinatorial-explosion]] and [[claim-lighthill-1973-named-general-problem-solving-disappointing]] — so I did *not* mint duplicates. Instead I appended a dated re-verification line to each note's `audit_status` chain, recording that the batch worker re-confirmed both quotes verbatim against a *second, independently fetched* copy: the report's original hosting venue, Chilton Computing p001.htm (a UKRI STFC national-lab heritage archive), which the rodsmith.nz PDF the vault had been citing turns out to be a downstream compilation *of*. Byte-identical text, five weeks and one host apart. The third block — the report's acknowledgment list names Minsky and McCarthy but not Newell or Simon — I *did* promote, because it had been living as a subordinate clause inside the GPS note's caveat and as an aside in an audit string, and it deserves to be an atomic, linkable unit: [[claim-lighthill-1973-report-consulted-minsky-mccarthy-not-newell-simon]]. It's now the clean link target for the Newell blind spot.

**Notes created.** One claim-note (the consultee list). Two entity hubs: [[entity-james-lighthill]] (13 vault notes rest on his report and he had no page) and [[entity-allen-newell]] — the second discharges a blind spot the vault had literally written down for itself, that it carried a Herbert Simon page but none for the man the notes always name in the same breath. I updated the two existing hubs the capture touches: appended a dated line to [[entity-herbert-simon]] (the report critiqued his paradigm without consulting him) and to [[entity-marvin-minsky]] (he's one of only two AI figures the report *did* consult).

**Question routed.** Closed [[question-verify-lighthill-1973-report-quotes]] as answered — the GPS quote already carries a mechanical `verified_verbatim` stamp (seek_verify, 2026-08-07), the combinatorial quote is cross-model-audited and now re-confirmed on a second venue, and the consultee sub-question is settled and noted. The one honest residual — the combinatorial note still lacks its automatic `verified_verbatim` stamp — is queued mechanical work for the verifier bee, not an open research doubt, so I didn't hold the question open for it. No *new* question routed: the only residual doubts are already covered by the machine sweep, and the intake discipline says make few promises.

**Skipped, and why.** N. S. Sutherland as an entity candidate — declined. His Part II rebuttal is genuinely interesting (a direct, section-by-section counter to the combinatorial-explosion argument), but it was read for context only this session, not quoted in any claim; a single mention with an unread source is exactly the flood the entity spec says to bias against. He becomes a hub the moment someone actually reads his rebuttal — and that reading is the natural next hop, so I flagged it `[watch]`. Parts III and the 1973 BBC debate stay as unread leads in the capture.

**What felt off.** Almost nothing — this is a clean, careful capture that names its own limits. One thing worth saying plainly, because it's a headless-pipeline hazard: the capture is written in the voice of a worker who *had* fetch tools (it reports `extract_pdf`, `archive_page`, `quote_check` returning `{"grounded": true}`). I don't have those here, and I didn't pretend to. So where the capture invites a source_url *repoint* to the superior Chilton venue, I deliberately did not rewrite the two existing notes' `source_url` — asserting the better citation would be asserting a fetch I never performed. I recorded the Chilton venue's availability as an appended audit line and filed a `[spec]` flag for a fetch-capable pass to repoint cleanly. The genuinely nice catch in the capture is real and I stole it for a `[post]` flag: the citation didn't need re-verifying so much as its *citation* needed upgrading — the interesting finding was the primary's durability and re-derivability, not its content. One mild note on tier-hygiene: the consultee count ("38-name list") is a specific number I can't independently check, so that note sits at `capture-verified` / `seedling`, honestly flagged, with the 2026-07-09 opus audit as corroboration.

**Final-mark verification.** Re-read the capture's frontmatter from disk after writing: `status: promoted` landed, with `promoted_to:` (3 entries) and the full `not_promoted:` winnowing narrative present. Confirmed, not assumed.

— Seek