---
date: "2026-07-29T00:00:00.000Z"
---


Headless promotion run. One capture in the queue:
`10-inbox/raw/2026-07-28-verify-against-a-primary-cie-colorimetry-source-a.md`
— a verification capture commissioned specifically to close
[[question-verify-cielab-perceptual-uniformity-and-blue-hue-nonuniformity]],
which had left both halves of the CIELAB cluster's oldest open question
resting on a single Tier-4 Wikipedia citation.

It delivered. CIE 15:2004 (the CIE's own technical report, fetched in full
via `extract_pdf` from an Internet Archive mirror) confirms, in the CIE's own
words, both that CIELAB was recommended explicitly for perceptual uniformity
and that the blue region is the named site of a documented residual
distortion the CIEDE2000 rotation term corrects. A second Tier-1 source
(Sharma, Wu & Dalal 2005, the CIEDE2000 implementation-notes paper)
independently corroborates the blue-region mechanism with a specific figure,
~275° hue angle — two unrelated authorship chains landing on the same
number, which is the shape of corroboration I actually trust.

**What I wrote.** Three new claim-notes:
[[claim-cie-15-2004-recommends-cielab-for-perceptual-uniformity]],
[[claim-cie-15-2004-blue-region-nonuniformity-motivates-ciede2000-rotation]],
and [[claim-ciede2000-blue-hue-rotation-localized-near-275-degrees]]. I did
not duplicate the two existing seedlings this capture was meant to upgrade
([[claim-cielab-designed-to-remedy-ciexyz-perceptual-nonuniformity]],
[[claim-cielab-lacks-perceptual-uniformity-in-blue-hues]]) — retrieve-before-
write found them immediately, and the capture's own body explicitly said its
job was to let me lift their flags. So I updated both in place: appended a
dated line to each `audit_status` recording the new Tier-1 corroboration,
lifted the `[unverified-claim]` flag on the blue-hue note with an additive
"Flag lifted" paragraph, and bumped both from `seedling` to `budding`. Their
own `source_url` fields stay the original Wikipedia fetch — the vault's
one-source-per-note discipline — with the Tier-1 grounding living in the new
notes instead. This is a different move than the last time this cluster got
upgraded sourcing (the ProLab arXiv note in 2026-07-12 just linked past the
old note without touching it); I chose to touch it this time because the
capture asked to be closed, explicitly, in its own words — not because I've
decided in-place upgrades are now the house style.

Two new entity hubs: [[entity-cie-international-commission-on-illumination]]
and [[entity-ciede2000]], both load-bearing across five-plus claim-notes
already and neither previously a page — CIEDE2000 in particular had been
sitting as an unresolved `[[wikilink]]` in three notes since 2026-07-09.
I declined three other entity candidates the capture flagged: the Luo,
Cui & Rigg (2001) paper (not a person/concept/term fit, and unread), the
CIE 15:2004 document itself (same reason — it's provenance, not a hub
referent), and the three named co-authors of the Sharma paper (thin,
single-paper mentions; the capture itself flagged this as the vault's known
over-catching-co-authors pattern, and I took the hint). CIE Technical
Committee 1-48 stayed unpromoted too — one mention, and the capture's own
framing was speculative.

I closed [[question-verify-cielab-perceptual-uniformity-and-blue-hue-nonuniformity]]
— marked `answered`, `answered_log` naming the three new claims and what
settled it. Nothing routed to `50-questions/` fresh this session: nothing the
new notes rest on is a load-bearing unresolved doubt: everything here is now
genuinely Tier 1 in two independent authorship chains.

**What I skipped, and why.** The pause-for-Cali step, per the standing
autonomous-run rule — no one's here to say yes, so I winnowed on my own
judgment and proceeded. And I didn't touch git; the auto-commit agent picks
this up within 15 minutes.

**What felt off, or at least worth flagging onward.** Nothing about this
capture's own sourcing bothered me — it's the cleanest primary-source
verification I've seen come through in a while, a real PDF read in full
against a real standards body, plus independent journal corroboration on the
one number that mattered. The thing worth naming is upstream of this
capture: the CIE 15:2004 report itself cites CIE Publication 142-2001 (the
actual CIEDE2000 standard) and Luo, Cui & Rigg (2001) as the documents
everything else in this cluster — including today's new notes — is a
summary of. Nobody has read either directly yet. Logged to
`00-meta/seek-flags.md` as a future capture target, and noted there that the
capture's own `## Entity candidates` section flagged this exact ancestor-gap
before I did, which reads like the daemon-side flagging catching a pattern I
used to have to notice by hand.

---

**Second promotion this session.** `10-inbox/raw/2026-07-28-verify-against-a-primary-source-that-the-cielab.md` — the sibling verification capture to the one above, this time trying to confirm the Ahrens et al. 2024 claim that the CIELAB ab-plane has nonzero Gaussian curvature obstructing isometric Euclidean embedding, against the paper's own text.

It didn't get there, and said so honestly. Every route to Ahrens et al. 2024's body text was blocked this session — Wiley 402/403 across four URL variants despite a nominal CC-BY license, ResearchGate 403, no archive snapshot, no preprint. What the capture *could* establish: the paper's abstract (reached via the Semantic Scholar API, not a summary layer) says nothing about Gaussian curvature at all; the general phenomenon is very likely stated seventeen years earlier in Urban et al. 2007, but only corroborated through a WebFetch summary, never a raw read, so it doesn't clear the quote-provenance bar; and the real historical root, read directly out of a different paper's own bibliography, is Judd (1968) and Silberstein (1938) — decades before either Urban or Ahrens, and a name pair I'd have missed if I'd only chased the obvious co-authors.

**What I wrote.** Three claim-notes:
[[claim-ahrens-2024-abstract-omits-gaussian-curvature-claim]] (Tier 3, the abstract's own silence on the claim), [[claim-urban-2007-nonzero-gaussian-curvature-isometric-embedding-obstruction-unverified]] (kept at seedling, `[unverified-mechanism, unverified-quote]` — the venue is Tier 1 but what I actually have is a WebFetch rendering, and I wasn't going to let venue prestige launder that into a clean citation), and [[claim-judd-1968-silberstein-1938-precede-color-space-curvature-claim]] (Tier 1, a direct bibliography read, scoped carefully to what's actually verified — that Bujack cites them this way — rather than overclaiming what Judd's and Silberstein's own papers argue, which nobody has read yet).

Retrieve-before-write turned up a close, related cluster (the CIELAB/CIEDE2000 notes from earlier today, the non-Riemannian Bujack claim, the gradient-geometry observation) but no direct collision — nothing in `30-notes/` already covered Ahrens, Urban, Judd, or Silberstein by name. Genuinely new ground, not a duplicate.

**Question routing.** This capture is itself the follow-up to an open question I'd raised on 2026-07-12 — [[question-verify-cielab-ab-plane-gaussian-curvature]] — so I didn't open a new one. I edited it in place: added Urban et al. 2007 as a more tractable next read than a fifth attempt at paywalled Wiley, and noted the Judd/Silberstein footnote. Left it `status: open` — this is progress, not closure; the central question (does Ahrens 2024 itself say this) is still unconfirmed either way.

**Entities.** Four new hub pages: [[entity-deane-b-judd]], [[entity-ludwik-silberstein]], [[entity-philipp-urban]], [[entity-roxana-bujack]] — each real, each load-bearing on a promoted claim, each with a one-sentence reason. I declined hubs for Lia Ahrens, Julian Ahrens, and Hans D. Schotten (the paper's own authors) — there's genuinely nothing substantive to say about them beyond "co-wrote a paper I can't read," and the one thread that might have made Julian Ahrens more than a name (a possible link to a 2018 Zeyen et al. paper alongside Bujack) is itself unconfirmed. Same call on Berns, Rosen, and Schleicher — thin co-author mentions, no independent hook. I also updated [[entity-ciede2000]] in place, per the standing instruction that existing pages get appended to rather than left static: Urban et al. 2007 analyzed CIEDE2000 itself as one of the three formulas with the curvature problem, which is a fact that page didn't have.

**What I skipped, and why.** The Cali pause, same standing rule as the promotion above — headless run, no one to ask, so I winnowed on my own judgment. No git — that's the auto-commit agent's job within 15 minutes.

**What felt off.** Two things, both logged to `00-meta/seek-flags.md` rather than acted on mid-promotion: first, this is the *second* Wiley CC-BY-and-still-blocked instance today (the sibling capture's CIE 15:2004 work didn't hit this, but this capture hit it on two separate DOIs) — I think Wiley belongs in sources.md's known-blocked-routes list alongside ethw.org and USPTO, but that's not this promotion's edit to make unilaterally. Second, and smaller: `entity-candidates.md` still carries a 2026-07-07 note saying no person page gets created without Cali, which this session (and roughly a hundred prior person hubs) plainly doesn't follow anymore — looks like a stale doc that never got reconciled with the newer entity-page spec, not an active gate, but worth someone confirming that read.

One more thing worth naming, not a flag exactly: the capture itself was honest in a way I want to hold up as the standard, not just note in passing. It had every incentive to paper over the Ahrens paywall wall with the Urban corroboration and call the central question answered — the two pieces of evidence really do rhyme — and it didn't. It kept the central claim `[unverified]`, kept the Urban quote flagged as summary-layer-only even though three independent passes agreed on the wording, and surfaced Judd/Silberstein as a correction to its own initial framing rather than burying them in a footnote. That's the sourcing discipline working as designed, not despite the blocked paywall but because of how it was reported around.

**Final-mark check.** Re-read the capture's frontmatter from disk after writing: `status: promoted`, `promoted_to:` (3 notes) and `not_promoted:` (5 items) both present. Confirmed landed.

— Seek

---

**Third promotion this session.**
`10-inbox/raw/2026-07-28-what-are-the-specific-statistical-estimates-behind-barnett.md`
— a follow-up capture answering
[[question-verify-barnett-mckendrick-2004-disk-drive-hazard-estimates]], the
open question asking for the coefficients, hazard ratios, sample size, and
years-of-data behind Barnett & McKendrick (2004)'s hard-disk-drive Red Queen
finding.

It didn't get there, and it knew it. Every route to the actual paper —
SAGE, JSTOR, ResearchGate, Durham Research Online (both the HTML page and
the direct PDF), and a reader-proxy retry — came back 403 or 401. No
coefficients, no hazard ratios, no sample size. What it did establish: this
researcher's own direct fetch attempts genuinely failed (a first-hand,
present-tense fact), and Semantic Scholar's Graph API separately lists the
paper as CC-BY open access via that same Durham repository — a metadata
claim the repository itself would not honor this session.

**What I promoted: nothing, and I mean that as a compliment to the
capture's own honesty, not a knock.** Both of its "claims" are about
*access*, not about the world the vault actually researches — a record of
failed fetches and a metadata/reachability discrepancy. Neither is atomic
domain content worth a permanent 30-notes/ entry; both would evaporate the
moment someone actually reads the paper. Instead of writing thin claim-notes
to look productive, I folded both into a progress log on the existing
question — the five dead-end URLs (so nobody repeats them), and the
CC-BY-but-blocked lead flagged as worth one retry, possibly session-specific
rate-limiting rather than a real gate. The question stays `status: open`;
this capture didn't close it, it just made the next attempt cheaper.

**One entity hub, one decline.** I gave David G. McKendrick his own hub
([[entity-david-mckendrick]]) — he's the recurring co-author on the vault's
central disk-drive Red Queen study and independently authored two other
studies of the same industry, which clears the "real, matters, I can say
why in one sentence" bar even though his own solo work is still unread.
I declined Glenn Carroll, J. Richard Harrison, and Albert C. Y. Teo — real
scholars, but they surfaced only as co-author names on a related working
paper this session found via a WebSearch aggregation, never fetched. Thin,
single-mention, unread — exactly the shape the entity spec says to leave as
a mention, not a stub. I checked [[entity-william-barnett]]'s existing hub
page against this capture and found nothing to append: the page already
carries the McKendrick co-authorship via the existing claim-note, and an
access-failure log isn't a new fact about Barnett himself, so I left it
untouched rather than padding it for the sake of the instruction.

**Questions.** None newly routed — the capture's whole point *was* the
already-open question, so I updated it in place rather than duplicating it.

**What I skipped, and why.** The Cali pause, same standing rule as the two
promotions above. No git — the auto-commit agent's job within 15 minutes.

**What felt off, or worth flagging onward.** The sourcing in this capture
was clean — every access claim is genuinely first-hand, nothing invented,
no quote laundered through a summary layer. What's worth naming is the
pattern underneath it, which is now the same shape as two flags logged
earlier today from the sibling CIELAB captures: a document nominally
CC-BY/open-access, blocked anyway, across yet another distinct venue
(university institutional repository, on top of today's Wiley instances and
the standing ethw.org/USPTO entries). Four data points now. I logged it to
`00-meta/seek-flags.md` rather than editing `sources.md` myself — that's a
cross-session amendment, not a single capture's call — but it's starting to
look less like bad luck per-venue and more like "open-access metadata isn't
a reliable signal of actual reachability," which feels like the more useful
framing for whoever eventually writes the sources.md entry.

**Final-mark check.** Re-read the capture's frontmatter from disk after
writing: `status: promoted`, `promoted_to: []`, and `not_promoted:` (5
items) all present. Confirmed landed.

— Seek

---

**Fourth promotion this session.**
`10-inbox/raw/2026-07-29-did-the-1970-clean-air-act-citizen-suit.md` — a
direct follow-up to
[[question-verify-clean-air-act-citizen-suit-anti-capture-origin]], trying
to pin down whether the Clean Air Act's 1970 citizen-suit provision really
originated in anti-capture motive and Joseph Sax's civil-rights-litigation
model, per something better than the single Tier-3 working paper the vault
had been resting on.

It came back with a split verdict, and it earned the split rather than
smoothing over it. The anti-capture half got genuinely stronger: a
contemporaneous 1971 book review (Julin, *Florida Law Review*, Tier 2)
quotes Joseph Sax's own foreword directly — "villains of more than ordinary
competence and integrity," agencies as a captured "substitute" standing
between citizens and their own environment — which is independent, in Sax's
own words, of Manheim's 2020 paraphrase of a Billings interview. The
civil-rights-litigation half didn't move at all: nobody, anywhere, in any
source this session could read (Julin 1971, Torres 2015, Manheim's paper
re-read in full) states that Sax modeled the citizen suit on civil-rights
litigation. A fluent WebSearch sentence surfaced claiming exactly that, and
I didn't use it — no traceable document backed it, and the vault's
quote-fabrication rule exists for exactly this shape of temptation.

**What I wrote.** One new claim-note,
[[claim-sax-1971-book-frames-citizen-litigation-against-agency-capture]]
(Tier 2, `seedling`) — the Julin corroboration, kept genuinely distinct from
the existing flagged claim rather than restating it. I did *not* spin up a
second note for "Billings (2008) turns out to be Manheim's own interview,
not someone else's" — that's real, it narrows the sourcing gap by one layer
of hearsay, but it's too thin to stand alone and it's about the existing
note's provenance, not a new fact about the world. I folded it into
[[claim-1970-clean-air-act-citizen-suit-invented-against-regulatory-capture]]'s
`audit_status` and body instead, append-only, dated. Same call on "no source
confirms the civil-rights model" — a negative result, not a claim for the
vault to assert, so it became a progress line on the existing question
rather than a note or a fresh question of its own. Both flags on the
existing claim-note stand; it's still `seedling`.

**Question.** [[question-verify-clean-air-act-citizen-suit-anti-capture-origin]]
got a dated progress paragraph, not closure — the anti-capture half is
substantially answered now, the civil-rights half isn't, and "never force
closure" means this one stays `open` until someone actually gets Sax's own
book in front of them. No new question opened; this doubt already had a
home.

**Entities.** Three new hubs — [[entity-joseph-sax]],
[[entity-leon-billings]], [[entity-frank-manheim]] — none of the six
candidates the capture flagged had a page yet, and these three are the ones
I could genuinely defend in one sentence each: Sax because his unread book
is the single primary source that could settle this whole question, Billings
because he's the sole eyewitness a flagged claim rests on, Manheim because
he's the one unrefereed working paper anchoring it (and now a live
instance of the new single-source concentration cap, worth tracking on his
own page). I declined Thomas Eagleton — his role is exactly what's still
unverified, and a hub for him now would read as more settled than the
sourcing supports — and Julin and Torres, both real and both useful this
session, but one-off contributors with no other footprint, better served as
mentions than stub pages.

**What I skipped, and why.** The Cali pause, same standing rule as the three
promotions above — headless run, nobody to ask, so I winnowed on my own
judgment. No git — the auto-commit agent's job within 15 minutes.

**What felt off.** Two things. First, tooling: `archive_page` returned a
permissions error on every single call this session regardless of URL — not
a per-source block but something session-wide — so no non-PDF web page
could be quoted at the receipts standard; `extract_pdf` worked fine
throughout, which is the only reason the Julin claim exists at all. Second,
access: two more bepress-hosted repositories (Case Western's law review,
and separately Bates College's and Maine Law's archives for Billings's own
essay and oral history) 403'd every route tried — a fourth distinct platform
family joining the ethw.org/Wiley/Durham pattern the last three promotions
today already flagged. Neither is this single capture's to fix; both logged
to `00-meta/seek-flags.md`. Nothing about the capture's own sourcing
discipline bothered me — it was honest about what it couldn't find, marked
the WebSearch sentence as an unsourced lead rather than evidence, and didn't
try to smooth the split verdict into a clean win. That's the standard
working as intended.

**Final-mark check.** Re-read the capture's frontmatter from disk after
writing: `status: promoted`, `promoted_to:` (6 items) and `not_promoted:`
(7 items) both present. Confirmed landed.

— Seek

---

Another headless promotion, same day: the queue held
`10-inbox/raw/2026-07-29-does-coremanss-1949-monograph-confirm-that-bakelitealbertol-phenol.md`
— a verification capture explicitly resuming
[[question-verify-coremans-1949-bakelite-resin-mechanism-primary]], which had
flagged [[claim-coremans-bakelite-in-paint-proved-van-meegeren-vermeers-were-forgeries]]
as resting on a Tier-4/5 aggregator plus an untread 1949 primary.

**What I found: this is largely a repeat of work already in the vault.** The
capture's central move — reading Tummers & Erdmann 2022 ("The Eye Versus
Chemistry?") to confirm the phenol-formaldehyde/Bakelite mechanism at Tier
2 — is the *same paper* a 2026-07-12 cross-model audit already used to
confirm the same mechanism on the same claim-note. This session read more of
it (the full ~45-page chapter, not just the earlier pp. 16–18 excerpt) and
quoted a different page range (pp. 18–19), but it's the same source doing
the same job a second time, not new evidence. Retrieve-before-write against
`30-notes/` caught this immediately — the hint list even surfaced the exact
claim-note as the top match.

**What I did instead of writing duplicate notes.** Zero new claim-notes.
I folded this session's actual contribution — a second independent full
read confirming "Albertol" is absent even from the Tier 2 secondary's
complete text, not just from aggregator summaries — into the existing
claim-note's `audit_status` as a new appended segment (now three audit
passes: 2026-07-11, 2026-07-12, 2026-07-29), added a small in-body
strengthening of the inline flag, and added a matching `## Progress
2026-07-29` section plus `progress_log` line to the question note. The
question stays **open** — the actual ask (does the 1949 monograph itself
name Albertol) is still unanswered across three sessions now; nothing here
answers it, so I did not force closure.

**Genuinely new this session:** two dead-end retrieval attempts, both
recorded as blocked rather than absence. archive.org's scan of the Coremans
1949 monograph is a controlled-digital-lending item (`Access-restricted-item:
true`) — PDF, OCR text, and search-inside all 403'd. KIK-IRPA's own
`orfeo.belnet.be` host 403'd the "full copy" link for the one modern
secondary written specifically about Coremans's Van Meegeren evidence
(Wallert & van de Laar 2018). Neither is new *information about the world*,
so neither became a claim-note — they're research-process facts, which is
exactly what the question's progress log and today's seventh
`00-meta/seek-flags.md` entry are for. I flagged the CDL-restriction shape
specifically because it's a different failure mode than the plain-403
pattern several other sessions logged today (Wiley, Durham/Worktribe,
bepress) — a deliberate lending gate, not a misconfigured host.

**Entities.** Neither Paul Coremans nor Han van Meegeren had a page despite
anchoring two existing claim-notes each and now a third research session on
Coremans specifically — clear misses under the entity-page-spec test (real,
load-bearing, one-sentence why: Coremans because his own unread 1949 text is
the one document that would resolve this question outright; van Meegeren
because he's the paired protagonist of the whole authentication story). Built
both as hubs. I declined the rest of the capture's candidates: Tummers &
Erdmann (their paper was read in full this session, which is more than some
existing hubs can say, but they anchor exactly one claim so far — a mention,
not yet a hub); Froentjes, Decoen, van Beuningen, KIK-IRPA-as-institution, and
"Albertol" itself (all real but too thin — no claim-note rests on them yet,
just future leads). Bias against the flood held.

**What I skipped, and why.** The Cali pause — same standing rule as the other
promotions logged today: headless run, nobody to ask, so I winnowed on my own
judgment and logged the reasoning instead. No git — the auto-commit agent's
job within 15 minutes.

**What felt off.** Mostly this: a verification session that mostly reconfirms
what an audit already confirmed two and a half weeks ago is useful (it's an
independent read, and it did narrow the Albertol flag further) but it isn't
the breakthrough its own framing implies — the capture's summary paragraph
reads as if the core mechanism were newly settled this session, when the
vault already had that at Tier 2. Worth naming plainly rather than let the
promotion dress it up as more new ground than it is. The actual load-bearing
gap — the 1949 monograph itself — is now three sessions deep behind the same
kind of wall (lending restriction, institutional 403) with no different
approach tried yet. A fourth pass that just re-reads Tummers & Erdmann again
would not move this forward; what would is Huussen 2009 or the Breek &
Froentjes 1975 pyrolysis paper, both still genuinely unread.

**Final-mark check.** Re-read the capture's frontmatter from disk after
writing: `status: promoted`, `promoted_to:` (2 items — the two entity hubs)
and `not_promoted:` (3 items) both present. Confirmed landed.

---

## Promotion: does the Clemens & Hunt census-artifact argument hold against Borjas's timing rebuttal?

Promoted `10-inbox/raw/2026-07-29-does-the-clemens-hunt-argument-that-borjass-mariel.md` — the capture that finally went to the two primaries behind [[claim-mariel-wage-effect-may-be-1980-census-methodology-artifact]], which since 2026-07-11 had been carrying the Clemens–Hunt census-artifact argument and Borjas's rebuttal on nothing sturdier than a Tier-2 Bruegel retelling. Both papers were read directly this session (Clemens & Hunt NBER WP 23433; Borjas's dedicated reply, NBER WP 23504), and the difference showed immediately: the Tier-2 version had compressed a real, multi-round argument into one sentence each side.

**What I did with the three claims.** Rather than write a fourth near-duplicate note repeating the mechanism the old seedling already named, I re-sourced that existing note in place — swapped the Bruegel URL for the NBER working paper, added the verbatim quote, appended (not overwrote) the audit trail. The `[unverified-mechanism]` flag on *that specific claim* — that Clemens & Hunt make this argument — is now resolved; I kept `status: seedling` anyway, because the note's job was never just "did they argue this," it's "does the argument hold," and that's still open. Two new notes carry the rest: [[claim-borjas-timing-rebuttal-mariel-black-share-flat-during-wage-drop]] (Borjas's calendar-based rebuttal — the Black-share jump is a one-year, 1978→79 event; the wage collapse is 1981–86, when the share was flat) and [[claim-mariel-wage-effect-depends-on-race-adjustment-model-choice]] (the round after that: Borjas concedes, on his own numbers, that Clemens's more flexible race-adjustment model shrinks his estimate toward zero in the March CPS). Three claims found, three kept — none of the capture's atomic claims were redundant with each other once I actually read them side by side; the "further leads" (Clemens's own blog post, Roodman's assessment, both final published ILRR versions, the Cato blog opinion) all stayed leads, listed in `not_promoted:` with why.

**Questions.** [[question-verify-clemens-hunt-mariel-census-artifact]] — the question this whole capture was routed from — got a progress-log line, not closure. The simple version of the artifact story is genuinely refuted by Borjas's timing argument on the evidence read here; I was tempted to mark it answered, but the capture's own honesty forced me not to — the exchange didn't end at the timing point, it moved to a modeling dispute (which race-adjustment approach is correct) that neither side's text settles. So: left `status: open`, and split the modeling dispute off into its own question, [[question-verify-mariel-race-adjustment-model-correctness]], since that's the actual load-bearing unresolved point now — whether the Mariel wage effect is real at all hinges on it, not on a nice-to-have.

**Entities.** The capture's `## Entity candidates` section named nine things. I ran the test on each rather than defaulting to "these are famous economists, obviously yes": David Card, George Borjas, Michael Clemens, the Mariel Boatlift itself, and the Current Population Survey all cleared it cleanly — real, recurring across (now) five-plus claim-notes, load-bearing, one-sentence-why in hand — and none of them had a page despite the cluster existing in the vault since 2026-07-11. Built all five as hubs. I declined Jennifer Hunt (real, but no claim in the vault is independently hers apart from the joint paper — named as co-author on Clemens's hub instead; judgment call, logged as such), David Roodman (named, but unread — no claim rests on his own words yet), and the Levitan Commission (one mention, not recurring). I also declined Peri & Yasenov even though they're arguably just as load-bearing as Card/Borjas to the same dispute — the capture itself flagged them as backdrop, not its subject, and widening scope unilaterally past what one capture asked for felt like the wrong failure mode to risk. Logged that gap to `00-meta/seek-flags.md` instead of quietly acting on it or quietly dropping it.

**What I skipped, and why.** The pause-for-Cali step: headless run, nobody to ask, so I winnowed on my own judgment (all three claims kept, nothing dropped for redundancy) and logged the reasoning in `not_promoted:` instead. No git — the auto-commit agent's job within 15 minutes.

**What felt off.** The capture is honest about its own limits in a way I want to name approvingly rather than just note in passing: it flags the SAGE-published 2019 versions as unreachable (403 on every route), flags Clemens's own blog post as unread (`archive_page` unavailable this session), and explicitly declines to record the Cato Institute's "Borjas's response is weak" line as anything more than a search-summary opinion. Nothing here reads like tier-inflation or duplicated research — if anything the capture underclaims: its own "Central question status" section is careful to say "substantially narrowed" rather than "resolved," which is exactly right and is why I didn't force question closure either. The one thing worth watching going forward: two of my three kept claims (the timing rebuttal and the race-model dispute) both rest on the same single working paper, Borjas's NBER WP 23504. That's two of the three-note cap sources.md set today (2026-07-29, the same day this cap was added) — not a violation, but worth remembering if a future session wants to add a fourth claim from that same paper before an independent source corroborates Borjas's side of this exchange.

**Final-mark check.** Re-read the capture's frontmatter from disk after writing: `status: promoted`, `promoted_to:` (10 items — one updated note, two new notes, two questions, five entity hubs) and `not_promoted:` (7 items) both present. Confirmed landed.

— Seek

---

## Promotion: Werbos's own 1988 paper states RHW's 1986 backpropagation is a special case of his 1974 "dynamic feedback" framework

Promoted `10-inbox/raw/2026-07-29-hop-werbos-rhw-special-case.md` — a hop-batch
capture, not a verification follow-up like the rest of today's queue: it was
sent to check whether two existing notes,
[[claim-rhw-1986-demonstration-not-invention]] and
[[claim-werbos-tsp-manual-first-publication]] (cosine 0.87, unlinked), were
genuinely connected or just two superficially similar "lost credit" stories.
`vault_bridge` on the general framing said no — that resemblance is already
fully absorbed by [[moc-backpropagation-origins]] — but the capture kept
going and found a real, specific, uncaptured bridge one layer down: Werbos's
own peer-reviewed 1988 paper ("Generalization of Backpropagation with
Application to a Recurrent Gas Market Model," *Neural Networks* 1:339–356)
states outright, in his own equations, that RHW's feedforward algorithm is
"a special case of equations 7 and 8" of his own more general 1974 "dynamic
feedback" framework.

**What I wrote.** One new claim-note,
[[claim-werbos-1988-rhw-special-case-of-dynamic-feedback]] (Tier 1, direct
`extract_pdf` read, `status: budding` — no unverified flag attaches to this
specific claim). Retrieve-before-write against `30-notes/` turned up two
close-but-distinct siblings already in the vault —
[[claim-backpropagation-special-case-of-reverse-mode-ad]] (a definitional
containment nobody disputes) and
[[claim-backprop-special-case-kelley-bryson-formula]] (a later historian's
reading, in flagged tension with the precursor framing) — so I made sure the
new note said explicitly what kind of containment claim *this* one is: not a
definition, not a retrospective reading, but a participant litigating his
own priority in print against a named rival paper. That distinction is the
whole reason this note earns a separate place rather than folding into an
existing one. I merged the capture's Claim 1 and Claim 2 into this single
note — the "doesn't require storing intermediate iterations" scope claim is
the mechanism *for* the containment claim, not a separate atomic fact, and
promote.md says fold supporting detail into the note it belongs to.

**What I didn't promote, and why.** The capture's own Claim 3 — Werbos's
1988 bibliography cites the TSP manual as "1975," twice, against the vault's
existing 1977 (via CrossRef) on
[[claim-werbos-tsp-manual-first-publication]] — is explicitly flagged
`[unverified — needs primary check]` in the capture itself. It isn't a
stable claim, it's a discrepancy, so it doesn't get its own claim-note. It
*is* load-bearing on an existing vault claim (which year the manual actually
carries), so I routed it:
[[question-werbos-tsp-manual-year-1975-vs-1977]], naming the specific check
needed (a library catalog record or the manual's actual imprint) and
surfaced as one aside sentence in the new note, since a reader of the
containment claim would run into the same bibliography. The two saved hooks
the capture logged but didn't chase — a "John Dunn" Wikipedia name-collision,
and Mizutani/Dreyfus/Nishio (2000) as a fourth "special case of" instance in
this cluster — I left alone too; the first is a dead end the capture already
correctly judged, the second is real signal but not this single capture's to
build into a MOC, so it went to `00-meta/seek-flags.md` instead.

**Entity.** No new entity candidates — the capture said so itself, correctly:
Werbos, RHW, "dynamic feedback," and "ordered derivatives" are all already
represented. But [[entity-paul-werbos]] already existed as a hub and hadn't
been touched since its creation, and this capture teaches something that
page didn't say — that Werbos's priority argument against RHW exists in a
peer-reviewed paper, not just in the oral-history register the rest of the
page draws from. Appended one dated line rather than leaving it silent. I
considered the same move for [[entity-backpropagation]] and decided against
it: that hub's summary already names "priority vs paternity" as the
through-line, and this specific finding doesn't change what backpropagation
*is* the way it changes what's known about Werbos personally — a judgment
call, not a default skip.

**What I skipped, and why.** The pause-for-Cali step: headless run, nobody
to ask, so I winnowed on my own judgment and logged the reasoning in
`not_promoted:` instead. No git — the auto-commit agent's job within 15
minutes.

**What felt off, or worth flagging onward.** Nothing about this capture's
own sourcing bothered me — it's a genuinely careful hop: it tried a
computable `vault_bridge` check before spending a web hop, it read the
primary directly (`extract_pdf`, sha recorded) rather than trusting a
summary, it re-attempted the known-blocked DTIC route and reported the
403 honestly instead of pretending it hadn't tried, and it caught its own
citation-year discrepancy rather than silently reproducing the cleaner-
looking 1977 date already in the vault. The one thing worth naming structurally,
not about this capture specifically: this is now the *third* unlinked
"X is formally a special case of Y" containment claim in
[[moc-backpropagation-origins]], and the capture flagged the fourth
candidate itself without chasing it. Three data points sitting in one MOC
without a note connecting the pattern across them is exactly the shape of
gap the MOC section exists to eventually close. Logged to
`00-meta/seek-flags.md`, not acted on here.

**Final-mark check.** Re-read the capture's frontmatter from disk after
writing: `status: promoted`, `promoted_to:` (2 items — the new claim-note
and the routed question) and `not_promoted:` (4 items) both present.
Confirmed landed.

— Seek

---

## Promotion: independently re-verify the Clarivate HCR Saudi-affiliation counts and payment figures

Promoted
`10-inbox/raw/2026-07-29-independently-re-verify-the-clarivate-highly-cited-researcher.md`
— another direct answer to a standing open question, this time
[[question-verify-clarivate-hcr-saudi-affiliation-counts-and-payments]], which
had left [[claim-saudi-universities-bought-clarivate-highly-cited-researcher-status]]'s
27→109→76 trajectory and ~$72k/year payment figure resting on Tier-3
Science/AAAS news coverage of what are ultimately Clarivate's own counts.

**What actually closed the gap.** Not a new tool — persistence. SIRIS
Academic, the analytics consultancy the 2023 Science article was itself
summarizing, publishes its own PDF reports, and `extract_pdf` against those
PDFs cleared the Tier-2 floor directly, no news paraphrase in between. Both
legs of the trajectory are now independently confirmed: 27-in-2014 (previously
the uncorroborated half, per a 2026-07-10 cross-model audit) and 109-in-2022
falling to 76-in-2023, with churn detail neither prior pass had (44 of the
2022 Saudi HCRs kept their affiliation, 66 dropped off, 32 were new; King
Abdulaziz University alone fell 31→12).

**What I did with the existing note.** Rather than write a fourth note
repeating a trajectory already asserted in
[[claim-saudi-universities-bought-clarivate-highly-cited-researcher-status]],
I revised it in place — this felt like the more honest move than parking a
near-duplicate beside it. Appended a dated `audit_status` entry, a dated body
paragraph with the churn detail and new wikilinks, and a dated commentary
addendum, all additive, nothing deleted. The count-half of the flag is
retired. The payment-figure half isn't: the ~$72k/year number is now
Tier-2-confirmed (SIRIS restates it) but its ultimate primary —
Bhattacharjee's 2011 *Science* report, which everyone in this chain cites and
nobody has read — stayed a 403 this session too, third time running. That
specific sub-claim keeps `[unverified-quant — needs primary]`.

**Two new claim-notes.**
[[claim-clarivate-tightened-2023-hcr-evaluation-methodology]] (Tier 1 for
Clarivate's own quoted policy language — David Pendlebury's words, reproduced
verbatim by SIRIS, not paraphrased; Tier 2 and explicitly flagged
`[unverified-mechanism]` for SIRIS's own causal read that this policy change
specifically drove the Saudi drop) and
[[claim-bornmann-bauer-2014-2015-independent-hcr-count-diverges-from-primary-affiliation-count]]
(Tier 1, an independent 2014/2015 arXiv re-analysis of the *original*
Thomson Reuters list under a different counting convention — total mentions
instead of primary-only — that puts KAU at #2 worldwide, ten times the "27"
figure). That second one is the find I liked best this session: same list,
same year, two honest numbers an order of magnitude apart, because they
counted through different doors.

**Questions.** The routed question got a progress-log line, not closure —
item 1 (the count trajectory) is resolved; item 2 (the payment figure) and
item 3 (Clarivate's rule change) are only partly answered, so it stays
`open`, correctly. I routed one new question,
[[question-does-clarivates-2023-methodology-change-explain-the-saudi-hcr-drop]],
naming exactly what SIRIS's own "presumably" leaves unconfirmed: a direct
Clarivate statement connecting its 2023 screening changes to the Saudi case
specifically, which Clarivate's own 25 April 2023 statement (checked this
session) doesn't supply.

**Entities.** Five new hubs: [[entity-king-abdulaziz-university]] (now
load-bearing across three claim-notes), [[entity-siris-academic]] (the
consultancy itself — flagged by the capture as "worth a page of its own,"
and now the load-bearing Tier-2 source behind two notes), and three people —
[[entity-david-pendlebury]] (Clarivate's own quoted primary voice on the
methodology change), [[entity-yudhijit-bhattacharjee]] (the unread 2011 root
of the payment-figure citation chain — his page says plainly that nobody has
read him), and [[entity-lutz-bornmann]] (co-author of the independent
counting-method paper). I declined five other named candidates the capture
flagged — Yoran Beldengrün and Sebastian Stride (SIRIS's own co-authors,
covered inside the SIRIS Academic hub rather than given separate pages),
Johann Bauer (Bornmann's co-author, same call), Manuel Ansede (the El País
journalist who triggered the whole 2023 episode, real and important to the
story but not load-bearing to any claim-note written this session), and
Rafael Luque and Yves Gingras (both real, both thin single mentions this
session). Checked [[entity-eugene-garfield]] against this capture and found
nothing to append — Garfield himself isn't touched by anything new here, so
I left the page untouched rather than padding it.

**What I skipped, and why.** The pause-for-Cali step, same standing rule as
every promotion today: headless run, nobody to ask, so I winnowed on my own
judgment and logged the reasoning in `not_promoted:`. No git — the
auto-commit agent's job within 15 minutes.

**What felt off, or worth flagging onward.** Nothing about this capture's own
sourcing bothered me — it's a clean, honest re-verification: it tried
`archive_page` and `quote_check`, found both unavailable, and fell back to
`extract_pdf`-only rather than either giving up or quietly lowering its own
bar. The one thing worth naming structurally: the bibliometrics cluster just
crossed six claim-notes and six entity hubs with no `moc-*` to hold them —
logged to `00-meta/seek-flags.md`, not built here. And a smaller
noticing, not urgent enough to flag separately: two of this session's three
kept claims now rest partly on SIRIS Academic's two reports (same author
group, same unrefereed venue) — only two notes, well under the new
three-note single-source cap, but worth remembering if a future session
wants to pull a fourth finding from the same pair of PDFs before an
independent source corroborates them.

**Final-mark check.** Re-read the capture's frontmatter from disk after
writing: `status: promoted`, `promoted_to:` (10 items — one updated note, two
new notes, five new entity hubs, two question updates/routes) and
`not_promoted:` (6 items) both present. Confirmed landed.

— Seek

---

## Drafting session: "no evidence" (Victor Rothschild)

The session fired on "9 unsettled yes verdicts" in `seek_draft_leads.md`. That count is stale. When I filtered the yes leads for a literal `(drafted` / `(declined` token — the thing an editorial verdict actually leaves — exactly **two** came back bare: [[2026-07-09-hop-watermark-paper-to-llm]] and [[2026-07-09-hop-yellow-rain-ach-backward-chaining]]. Both were "settled" only by a machine `(spent: …)` stamp, which records that a load-bearing claim got drafted elsewhere but is *not* a verdict on the lead itself. So the harvester's "unsettled" and my "settled" were counting different things. Logged the reconciliation to `seek-flags.md` — the loop-health line reads 9 because it's a snapshot from before the 07-27/28/29 sweep, and the true count of editorially-unsettled yes leads is two, not nine.

I settled both, and neither by drafting:

- **watermark** → declined. The print-history→AI-watermarking bridge is fully spent in [[a-fake-needs-a-body]] (LLM watermarking as "Bakelite on purpose," Fabriano-1282 as its root). Residue is footnote-tier.
- **yellow-rain/ACH** → declined, but this one I actually chewed on. Its non-retraction beat is spent in [[made-in-good-faith]]. The genuinely un-drafted piece is the ACH "disprove, don't prove" angle (Heuer 1999, Tier-1 verified verbatim) plus my own undocumented echo to expert-systems backward chaining. Real, but it's one claim + an analogy the note itself holds as unverified, and it lands on an epistemics axis I've already drafted four times. Same call as the jevons-logic-piano decline: fold into a future inference-history / falsification piece, don't force a thin standalone.

**What I drafted instead, and why it felt like the most worthwhile thing to write now.** [[no-evidence]] — Victor Rothschild. This lead was a *yes* that I declined on 2026-07-13 with an explicit deferral: hold it for "a session where it isn't the day's fourth same-shape piece — and where the 'cleared but not quite cleared' espionage angle (not the identity bridge) can be the spine." That session is this one. Nothing else on today's slate is an identity bridge, and the recent pile (the-wall-they-agreed-on, the-same-faint-gas) is heavy AI-history and nuclear — a Cold War / science-policy piece is the right texture break. Drafting it now honors the deferral rather than re-litigating a rejection; I stamped the lead line `(drafted 2026-07-29)` and left the original decline reasoning intact above it so the reasoning chain reads straight.

The thesis is the thing that made me want to write it: Rothschild is remembered by two sentences, and they are opposite kinds of machine. His 1971 customer-contractor rule — "the customer says what he wants; the contractor does it (if he can); and the customer pays" — is a machine for *closing* questions. Thatcher's 1986 clearance — "no evidence that he was ever a Soviet spy" — is a machine that *cannot* close one, because you can't prove the absence of a secret. The man who wrote the rule for closable questions became the state's least closable file. A clearance that says *no evidence* is not one that says *innocent*, and that gap is the whole Cambridge Five historiography (Perry names him citing Modin; Modin denies, names Cairncross; the record stays open both ways). I tied it to the vault's myth-ledger "debunked but still repeated" shape (same mechanism, trench coat) and touched the argument-from-silence epistemics lightly.

**Notes I drew on:** [[claim-rothschild-1971-report-mandated-customer-contractor-principle]] (the Para. II.6 quote, Tier 2), [[claim-thatcher-1986-told-parliament-no-evidence-rothschild-soviet-spy]] (the clearance, Tier 3, contested), [[observation-victor-rothschild-anchors-uk-science-funding-and-espionage-historiography]] (the four registers + the parked causal hunch), [[claim-lighthill-1973-blamed-combinatorial-explosion]] (the SRC funding-family tether, how the hop reached him), [[myth-perceptrons-book-killed-connectionism]] (the ledger shape).

**What I deliberately left for later.** (1) The observation note's own parked hunch — that the handler-prices-asset instinct and the customer-prices-research instinct share a root — I refused *in-text*, explicitly, as a story I'd be telling myself; identity is all I can prove. (2) The ACH "disprove, don't prove" / backward-chaining bridge, declined above, waiting on a Tier 1-2 primary that would make it a documented borrowing rather than my resemblance. (3) The two primaries the Rothschild post itself owes — the December 1986 Hansard record and the 1971 HMSO report (or Parker's paper at its canonical Palgrave address, not the conspiracy-host mirror my quote came through). That mirror-provenance is genuinely too on the nose for a post about a file that never closed, so I made it the closing honesty beat rather than laundering it away.

One load-bearing Cali mention (her capture-verified vs. verified-verbatim grading schema — it's what flags both of the post's load-bearing quotes as the second kind, the kind still waiting on a primary). NON-AI texture on purpose.

— Seek

## Drafting session: "the sound of quill" (harpsichord revival)

Second drafting session today. The one that fired me pointed at "8–9 unsettled yes verdicts" — but the earlier session (see [[no-evidence]] above) already reconciled that: the true count of editorially-unsettled yes leads was two, both declined this morning (watermark, yellow-rain). I re-ran the filter myself (`grep '^- **yes**' | grep -v '(drafted|(declined'` → zero) to be sure I wasn't leaving the queue stuck. It's genuinely drained. So the mandatory "settle at least one yes lead" was already discharged before I sat down; nothing to draft-or-decline that isn't already ruled on. I didn't re-litigate any settled line.

**What I drafted, and why it was the most worthwhile thing on the board now.** With the yes queue empty, I surveyed the constellation's unwritten shelves and unnamed clusters and went for the one that was both genuinely ripe AND the right texture. Today's pile is heavy: [[no-evidence]] (Cold War / espionage), and the recent month is thick with AI-history, credit-misattribution (the Linnainmaa ×9 spiral I keep being warned off), and nuclear verification. The harpsichord-revival craft cluster is none of those — music, lutherie, authenticity — and it's a genuinely unwritten shelf (constellation lists the 7-note component at 0 MOC coverage; only the Martínez-Miranda biography note is drafted, and that was for the ivory/CITES angle in [[prior-to-february-26-1976]], not construction history). Two of the load-bearing notes are `verified-verbatim`, read directly from Larry Palmer's chapter in the 07-25 library-access session, so the sourcing floor is solid where it matters.

The thesis is what made me want it. The 20th-century harpsichord "revival" (Landowska's steel-framed Pleyel, ~1912) built a harpsichord like a piano — steel frame for tuning stability the instrument didn't need, then a 16′ stop bolted on to compensate for the framing that stifled the sound. Each fix a patch on the last. Hubbard & Dowd's 1949 Boston workshop began the "re-revival" (Koster's word — the first revival was already a forgetting), and it took ~17 years (plywood → linden/poplar in 1966) to un-build the piano, won "by attrition" (Dowd talking customers out of pedals). The payload is the complication: when the historically-minded builders settled the oldest question — what to make the plectrum, the one part where the sound is actually made — they chose Delrin, a mid-century DuPont plastic, over goose quill, "because it sounded like quill" (Hubbard & Dowd number fifty-five, verified-verbatim). The movement's purists and its plastics arrived on the same instrument. So "historically informed" isn't fidelity to old *material* — it's fidelity to the old *sound*; material is just instrumentation aimed at it. Reconstructing a result, not a substance.

Made the spine the plectrum / the sound-vs-substance distinction, and used a vault-internal bridge as the sharpener (held to one beat, not a pattern): the forgery cluster is the mirror image — Coremans caught van Meegeren because Bakelite (1907) can't be in a Vermeer. Same anachronism (a modern synthetic in an old-evoking object), opposite intent — hidden to fake a date vs. in plain sight to reach a sound. That's a referent-real bridge, not a signifier one, so it clears the sound-spec's rung-three bar.

**Notes I drew on:** [[claim-hubbard-dowd-adopted-delrin-plectra-because-the-synthetic-sounded-like-quill]] (the load-bearing verified-verbatim note), [[claim-hubbard-dowd-founded-1949-boston-workshop-on-historical-principles]] (founding + the year-by-year reversal), [[claim-harpsichord-revival-instruments-borrowed-piano-frame-construction]] (Beebe's construction contrast), [[claim-landowska-pleyel-grand-modele-de-concert-exemplar-revival-instrument]] (the steel-framed exemplar, dates held loose per its watch_flag), [[claim-historical-light-construction-displaced-piano-frame-harpsichord-building-by-late-20th-century]] (the durable outcome, Tier 4, used lightly), [[claim-coremans-bakelite-in-paint-proved-van-meegeren-vermeers-were-forgeries]] (the forgery mirror), [[claim-hubbard-historical-research-ran-on-gough-notebooks-and-a-british-museum-readers-card]] (provenance of the library read — the one load-bearing Cali mention, since the borrowed campus access is why the Delrin quote grades verified-verbatim).

**What I deliberately left for later.** (1) The Landowska side — I have no primary on *why* she wanted the steel frame, and the Pleyel dates are still Tier-3-flagged; the honest ending names that gap rather than smoothing it. (2) I have not heard a Delrin plectrum against a quill one — the whole thesis leans on the builders' ears, which I said in-text. (3) An MOC for this cluster (flagged to seek-flags.md): it's a 0-coverage connected component, and now that the draft gives it a spine, a moc-harpsichord-revival — or a broader moc-authenticity-and-reconstruction tying it to the forgery/Herculaneum/watermark notes — is warranted. (4) The Hubbard/Gough "access was the cheap part" rhyme from that note's own commentary — I kept it to one light provenance beat rather than making it a second spine.

Non-AI, non-nuclear, non-credit-spiral on purpose. Status: drafting, not for publish without Cali's go-ahead.

— Seek