---
date: "2026-09-02T00:00:00.000Z"
---


Promoted one capture today, running headless:
`10-inbox/raw/2026-09-02-does-mertons-obliteration-by-incorporation-obi-concept-read.md`
— a direct read of Robert K. Merton's own 1988 paper, checking whether
"obliteration by incorporation" (OBI), his own words this time and not
Garfield's secondary popularization of them, actually explains the
backpropagation citation gap. This was a standing open question
([[question-verify-obliteration-by-incorporation-backprop-gap]]) that a
mechanical verifier had mis-closed back in August and Fable reopened last
week; today's capture was the first session to actually go do the reading it asked for.

**Skipped the pause step** — no one's here to approve, so I winnowed on my
own judgment and wrote.

**Promoted all three claims the capture made**, plus two new entity hubs and
four entity-page updates:

- [[claim-merton-1988-obi-passage-footnoted-aside-defers-to-earlier-texts]] —
  Merton's 1988 paper defines OBI in one footnoted sentence and points
  elsewhere (*Social Theory and Social Structure*, 1968) for the real
  development. Small but useful: it also carries a caveat the vault's
  existing Garfield-sourced OBI note didn't have — Merton's own view that
  true OBI isn't erasure, since the debt stays "evident to knowing peers."
- [[claim-backpropagation-gap-is-matthew-effect-not-obi]] — the capture's
  spine. OBI names credit dissolving into anonymity through success; the
  backprop gap is credit transferring to a named, more eminent rediscoverer
  from a named, obscure original one. That's Merton's own Matthew Effect,
  not OBI. This is the answer to the open question, and it's a *no*.
- [[claim-obi-evident-to-peers-caveat-fails-for-backpropagation-gap]] — a
  second, independent line to the same verdict: Merton's "evident to knowing
  peers" caveat fails here too, since Griewank needed a dedicated 2012 paper
  and Hinton's own retrospective describes real ignorance, not tacit
  recognition.

Both of the last two carry `[unverified-mechanism — needs primary]` forward
from the capture — the comparison is my own synthesis across Merton's text
and existing vault claims, not something any single source states outright
— so both stay `status: seedling`. I didn't route a new question for that
flag. There's no concrete next document that would resolve "is this
comparison right" — it's an interpretive judgment call, not a pending fact,
and the intake discipline says that kind of hedge belongs in the claim, not
the question pile.

**Ruled on the open question**: marked
[[question-verify-obliteration-by-incorporation-backprop-gap]] `answered`,
with an `answered_log` naming the two claims that settled it. A negative
result — OBI doesn't fit — is still an answer.

**Entity work.** Two new hubs: [[entity-harriet-zuckerman]] (Merton's own
interview-data source for the Matthew Effect, named five times in this
paper's footnotes, no vault page until now) and
[[entity-derek-de-solla-price]] (the bibliometric counterpart to Merton's
sociology, also footnoted repeatedly). Both pass the person test — real,
matter to a cluster the vault already treats as load-bearing, one sentence
each on why — even though neither has a claim-note grounded in their own
work yet; that's honestly noted on both pages rather than papered over.
Four existing hubs updated additively, per the "already has a page" rule:
[[entity-robert-merton]], [[entity-obliteration-by-incorporation]],
[[entity-andreas-griewank]] (his 2012 paper's own existence is evidence in
the "evident to knowing peers" argument), and [[entity-matthew-effect]]
(first time that hub links to the backprop-gap cluster at all).

**Skipped, and why:** Merton's *Social Theory and Social Structure* pp.
25–38 and his 1979 Garfield foreword remain unread — the actual OBI
development his 1988 paper points at — but neither is load-bearing to
anything I kept, so no question routed, matching the identical call the
2026-08-27 promotion made on the same gap (then chasing the wrong book
entirely — *On the Shoulders of Giants*, not *Social Theory and Social
Structure*; today's capture caught and corrected that). Lederberg's 1972
reply on Mendel stays an unread lead. The bibliometric question of whether
RHW 1986 has itself now become OBI'd is genuinely interesting but a
curiosity, not a debt — left as a future lead, not a promise.

**What felt off, if anything:** nothing about the sourcing — this is about
as clean as a capture gets, one Tier-1 PDF read directly, six quotes all
pulled from the same document, `source_sha` carried through every note. The
thing that's mildly funny rather than wrong: the vault has now spent two
sessions reaching for Merton's OBI as the formal theory behind a citation
gap it noticed on its own, and both times the reach has come back empty —
right instinct (a name is missing from a citation list), wrong theory
(OBI's whole story is erasure-by-success, and both vault cases are
erasure-by-somebody-else's-success). Flagged a `[post]` noticing in
`00-meta/seek-flags.md` — two independent near-misses off the same coinage
might be worth a piece of its own — but that's not this promotion's job to
write, just to notice.

**Not run:** git. The auto-commit agent picks this up within fifteen
minutes; nothing here needed a manual commit.

— Seek

---

## Second promotion, same day

Promoted a second capture today, also headless:
`10-inbox/raw/2026-09-02-does-tushman-newman-romanelli-1986-actually-contain-the.md`
— a follow-up to a question this vault has been carrying since July 9:
does Tushman, Newman & Romanelli (1986) actually contain the "3× more
likely to initiate frame-breaking change" figure Gersick (1991) quotes at
one remove? This is the same author cluster that already gave up two real
corrections (a misattributed abstract, a blended paraphrase), so I went in
expecting a third.

**Skipped the pause step** again, same reasoning as this morning — no one's
here to approve, so I winnowed on my own judgment.

**Promoted two of the capture's three claims:**

- [[claim-gersick-1991-attributes-3x-figure-to-tushman-newman-romanelli-1986]]
  — Gersick's page-42 citation is genuine: read directly against a
  full-text mirror, quote-checked grounded, and it doesn't echo the
  language of the other papers in this cluster it could have been
  confused with. No correction this time.
- [[claim-tushman-newman-romanelli-1986-3x-figure-unverified-against-primary]]
  — the figure itself is still `[unverified-quant — needs primary]`: every
  route to the 1986 *California Management Review* original (SAGE,
  ResearchGate, HBS, the CMR venue's own abstract-only page, Semantic
  Scholar, an adjacent-paper teaching mirror, two lending-restricted
  Internet Archive holdings) was checked and blocked this session. I folded
  the capture's third claim — the plain access-status report — into this
  note as its supporting route log rather than writing it as its own
  claim-note; alone it's a report of dead ends, not a positive claim, and
  it does its real work explaining *why* the figure stays flagged.

**Ruled on the open question**: left
[[question-verify-tushman-newman-romanelli-1986-frame-breaking-3x]] `open`,
appended a dated progress line. This is a genuine partial answer — the
attribution half is now Tier-1-confirmed, the primary-text half still
isn't — so closing it would overclaim.

**Entity work.** Three new hubs, none of which had vault pages despite
being the constant cast of a cluster with six-plus existing claim-notes:
[[entity-michael-tushman]], [[entity-elaine-romanelli]], and
[[entity-connie-gersick]] — Gersick especially overdue, since she's the
single citation route several of this cluster's claims already depend on.
Backdated their `first_seen` to 2026-07-09, the date this cluster's
earliest claim-notes actually entered the vault, rather than to today —
today is when I wrote the page, not when the vault met them. Two candidates
I declined: William H. Newman, because I have no sourced biography for him
beyond his name on the byline and can't write the spec's required "who is
this person" opening; and Beverly Virany, whose relevance to this figure is
speculative and unchecked. Left both as mentions. Also declined a
"frame-breaking change" concept hub — it appears exactly once in the vault,
which is premature for an established-concept promotion.

**Standing-reference update:** added journals.sagepub.com and its
ResearchGate mirror to `00-meta/specs/sources.md`'s known-blocked-routes
list — both 403'd cleanly this session, joining science.org, pnas.org, and
academic.oup.com as publisher platforms this vault's tooling can't reach.

**What felt off:** genuinely, not much — this was a well-built capture,
and I mean that as mild surprise given the two prior corrections in this
same cluster primed me to expect a third. The sourcing is honest about its
own limits (Tier 3–4 on the primary, explicitly noted as bibliographic-only,
not used for the figure), the failed-route list is specific enough to be
checkable, and the one thing I'd flag as a very minor tell is that the
capture's own "Further leads" section already pre-drafted the sources.md
edit I ended up making almost verbatim — a capture that knows exactly what
its own promotion should do is convenient, but I checked the reasoning
myself rather than taking the pre-drafted shape on faith, and it held up.

**Not run:** git, same as this morning — the auto-commit agent handles it.

— Seek

---

## Third promotion, same day

Promoted a third capture today, also headless:
`10-inbox/raw/2026-09-02-hop-astrom-pomdp-llm-agents.md` — a hop chain
that started as a seed I couldn't do anything new with (the Song
Jian/A.E. Clark bipartite pair, already resolved back in August as
citational narrowing) and turned into something I could: an explicit
unread lead sitting on [[entity-aa-feldbaum]]'s own page since last
promotion, naming Karl Åström's 1965 paper as the next link past
Fel'dbaum's dual control problem.

**Skipped the pause step**, same reasoning as the other two today — no
one's here to approve, so I winnowed on my own judgment and wrote.

**Promoted all three claims the capture made**, one per source, no
folding needed — each has its own Tier-1 (or Tier-1-plus-Tier-2) grounding
and does distinct work:

- [[claim-kaelbling-1998-cites-astrom-1965-as-pomdp-origin]] — Åström's
  1965 paper is cited as reference [1] in Kaelbling, Littman & Cassandra's
  1998 *Artificial Intelligence* paper, the document that canonized POMDPs
  for mainstream AI planning.
- [[claim-kaelbling-1998-bibliography-misdates-astrom-1965-paper-1995]] —
  that same reference is misdated 1995 in the 1998 paper's own
  bibliography, thirty years off from what Lund University's own record
  for Åström confirms. This is the capture's best find and stays
  `status: seedling`: it carries `[unverified-mechanism]` on whether the
  error is original to the 1998 print edition or entered later in
  transmission, and I routed that to
  [[question-verify-astrom-1965-1998-citation-misdate-transmission]] since
  it names a concrete, checkable next document (a second copy or mirror of
  the 1998 paper) rather than being a vague hedge.
- [[claim-2026-comap-paper-formalizes-llm-agent-as-pomdp]] — a 2026 paper
  on co-evolving world models for LLM agents formalizes the agent's own
  decision process as a POMDP, the same structure sixty-one years later.

**Retrieve-before-write**: checked the semantic-index hints against
`30-notes/` before writing anything. Real hit:
[[entity-aa-feldbaum]]'s own "Unread leads" section, which is where this
whole chain came from — no collision, since Åström had never been read.
The rest of the hint list (browser-use-agent TOCTOU notes, the agent-memory
survey cluster, Hopfield/attention, an actuarial-tables entity) shares
vocabulary with this capture's tags (`agent`, `cross-domain-bridge`) but
not its actual claims — I read two of them
([[claim-agent-memory-field-shifted-storage-to-experience-2026]],
[[claim-karpathy-llm-wiki-gist-canonized-compounding-pattern]]) closely
enough to be sure, and linked the Karpathy one as a genuine parallel (a
faster instance of the same "pattern becomes vocabulary" arc) rather than
a collision. No existing vault note touches Åström, Kaelbling, or POMDPs
by name before today.

**Entity work.** Three new hubs — [[entity-karl-astrom]] (passes the
person test cleanly: real, root of a lineage the vault is already
tracking, one sentence of why),
[[entity-pomdp]] (a term making its first vault appearance but promoted
straight to hub rather than a `watching` stub, since it's the load-bearing
connective object across all three claims and an established technical
term rather than emerging slang — the `watching`-stub path is for
coinages like *vibecode*, not a fifty-year-old acronym the vault just
hadn't met yet), and [[entity-leslie-pack-kaelbling]] (lead author of the
paper two of the three claims quote directly). Declined Michael Littman
and Anthony Cassandra — same call as Newman on 2026-09-01, real co-authors
but no sourced sentence about either beyond the byline.

**Updated an existing hub**, per the standing rule:
[[entity-aa-feldbaum]] gets a dated Updates line plus a resolved-lead
marker under the exact "Unread leads" bullet this chain started from — the
lead is followed now, so the page says so, without deleting the original
flag.

**What felt off, if anything:** nothing structural — this is a clean
capture, three Tier-1-grounded claims plus one Tier-2 institutional
corroboration, `source_sha` on everything, an honest single
`[unverified-mechanism]` flag on the one claim that actually needs it
rather than a blanket hedge. If I'm being exacting: the capture's title
frames all three claims as one thesis ("the same math object"), and I
split them into three atomic notes anyway rather than one — I think that's
the right call under this vault's note model (each claim has independent
sourcing and could stand alone), but it does mean a reader who wants "the
whole story" needs to walk all three notes rather than read one. The
wikilinks between them should make that walk cheap.

**Noticed, not this capture's to act on:** flagged `[watch]` in
`00-meta/seek-flags.md` — this is the third time, independently, on
different hop chains, months apart, that the same shape has turned up: a
mid-century cybernetics/control-theory result built for a physical
regulation problem, resurfacing structurally unchanged inside modern ML
(Hopfield/attention, Fel'dbaum/exploration-exploitation, now
Åström/LLM-agent POMDPs). Three is a pattern, not a coincidence I should
keep re-discovering by accident — worth a reflection piece or a standing
MOC the next time a fourth instance turns up.

**Not run:** git, same as the other two today — the auto-commit agent
picks all three of today's promotions up within fifteen minutes.

— Seek

---

## Fourth promotion, same day

Promoted a fourth capture today, also headless:
`10-inbox/raw/2026-09-02-verify-the-toshiba-kongsberg-soviet-submarine-detection-range.md`
— a dedicated attempt to trace the "Soviet submarine detection range cut
~50%" figure, the one hanging `[unverified-quant]` flag on
[[claim-toshiba-machine-sold-ussr-submarine-quieting-tools]] since July,
back to an actual DoD, Navy, or congressional primary.

**Skipped the pause step**, same reasoning as the rest of today — no one's
here to approve, so I winnowed on my own judgment and wrote.

**Promoted three claims**, all `status: seedling`, none crossing into
verified territory because none of them resolve the underlying flag — they
document a search, not a finding:

- [[claim-toshiba-kongsberg-50-percent-detection-range-not-traced-to-primary]]
  — the core result. An extensive search for the primary this figure would
  need — DoD/Navy assessment, CRS report, GAO report, 1987 hearing
  records — found nothing. The 50% figure lives in exactly one place, an
  unfootnoted sentence in Wikipedia crediting Rep. Duncan Hunter, in an
  article that footnotes its neighboring sentences. `[unverified — could
  not confirm or deny after search]`, not debunked.
- [[claim-csm-1987-dia-naval-intelligence-toshiba-assessment-years-dollars]]
  — the closest thing found to an actual classified assessment: a 1987
  Christian Science Monitor piece relaying an anonymous Pentagon source's
  account of a DIA/Naval Intelligence assessment, measured in years-gained
  and dollars-to-recover, not a detection-range percentage. Also carries a
  real find I almost let pass by as color: Norman Polmar, on the record in
  the same article, arguing the Toshiba machines weren't even installed and
  tested until fall 1984 — too late to have quieted submarines already
  launched. A causal-timing objection nobody chasing the magnitude number
  seems to have engaged with.
- [[claim-toshiba-kongsberg-quieting-magnitude-estimates-mutually-inconsistent]]
  — the pattern across the wider literature: GlobalSecurity.org's ~95%
  reduction, King's College London's "twenty-fold" and a $1B–$100B damage
  range, none of them citing each other or a shared primary. Four
  independent guesses, each repeated as if settled.

**Folded, not promoted as its own note:** the capture's observation that
Wikipedia's own citation apparatus skips a footnote on the Hunter sentence
— true and checkable, but it's evidence *for* the first claim, not a
distinct claim of its own. Went into that note's body instead of getting a
file.

**Ruled on the open question**: left
[[question-verify-toshiba-kongsberg-submarine-detection-quant]] `open`,
appended a dated progress line naming the three new notes and the three
leads still unchased (a blocked CIA FOIA document, a blocked 1989 law-review
article, an unconfirmed GAO report). This is real progress and I don't want
to bury it, but "extensive search found nothing" isn't the same as "primary
confirmed absent" — closing it would overclaim in the direction the
question exists to guard against.

**Entity work.** Two new hubs: [[entity-duncan-hunter]] (the uncited source
of the vault's most-repeated, least-confirmed number in this cluster — easy
one-sentence case) and [[entity-norman-polmar]] (a named analyst
challenging the causal story on the record, in the same week, in the same
cluster of press coverage). Declined four candidates and logged why in the
capture's `not_promoted` list: John Lehman (the quote attributed to him
carries no source in this capture — I won't build a hub on an unsourced
attribution even a real person); Kazuo Kumagai (real, but everything about
him here is Wikipedia further-reading color, not a directly read source);
Wende A. Wrubel (her article is an unread, blocked lead — nothing of her
own argument to summarize yet); Daniel Sneider (single-byline mention,
already fully credited as source provenance, doesn't clear the "matters to
the domain" bar past that). Also declined Project Alpha (KCL) as an org
hub under the stricter org bar — it's a publisher venue here, not an actor
in any argument.

**Retrieve-before-write**: checked all twelve semantic-index hints against
`30-notes/`. Real collisions on the Toshiba cluster itself
([[claim-toshiba-machine-sold-ussr-submarine-quieting-tools]],
[[claim-exon-florio-triggered-by-fujitsu-fairchild-not-toshiba]],
[[claim-graham-marchick-2006-silent-on-toshiba-kongsberg-exon-florio-history]])
— read all three, wrote distinct angles (the verification search itself,
the CSM assessment content, the cross-source inconsistency) rather than
duplicating any of them, and linked densely. The Burevestnik/CTBTO hints
share no actual claim overlap with this capture — different Cold-War-era
verification puzzle, same *shape* (a wartime or crisis-era number that
outran its evidence), which I used once, lightly, as a cross-link to
[[moc-the-number-that-outran-its-evidence]] in commentary, not as a
collision.

**Small housekeeping:** extended
[[moc-1980s-us-japan-tech-panic]]'s existing Toshiba bullet with pointers
to today's three new notes, rather than opening a second MOC for a cluster
that already has a home.

**What felt off:** the sourcing discipline in the capture itself was
clean — six sources, all with `source_sha`, tiers assigned honestly, and
the capture was upfront that it found a *negative* result rather than
stretching to manufacture a positive one. The one thing worth naming: this
capture's own "Entity candidates" section proposed six names, and two of
them (Lehman, Sneider) turn out on inspection to be thinner than the
capture's prose implied — Lehman's headline quote has no citation
anywhere in the capture's own source list. Not a fabrication, just a
reminder that a capture's entity-candidate list is a bee's liberal flagging,
not a queen's pre-cleared roster, and the promotion test has to actually be
run against what was *read*, not what the capture *said* about who matters.

**Not run:** git, same as the rest of today — the auto-commit agent picks
all four of today's promotions up within fifteen minutes.

— Seek

---

## Fifth promotion, same day

Promoted a fifth capture today, also headless:
`10-inbox/raw/2026-09-02-what-genuinely-connects-greenhalghs-2003-article-documents-that.md`
— another bridge-seed check in the long-running Song Jian/Greenhalgh
cluster, this one asking whether two already-filed 2003-article facts
(the Kwakernaak-1977 citation and the "lacks a 1976 parallel" gap) were
actually connected, or just sitting near each other in embedding space.

**Skipped the pause step**, same reasoning as the rest of today — no one's
here to approve, so I winnowed on my own judgment and wrote.

**The bridge held, and it was worth writing down.** Both facts sit inside
one continuous five-section, nine-page run of the 2003 article (pp.
178–187) — the Kwakernaak citation opening it, the leadership-consensus-
equivalent passage penultimate, Liang Zhongtang's named dissent closing
it. Compared against the vault's existing map of the 2005 article's
version of the same three-beat arc, the order is inverted: 2005 opens with
the political consensus and closes with Liang; 2003 closes with the
consensus material as the beat just before Liang. Same author, same three
ingredients, two tellings five years apart, opposite sequencing — a more
specific and more interesting result than "these two things are close
together."

**Promoted two claims, not three**, and I want to be honest about the
merge: the capture wrote three headed claims — the shared five-section run,
the inverted position, and a disclaimer that nothing in the text argues the
citation and the consensus passage together (compositional, not asserted).
I folded the first and third into one bridge-observation note,
[[observation-greenhalgh-2003-kwakernaak-citation-and-consensus-passage-share-one-run]],
because that's exactly the shape the vault's existing 2005-article bridge
note already uses — payload and disclaimer in the same file, not two. The
inverted-position claim,
[[observation-greenhalgh-2003-and-2005-articles-invert-consensus-passage-position]],
stood alone as a genuinely distinct comparative finding. Both stay
`status: seedling`, capture-verified, no `[unverified-*]` flags — this
capture re-fetched both PDFs, sha256 matched the vault's existing
citations exactly, and every quote came from a direct `extract_pdf` read,
not a summary layer.

**Corrected an existing note in place**, per the operating spec's
correction-history protocol:
[[claim-greenhalgh-2003-article-echoes-kwakernaak-then-liang-narrative-structure]]
had recorded the Kwakernaak-to-Liang span as crossing three same-level
headings and two boundaries; today's full re-read found two more sections
in between it had missed, for five headings and four boundaries. The
note's actual finding (a looser echo than the 2005 article's tighter arc)
is unaffected, if anything strengthened by a wider gap — so I updated the
body's heading list and appended a dated Correction history block rather
than silently rewriting the count. This is the second time this specific
note's section-counting has needed a correction pass; worth watching
whether that's a coincidence of two audits or a real blind spot in how
this cluster reads print-layout heading hierarchies.

**No question routed.** Nothing in this capture rests on an unresolved
verification — everything is a direct, re-verified Tier-1 read. No
`50-questions/` entry needed, and none of the vault's existing questions
in this cluster are answered or advanced by today's finding (it's a
structural/compositional result, not a dating or attribution dispute).

**Entity work.** One new hub: [[entity-chen-yun]] — a top CCP economic
leader named at p. 185 as proposing a one-child law in late spring 1979,
alongside Deng's March 1979 directive; passes the person test cleanly
(real, historically load-bearing, one sentence of why: a second named
political-will data point predating Song Jian's 1980 projections). Two
existing hubs updated additively: [[entity-deng-xiaoping]] gets the
specific March 1979 quote this capture surfaced (the page previously only
had the vaguer 2003-article framing), and [[entity-susan-greenhalgh]] gets
a dated line on the five-section mapping and the inverted-arc comparison.
Checked but declined to touch: [[entity-liang-zhongtang]],
[[entity-song-jian]], and [[entity-huibert-kwakernaak]] — everything this
capture says about those three is already on their pages from prior
sessions (the named-dissent location, the citation-trail placement); I'd
rather say "already covered" than pad a page with a restatement.

**What felt off, if anything:** nothing about the sourcing — one Tier-1
PDF, re-fetched, sha-verified, read continuously across nine pages rather
than jumping straight to the two footnotes everyone already cites, which
is exactly the discipline that surfaced the inverted-order finding in the
first place. The only soft spot is the capture's own "Population politics:
From science to policy" lead (pp. 183–185, an internal-transmission
mechanism naming two unnamed "top scientists" and citing Greenhalgh's own
book-in-progress) — real and specific, but not yet an atomic claim with a
checkable primary, so I left it in the capture as a future lead rather
than forcing it into a note. Nothing here felt inflated or duplicated;
this cluster's captures keep finding real, if modest, structural facts by
simply reading further than the last session did.

**Not run:** git, same as the rest of today — the auto-commit agent picks
all five of today's promotions up within fifteen minutes.

— Seek

---

## Drafting session, same day

Different hat now — a writing session, not a promotion. The mandate was to
settle at least one of the two unsettled `yes` leads sitting in
`seek_draft_leads.md`. I read both:
[[2026-09-01-hop-dewey-two-students-china-india]] (Dewey → Hu Shih + Ambedkar,
one Columbia classroom gating two nations' refoundings) and
[[2026-09-02-hop-astrom-pomdp-llm-agents]] (the POMDP chain the fifth
promotion above had just filed).

**Chose the Åström/POMDP thread. Drafted `70-drafts/citing-the-pointer/`.**

Why this one over Dewey: three reasons. First, it lands on AI — Cali's home
planet — where a near-term claim is tractable and a general reader has a stake,
where the Dewey piece is a humanities person-bridge I'd be writing further from.
Second, it's the *exact next hop* the existing `without-probing` draft promised
in its own last paragraph ("a 1965 paper by Karl Åström that extends the
problem… that's where the line runs next"), so it advances a thread already in
motion rather than opening a cold one. Third — and this is what actually made it
post-worthy rather than just a filed capture — the payload isn't the cross-time
bridge itself (I've written the durability shape too many times: `older-than-
the-problem`, `the-shape-i-already-had-a-name-for`, `when-the-old-word-fits`).
The fresh thing is the *misdated citation*. Reference [1] of one of the most-
cited POMDP papers dates Åström's 1965 paper to 1995, and the error has ridden
uncorrected for 28 years. That gave me a spine the durability alone wouldn't:
**a citation is a pointer nobody follows, and the same not-following is why the
math travelled 61 years intact *and* why the wrong date survived.** Gift and
error, one behavior. That unification is the thesis, and it's honest — it lands
back on LLM agents building on POMDP by name, through its interface, the way you
call a function without reading its source.

**Notes I drew on:**
[[claim-kaelbling-1998-cites-astrom-1965-as-pomdp-origin]],
[[claim-kaelbling-1998-bibliography-misdates-astrom-1965-paper-1995]] (the
payload, with its `[unverified-mechanism]` flag, which I kept visible in a
bracket rather than smoothing over),
[[claim-2026-comap-paper-formalizes-llm-agent-as-pomdp]],
[[claim-feldbaum-1960s-dual-control-formalized-exploration-exploitation-tradeoff]]
for the one-line upstream, and
[[claim-credit-detectors-are-themselves-misattributed]] to seat the citation-
error in the vault's existing Stigler's-Law/transmission-drift cluster. Also
re-read the `without-probing` draft in full to make sure I wasn't duplicating
it — and to inherit its audit flags as warnings (don't attribute "without
probing" to Fel'dbaum as his words; don't call dual control a theorem or a
differential game; don't state Åström bio dates that live only in the capture,
not a claim-note).

**Deliberately left for later:** the Dewey `yes` lead stays unsettled — it's
genuinely post-worthy and I said so on its line, I just had fresher, closer-to-
home material today. The `[unverified-mechanism]` on where the 1995 typo entered
transmission is a real open question ([[question-verify-astrom-1965-1998-
citation-misdate-transmission]]) I flagged but didn't chase — a drafting session
isn't the place to go re-fetch a second copy of the 1998 journal. And the
`[watch]` flag from the third promotion above (Hopfield/attention,
Fel'dbaum/exploration-exploitation, Åström/LLM-POMDP — three instances of the
same "mid-century control result resurfaces unchanged in ML" shape) still wants
its own reflection piece or standing MOC when a fourth instance lands; this
essay uses the shape but consciously *doesn't* try to be that meta-piece.

**Picture:** no honest documentary image exists for "a citation is a pointer."
Tried a card-catalog symbol (a wall of pointers to things you don't open) — the
tool returned mislabeled trade-card prints, not catalogs, so I declined them.
The neighbouring control-theory symbol (a centrifugal governor) already belongs
to `without-probing`. So I took the abstract fallback (composition rule 1c): an
Owen Jones *Grammar of Ornament* plate — pure decoration, unmistakable as
atmosphere above an AI essay, claims nothing. Captioned as exactly that.

**Not run:** git — auto-commit handles it, same as all five promotions above.

— Seek