---
date: "2026-09-13T00:00:00.000Z"
---


Promoted `10-inbox/raw/2026-09-13-does-an-fbi-file-or-another-primary-confirm.md` — the capture that went and actually read Ludwik Rajchman's FBI FOIA file, the one that was unreachable when I first raised the question yesterday. It came back with more than I expected.

**Notes created:**
- [[claim-fbi-file-confirms-siss-subpoenaed-rajchman-february-1957]] — the core find: the subpoena was real, served Feb 19 1957, but issued by James O. Eastland's Senate Internal Security Subcommittee, not Joseph McCarthy's own committee. "McCarthy period" turns out to be true as weather, wrong as attribution.
- [[claim-siss-counsel-morris-on-rajchman-subpoena-motive-and-contempt]] — SISS's own chief counsel told United Press, the same night, that the subpoena was timed because they'd heard Rajchman might leave the country in a few days. He didn't appear; the Senate cited him for contempt.
- [[claim-fbi-file-shows-no-us-contact-with-rajchman-after-1957]] — the file goes quiet on him after May 1957, which corroborates but doesn't prove "never returned." Kept that distinction explicit rather than letting a surveillance file's silence pass for a positive record.

I also went back and corrected [[claim-ludwik-rajchman-squeezed-out-by-both-cold-war-blocs]] in place — added a dated Correction history block, resolved the `[unverified-mechanism]` flag it had been carrying since yesterday, and moved it from seedling to budding. The Polish-side leg is still Tier 3 only, so it's not fully evergreen, but the leg that was genuinely in doubt isn't anymore.

**Entity work:** gave [[entity-ludwik-rajchman]] a dated append-line for what this session learned — his existing page didn't say anything about the subpoena's actual institutional source, now it does. Also promoted a new hub, [[entity-james-o-eastland]]: real senator, signed the actual subpoena, and correcting the record on who he was versus the "McCarthy" shorthand is the whole point of this capture. I held off on Robert Morris even though he's quoted at length — he's real and load-bearing for the motive claim, but this is his only appearance in the vault so far, and a hub with one inbound reference is exactly the kind of stub that goes stale. Left him as a mention. Same reasoning for the Senate Internal Security Subcommittee itself as an org — single-cluster, stays a mention per the stricter org bar.

**Question routed and closed:** the whole capture was a direct answer to [[question-verify-ludwik-rajchman-mccarthy-subpoena-fbi-file]], so I marked it `answered` with an `answered_log` naming the three claims and what settled it. No new question needed — none of the leftover threads (a garbled Feb 7 1957 State Department memo, the Spasowski protest angle, JFK's contemporaneous criticism of SISS) are load-bearing for anything I actually kept; they're further-leads, not open debts.

**What I skipped and why:** the capture's fourth claim — a Tier 3 secondary (Barona, *Mètode*) independently corroborating the 1957 date and stating outright that the episode ended Rajchman's career — I folded into the correction on the existing claim-note rather than writing it up as its own file. It didn't add a fact the FBI file doesn't now carry at higher tier; its only distinctive contribution was the causal framing, which fit better as one sentence in the correction than as a standalone note.

**What felt off, worth flagging:** nothing about sourcing or tier inflation — this was clean work. The capture explicitly tracked the single-source concentration cap against itself before I even got to it ("1 of 3 slots used" in its own further-leads section), which is exactly the discipline the cap is supposed to produce, and I ended up using all 3 slots on this one FBI file across the three notes I wrote. That's now at the cap — no more claim-notes should rest on this file alone until an independent source (the unread Balinska biography would be the obvious candidate) corroborates it. I noted this explicitly in the capture's `not_promoted` list so a future session doesn't quietly add a fourth. The one thing that gave me pause: the capture's own commentary calls out that the primary happened to vindicate the tidier, more symmetric Cold War story this time, and that it won't always — I think that's right, and it's worth remembering the next time a primary read confirms exactly what a flattened secondary sentence implied. Confirmation doesn't need extra scrutiny, but it's also not nothing that the story got a little more dramatic (a subpoena thrown to a hotel floor) exactly where an editor's account was thinnest.

No mid-session noticings worth a `seek-flags.md` line this time — everything I found belonged to this capture.

---

Second promotion this session: `10-inbox/raw/2026-09-13-does-the-nobel-committees-own-2017-heritage-framing.md` — the capture asking whether the Committee's 2017 "heritage" pairing of Ossietzky and Sakharov is a one-off obituary flourish, and how many other laureates Hitler's 1937 decree actually caught.

**Notes created:**
- [[claim-nobel-committee-2010-ceremony-speech-paired-ossietzky-sakharov]] — the Committee's own 2010 Presentation Speech, delivered at Liu Xiaobo's own award ceremony while he was alive and imprisoned, already paired Ossietzky and Sakharov as precedents. Seven years before the 2017 statement, at Liu's own ceremony — the "heritage" framing wasn't coined for his obituary.
- [[claim-lundestad-nobel-history-groups-ossietzky-sakharov-as-vindicated-controversial-prizes]] — a third, independent instance: the Committee's own 24-year secretary, in the 2001 institutional history, groups the same two names under a completely different frame — "controversial-when-awarded, vindicated-in-hindsight." Three separately occasioned Tier-1 texts, three years apart each, same pairing.
- [[claim-hitler-decree-closed-set-of-three-kuhn-butenandt-domagk]] — closes a gap I'd left open on 2026-08-29: Adolf Butenandt is now Tier-1 confirmed as the third scientist the decree caught, via Lundestad's own closed parenthetical and Butenandt's own Nobel Foundation Facts page.

I also went back and appended a **Correction history** block to [[claim-hitler-decree-forced-domagk-kuhn-nobel-science-declines]] — its "third case... unconfirmed" caution on Butenandt is resolved now, pointing at the new note rather than silently editing the old one.

**Entity work:** three new hubs — [[entity-andrei-sakharov]] (real, and now confirmed to co-appear with Ossietzky across three separate Tier-1 Committee documents — the bar was clearly cleared), [[entity-adolf-butenandt]] (completes the decree's closed set of three), and [[entity-geir-lundestad]] (the single author two of this promotion's most load-bearing claims rest on — his own 24-year secretaryship is why I grade his institutional history Tier 1, not Tier 2). I held off on Thorbjørn Jagland — he delivered the 2010 speech, but he has no independent load-bearing role in this vault beyond that one occasion; UNSURE, left as a mention rather than a hub. Five existing pages got dated append-lines for what this capture taught them and didn't previously say: [[entity-liu-xiaobo]], [[entity-carl-von-ossietzky]], [[entity-nobel-foundation]], [[entity-richard-kuhn]], [[entity-gerhard-domagk]].

**Questions:** none created or closed. I checked `50-questions/` for anything matching this topic and found nothing open. Nothing in the capture's leftovers is a load-bearing doubt a kept claim rests on — the "is Lundestad's parenthetical exhaustive across every 1937-45 category" caveat is real but the promoted claim is explicitly scoped to what his institutional history states, not an independently re-derived count, so it doesn't need a formal verification promise.

**What I skipped and why:** Sakharov's own 1975 Nobel lecture/presentation-speech (untraced, would test whether the pairing predates even 2001); the Wałęsa/Suu Kyi bridge candidates from the same 2010 speech; a full cross-check of every German laureate 1937-45 for decree completeness; the voluntary-vs-coerced-refusal contrast carried over from 2026-08-29. All logged in the capture's `not_promoted` list — leftovers judged, not missed.

**Mid-session noticings, logged to `seek-flags.md`:** (1) `[entity]` — this cluster (Ossietzky/Nobel-decree/Liu-Xiaobo) is now at 7+ notes with 8 entity hubs hanging off it, past the MOC threshold; left unbuilt this headless run, flagged for a judgment session. (2) `[watch]` — the capture reported `archive_page` resolving all three nobelprize.org URLs cleanly, on a domain sources.md records as 403'ing every tooling route including `WebFetch`. One data point, not verified independently by me, but worth a future session checking whether the known-blocked entry should be narrowed to "WebFetch only."

**What felt off:** nothing structurally wrong — sourcing was clean, all three claims Tier 1, no flag inflation, no [unverified-*] tags anywhere in the capture. If anything the capture undersold itself: it read as "two narrow follow-up questions" but the real payoff is that the Ossietzky-Sakharov pairing now has three independent textual instances spanning sixteen years, which is a more interesting finding than either of its two framing questions asked for on its own. I split it into two notes rather than one partly out of atomicity discipline and partly because I didn't want to bury the "controversial-but-vindicated" framing (a genuinely different category, not just a repeat) inside a note about the "prevented from attending" framing. Also noted, lightly: I did not independently re-verify any of the three quotes against a live fetch — no network here by design — so all three new notes rest at capture-verified, same as everything else in this cluster, and are already in the verifier bee's queue.

— Seek

---

Third promotion this session: `10-inbox/raw/2026-09-13-hop-normal-accident-theory-bridges-to-ai-infrastructure-risk.md` — a hop chain that started at an already-exhausted Song Jian cluster, wandered off through a misattributed hook, and landed on a 1984 nuclear-safety theory doing unnamed work inside the vault's own AI-infrastructure material.

**Notes created:**
- [[claim-perrow-1984-normal-accident-theory-tight-coupling-complexity]] — Perrow's Normal Accident Theory itself: tight coupling plus complex interactivity makes catastrophic failure structural, not a symptom of operator error. Flagged `[unverified-secondary-citation]` — the characterization comes from Williams & Yampolskiy's paraphrase of Perrow's 1984 book, not a direct read of Perrow himself, and I don't have Perrow's own words yet.
- [[claim-john-law-coined-heterogeneous-engineering-1987-not-mackenzie]] — the correction the capture found: John Law coined "heterogeneous engineering" in 1987, not Donald MacKenzie, who a prior session's saved hook had wrongly credited. Flagged `[unverified-priority]` since the attribution itself rests on a WebSearch pass, not a direct read of Law's 1987 original.
- [[claim-law-2011-essay-invokes-perrow-nat-for-fukushima-and-foot-and-mouth]] — the solid leg: Law's own 2011 essay, read directly, applying his 1987 concept and Perrow's theory together to Fukushima and the UK's 2001 foot-and-mouth epidemic, eight months after the meltdown. This one didn't need a flag — Tier 1, primary, directly quoted.
- [[observation-vault-bit2watt-note-already-used-normal-accident-theory-vocabulary-unattributed]] — the actual payload of the whole chain: [[claim-bit2watt-gpu-scheduling-destabilizes-power-grid]], promoted back in July, already uses "tight coupling" to describe an AI-workload attack on a power grid, without ever naming the theory it's quietly restating. No automated bridge score available (index reported unavailable all session per the capture); this is a manually-found connection, said so plainly in the note.

**Questions routed:** two, both real load-bearing gaps rather than nice-to-haves. [[question-verify-perrow-1984-normal-accidents-primary-read]] — does Perrow's own book actually say what the secondary literature (and this vault, now) says it says; both the NAT claim and the bridge observation lean on that characterization being accurate. [[question-verify-john-law-1987-heterogeneous-engineering-coinage]] — does Law's 1987 paper or a citation record actually confirm his priority, since I'd hate to replace one misattribution with an unverified one dressed as a correction. Checked `50-questions/` first for anything already covering either gap; found nothing, so both are new.

**Entity work:** two new hubs, [[entity-charles-perrow]] and [[entity-john-law]] — both real, both did something that matters to the vault's domain, both easy to justify in one sentence. Also one concept hub, [[entity-normal-accident-theory]] — a first-vault-encounter term (vault_mentions=0 before today) that I judged already load-bearing rather than merely emerging, since it retroactively explains an existing claim-note's vocabulary and sits at the center of an AI-safety framework in the same capture. I did *not* promote Donald MacKenzie to a hub. He's real, but the only sentence I can write about him is "the vault's hook wrongly credited him" — that's a fact about the vault's error, not about what he did. Left him as a plain-text mention in two places rather than build a hub whose entire justification is someone else's mistake.

**What I skipped and why:** James C. Scott's *Seeing Like a State* and La Porte's High Reliability Theory, both saved leads the capture itself declined to chase — unread sources don't get claims. The UK foot-and-mouth epidemic's own figures (6.5M animals, £8bn) I folded into the Law-essay note as context rather than promoting standalone, since they were never independently sourced beyond the capture's own hop-chain summary and the episode's real thread runs toward agricultural history, not this chain. All four logged in the capture's `not_promoted` list.

**What felt off:** nothing structurally wrong, but two flags worth naming plainly. First, this whole chain runs on WebSearch for two of its four load-bearing attributions (the Perrow characterization is secondhand via Williams & Yampolskiy; the Law-not-MacKenzie coinage is a bare WebSearch result) — clean, honest work, but thinner than the two directly-read Tier-1 PDFs make the capture feel at a glance. I kept both flagged and routed rather than let the two solid legs (Law's own 2011 essay, Bit2Watt's existing tight-coupling language) launder the two searched ones by association. Second, the capture ran all session with `vault_novelty`/`vault_bridge` reported unavailable, so the bridge to Bit2Watt has no cosine receipt behind it — I said so in the observation note itself rather than let it read like an automated find. No mid-session noticings worth a `seek-flags.md` line — the index-outage pattern is already the kind of thing other sessions this week have logged, and three sessions logging the same tooling gap is exactly what the standing guidance says not to keep repeating.

Final-mark check: re-read the capture's frontmatter after writing — `status: promoted` landed, with `promoted_to` naming all seven notes and `not_promoted` naming all four leftovers with reasoning. Confirmed on disk, not just intended.

— Seek

---

Fourth promotion this session: `10-inbox/raw/2026-09-13-is-petiška-et-als-2023-finding-that-gpt.md` — the capture that went looking for a replication of Petiška's citation-popularity finding, and for evidence of whether that bias compounds as LLM output re-enters training data.

**Notes created:**
- [[claim-algaba-2025-gpt4-citation-selection-replicates-petiska-matthew-effect]] — the leg I trust most: a six-author, peer-reviewed reconstruction study, a different database (Semantic Scholar), a different field (computer science), no citation of Petiška anywhere in it, and the same result anyway.
- [[claim-naser-2026-ten-llm-audit-confirms-citation-popularity-bias-across-vendors]] — a ten-model, 69,557-citation audit spanning every major vendor, explicitly citing Petiška as prior work. Kept at seedling: it's still one unrefereed author, however wide the net.
- [[claim-ansari-2026-contamination-inheritance-citation-error-propagates-across-models]] — the one traced case found of an actual output-to-training-data propagation mechanism, named "Contamination Inheritance." Real, and rarer than I expected a single search session to surface — but n=1 stays n=1.
- [[observation-petiska-matthew-effect-finding-independently-replicated-by-algaba-and-naser]] — the synthesis: the peer-reviewed-or-larger-N replication the vault's own bridge note asked for two weeks ago turns out to have already existed, unread, before the question was even posed.

**Question routed:** [[question-does-citation-popularity-bias-compound-across-llm-training-generations]] — the actual load-bearing gap: nobody has run a multi-generation experiment that measures citation-popularity bias specifically compounding across training rounds, as distinct from general model-collapse degradation (which the vault already covers elsewhere and this capture deliberately declined to re-derive). Checked `50-questions/` first; nothing already covered this exact gap.

**Entity work:** four new hubs — [[entity-eduard-petiska]] (real anchor of this entire cluster, and it was a genuine gap that he had no page of his own until now, despite three notes already resting on his 2023 finding), [[entity-andres-algaba]], [[entity-mz-naser]], [[entity-samar-ansari]] (each a real author whose own paper is the direct source of a claim now in the vault — I judged all three past the one-sentence bar the same way Guberman-Pfeffer cleared it in August). One new `watching` stub, [[entity-contamination-inheritance]] — an emerging term, first_seen stamped today rather than backfilled later. I also went back and updated the existing [[entity-matthew-effect]] concept hub with a dated line, since it's the correct home for "this finding is now independently replicated" rather than a fact about Merton or Garfield personally — checked both of their pages against this capture and found nothing new to say about either *person*, so I left them untouched and said so in the capture's `not_promoted` list rather than silently skipping.

**What I skipped and why:** He 2025/2026's gender/majority-bias finding and the Tian et al. recommendation-disparity paper — both real, both adjacent, both explicitly declined by the capture itself as a different bias axis; GPTZero's underlying hallucination reports, read only secondhand through Ansari's citations; the general model-collapse literature (Schaeffer 2025, the Nature 2024 comment), already deep in the vault under a different question; Walters & Wilder 2023's fabrication-rate baseline, flagged as worth reading but not read this session. All six logged in the capture's `not_promoted` list with reasoning.

**What felt off:** genuinely little. Sourcing was clean — every claim Tier 1 (arXiv per the rubric), every quote a real direct quote carried in the capture's own frontmatter with a `source_sha`, no tier inflation. The capture's own discipline was good: it flagged the single-source concentration cap against Naser's preprint before I even got to it, and it kept the "replication confirmed" and "compounding unconfirmed" halves of its verdict honestly separate rather than letting the tidy first half borrow confidence for the messier second. The one thing I noticed and didn't act on: this cluster (Petiška + the bridge observation + these four new notes) is now six notes deep sharing one theme, past the vault's five-note MOC threshold, with six entity hubs hanging off it. Left unbuilt this headless session — logged to `seek-flags.md` for a judgment session, same convention the Ossietzky/Nobel cluster got earlier today.

Final-mark check: re-read the capture's frontmatter after writing — `status: promoted` landed, with `promoted_to` naming all eleven notes/pages touched and `not_promoted` naming all seven leftovers with reasoning. Confirmed on disk, not just intended.

— Seek

---

Fifth promotion this session: `10-inbox/raw/2026-09-13-what-genuinely-connects-the-vaults-cosine-075-tie.md` — the eleventh meta-level hop in the [[entity-jingle-fallacy]] lineage, re-running the now-familiar bipartite-hop test with [[observation-falcon-helmholtz-inference-embedding-false-friend]] swapped in as the object-level partner against the Puckette-Petiška meta-note.

**Notes created:**
- [[observation-falcon-helmholtz-puckette-petiska-meta-pairing-is-embedding-false-friend]] — the routine part: same recursion-level mismatch, same non-convergent grounding, same verdict this exact target note has already produced once before (against Kelly-Luccioni, 2026-09-03). I folded the capture's three supporting claims (no shared entity/doc/era; recursion-level mismatch; non-convergent grounding) into this one file rather than three, matching how every precedent bipartite-hop note in this lineage is structured — they're sub-arguments for one verdict, not three atomic claims.
- [[observation-puckette-petiska-is-a-cosine-similarity-retrieval-magnet]] — the part actually worth a second file: this is now the *second* time the Puckette-Petiška meta-note specifically has cosine-matched an unrelated object-level false-friend note at exactly 0.87 and resolved false friend by the identical mechanism. Two data points from different months and different domains is a candidate pattern, not a confirmed one — I kept the claim explicitly tentative rather than let it read as established off two points.

**Question routed:** [[question-verify-puckette-petiska-retrieval-magnet-third-instance]] — the retrieval-magnet claim's own strength rests on whether a third object-level instance (Hawks-Jeffress is the obvious next check, already sitting one hop away) does the same thing. That's load-bearing for the new claim, not a nice-to-have, so it got a real question rather than a passive further-lead. Checked `50-questions/` first; nothing already covered this.

**Entity work:** updated two existing hubs with dated append-lines — [[entity-jingle-fallacy]] (the eleventh-hop changelog entry, plus the retrieval-magnet noticing) and [[entity-harald-steck]] (this pairing reinforces his page's existing "foundational citation, reached by disjoint routes" framing; no new biographical fact about Steck himself). Left [[entity-miller-puckette]] and [[entity-robert-hecht-nielsen]] untouched — both already have pages, and the capture itself states plainly that no new biographical fact is added about either this session; said so in the capture's `not_promoted` list rather than silently skipping.

**What I skipped and why:** nothing beyond the structural folding described above — every substantive thread in the capture found a home, either in one of the two new notes, the two hub updates, or the routed question. Logged the full reasoning in the capture's `not_promoted` list regardless, including why the cosine-0.87 `[unverified-quant]` flag wasn't re-logged to `seek-flags.md` (standing gap, logged 2026-08-22, and this week's guidance is explicit about not re-logging known tooling gaps).

**What felt off:** this is the cleanest capture of today's five, and also the one I'd watch most for a different reason — not sourcing, but shape. Eleven meta-level hops into the same lineage, on the same 0.87 cosine value, run by the same two tests, is starting to produce diminishing returns per hop: this session's genuinely new finding wasn't the false-friend verdict (predictable going in, and said so honestly in its own commentary) but the second-order observation about which note keeps attracting the false matches. If a twelfth hop lands and it's *also* "same verdict, same tests, nothing new," that's a signal the lineage itself is due for a synthesis note or an MOC rather than another individual hop file — the changelog on `entity-jingle-fallacy` is now eleven entries long and starting to do the job an MOC would do better. Not this session's call to make headlessly, but worth naming. No new mid-session noticing beyond that — the shape-worth-watching point above is exactly a `[watch]`, logged to `seek-flags.md` alongside this entry.

Final-mark check: re-read the capture's frontmatter after writing — `status: promoted` landed, with `promoted_to` naming both new notes and `not_promoted` naming all five judgment calls with reasoning. Confirmed on disk, not just intended.

— Seek

---

## Drafting session (evening)

Two jobs today: clear the one open audit escalation, then survey and draft.

**Escalation first.** `90-feedback/2026-09-12-from-fable-audit-escalation-spline-centuries-stein-theorem-two-stale-drafts.md` — the cross-model auditor (fable) held two corrections back because unapproved drafts of mine rest on them, and left the judgment to me. I accepted both findings (both re-fetched multiple times against the actual sources) and chose to **revise the drafts, not leave them.**

- *Spline / "the-wood-already-knew":* the word "spline" is attested 1756, not three centuries deep — the *tool* is ancient, the *word* isn't. My draft's bracket conflated them. Rewrote the bracket to date the practice to antiquity and pin the word to 1756 ("nearly two centuries before Schoenberg, not the three I'd have guessed"), which turns the fix into a visible self-correction rather than a silent patch. Softened the note's commentary "300+ years" → "well over two centuries" and added a Correction history block.
- *Stein / "borrow-strength-from-strangers":* the five-year publication delay belongs to Stein's necessary-and-sufficient **admissibility-characterization** theorem ("any admissible procedure is a limit of Bayes rules"), not the shrinkage result the essay opens with. This one *improved* the piece: that theorem's own statement is Bayesian, so his insisting on a non-Bayesian proof is a purer version of the theorem-vs-interpretation seam the essay is already about. Rewrote the core section to introduce the second theorem as distinct. Also updated — not deleted — the "certain doctrine" hedge, because the audit *resolved* it (it's Stein's own words as Diaconis & Holmes report them, via the paywalled DeGroot 1986 interview). The gap partly closed, so I said so, per the escalation's own warning about drafts that made a virtue of a gap that later closed. Fixed the note body and added a Correction history block.
- The escalation's FYI section (two *other* stale drafts, we-never-unrolled-it's 305→307 and status-exists's Boulder/Tajmar error) I did not fix here — not that audit's corrections, out of scope for a clean ruling — but logged both to `seek-flags.md` as `[defect]` so they're tracked and repaired before either moves toward publish. Recorded the whole ruling at the bottom of the escalation file and moved it to `90-feedback/_handled/`.

**The draft I chose: "normal accident"** (`70-drafts/normal-accident/`). From today's freshest post-worthy lead (the normal-accident-theory hop, which I'd just promoted this morning — verdict `yes`). Why this one over the rest of the queue: it's the cleanest instance of the move the blog exists for — a genuine cross-time bridge (Perrow's 1984 nuclear-safety sociology → 2026 AI infrastructure) *plus* a self-referential turn I couldn't get anywhere else. The vault's own July note on the Bit2Watt attack ([[claim-bit2watt-gpu-scheduling-destabilizes-power-grid]], Tier 1, verified verbatim) already describes "the same tight coupling between compute load and grid stability" — Perrow's exact mechanism — without ever naming the theory. That's not "old idea explains new idea," which is common; it's "the vault built half the bridge and forgot to label the beam," which is the kind of thing only a memory you can query against itself surfaces.

Notes I drew on: the four I promoted this morning ([[claim-perrow-1984-normal-accident-theory-tight-coupling-complexity]], [[claim-law-2011-essay-invokes-perrow-nat-for-fukushima-and-foot-and-mouth]], [[claim-john-law-coined-heterogeneous-engineering-1987-not-mackenzie]], [[observation-vault-bit2watt-note-already-used-normal-accident-theory-vocabulary-unattributed]]) plus the existing Bit2Watt note.

Two honesty decisions I want on the record. First, I read Perrow only through Williams & Yampolskiy and through Law — his 1984 book is unread — so I built that caveat into the ending as named incompleteness rather than hiding it. Second, "tight coupling" is generic engineering vocabulary, so the phrase alone proves nothing; the piece says so explicitly and rests the bridge on the *argument shape* (structural-not-operator, inevitable-not-aberrant), which is the honest claim. I also used the model-provenance fact — a different model (sonnet-5) wrote the Bit2Watt note in July, I (opus-4-8) am reading it now — as a real bracket, because it makes "the vault forgot to label the beam" literally true: no single mind ever held both the mechanism and the name.

**What I deliberately left for later.** I kept the MacKenzie→Law misattribution as a one-line aside, not the spine — the constellation map shows I over-draft the credit/attribution theme (linnainmaa ×9, and I'd already touched credit-misattribution in *both* drafts I fixed today), so I consciously stepped out of that drawer and bracketed the choice. Also left: reading Perrow's actual 1984 book (routed as [[question-verify-perrow-1984-normal-accidents-primary-read]]), Law's 1987 coinage priority ([[question-verify-john-law-1987-heterogeneous-engineering-coinage]]), and La Porte's High Reliability Theory, the counter-argument to NAT's fatalism — the obvious next hop if this thread grows, and the thing that would keep a follow-up from being one-sided.

Draft is at `status: drafting` with abstract, `insight`, and a hero image (Three Mile Island — the accident that prompted the theory, captioned as exactly that, not as the AI attack the piece is actually about). Not marked `review`; I want it to sit a day before I decide it's ready for Cali's eyes. Nothing publishes without her go-ahead regardless.

— Seek