AA b16 — Finalizing the RiskyMAD paper (list of record, lives beside the drafts it plans)#

Created: 2026m08d04_13h48
Scope: everything that must happen before b16 (RiskyMAD) is published. Paper-scoped only.
Status: OPEN — this is the permanent finalizing list for b16, not a session snapshot.
Placement rule (LLoL, 2026m08d04): an AA file belongs next to the thing it plans.
This one sits at hell/mm/b/16/ — one level above the version folders, so it survives
oov3 -> oov4 — because what it plans is the b16 paper, not one draft of it.
Only larger-than-this-paper tasks go to the central registry at /_post/aa/.

Note

Why an AA file may live in HELL. The append-only rule in CLAUDE.md is scoped to LLogs, under source/matheology/hell/ll/. It has never applied to HELL as a whole — hell/mm/ holds actively rewritten drafts (this paper is being rewritten right now), and hell/aa/ has existed since 2026m06d03. So no rule needed relaxing for this file to exist here; what needed correcting was the claim that “HELL is append-only”, which over-generalised a rule that names hell/ll/ specifically.

And AA files do carry audit value. A task written down and then not done is evidence — of what was attempted, what was foreseen, and what was forgotten. That is often more revealing than a task never recorded. (Recorded because the opposite was claimed in session 2026m08d04 and was wrong.)

Where b16 lives — so nobody has to search HELL#

What	Where
Current draft (being edited)	`hell/mm/b/16/oov3/b16-riskymad-waiting-time_oov3_2026m07d18.rst`
Superseded drafts	`hell/mm/b/16/` — `mmv1/`, `mmv2/`, `mmv3/`, `oov2/`
Published floor page	`source/study/matheo/b16/` (`index.rst` + `b16-form-riskymad-mmv5.rst`)
Changelogs	`study/matheo/b16/b16-changelog-mmv5-to-oov2.rst`, `...-oov2-to-oov3.rst`
Session llogs (the reasoning)	`hell/ll/study/b/16/` — 12 files; most recent: `b16-oov3-execution-verification-llog_2026m07d25_11h39.rst`, `b16-oov3-ring-restructure-plan-llog_2026m07d18_11h00.rst`, `b16-oov1-master-plan_2026m07d15_19h30.rst`
Paper-local bibliography	`source/_bib/b16-nuclear-risk.bib` (24 entries); cross-cutting refs in `source/_bib/references.bib`
Figures	`source/_file/pdf/hell/mm/b/16/fig/`
Related POST AA entries	POST AA b17 (the three non-paper leftovers of the b16 revision session), POST AA b18 (build/link guards)
Cross-cutting citation work	POST AA b19 — the b16 BibTeX row below depends on its decisions 1 and 3

Task list#

k	s	Task	Notes
k6	s9	§1’s Arkhipov account is DOUBLED and both halves are live.	Opened 2026m09d04, from reading the built PDF. Section 1 opens with LLoL’s own draft paragraph (“During the Cuban Missile Crisis on Oct 27 …”, committed `ba83501`), then a transition rule, then the older P1 paragraph telling the same story again (“Pinned in the Sargasso Sea …”). Both render. A referee therefore meets the same account twice on page one, in two different voices, before anything else. This is a work-in-progress state that is invisible in the source (the two are separated by what looks like an editing marker) and unmissable in the PDF. The single largest thing standing between the current draft and a referee copy. LLoL’s to close — the three costed alternatives (P1 / P2 / P3) and the two epieikeia options are still in comment blocks immediately around the live text.
k5	s9	Typos and unfinished phrasing in the live §1 draft paragraph.	Opened 2026m09d04. In LLoL’s own draft text, which prints: “US Navy destoyers” → destroyers; and “had not gentle kind reasonably spoken up”, which reads as unfinished rather than as a phrasing choice. Not corrected — it is LLoL’s own uncommitted prose in a section he is actively rewriting, and the trifecta usage may be deliberate. Listed so it is not lost if the paragraph survives the rewrite.
k4	s8	Re-home the epieikeia argument, or decide to let it go.	Opened 2026m08d07 by the Section-1 compression. When the B-59 account moved to SIA-X7, the paragraph “No rule could have supplied that” went with it. That paragraph is LLoL’s own conceptual contribution of 2026m08d06, and it is the only place in the ring that argues judgement-when-rules-run-out is a function that can be staffed, trained and paid for — which is precisely what Sections 6 and 11 then ask the reader to fund. As it stands the ask has lost its stated warrant in the ring. Recommended: one sentence added where the ask is made (Section 6 or 11), not a restoration in Section 1 — which would undo the compression. Full paragraph preserved in SIA-X7. LLoL’s call on wording and placement.
k5	s9	[DONE 2026m08d04] Fix the stale section cross-references in the restored Declarations.	The Competing interests paragraph was ported verbatim from OOv2 on 2026m08d04 and still points at OOv2 numbering that OOv3 does not have: “Section 2.5” (simulations), “Section 4” (ResearchCity), “Section 2.6” (parameters pushed against the conclusion), “Section 4.0a” (reverse exposure), “Section 2.0a” (headline figure). OOv3 has flat sections 1–11 with no such subsections. Ported deliberately unchanged — re-deriving numbers during a port is how content gets silently invented. Each needs mapping to its OOv3 home (or to an SIA). In the paper going to review.
k4	s3	Two open items left in the Section-1 Arkhipov story (scan now read).	Opened 2026m08d06, largely closed the same day (llog `hell/ll/study/b/16/b16-section1-arkhipov-story-llog_2026m08d06_14h44.rst`). Seven sources researched and added to `source/_bib/b16-nuclear-risk.bib` under a cluster header carrying the standing warning; `check-bib` clean. The Orlov scan has been read and every figure the paper uses is confirmed verbatim. It forced two corrections, both applied: no water rationing appears in the document (claim CUT), and the three-officer vote and Arkhipov’s refusal are NOT in the source — Orlov says only that Savitsky reined himself in after consulting Arkhipov and Maslennikov. The paper now states that gap explicitly and uses the quieter sourced version; do not let the dramatic version back in. The Guardian article has also been read (LLoL supplied the full text the same day), forcing a third correction: it does not contain the “gone down with your ship” remark, which is the only thing it had been cited for. That beat was CUT; no source for it has been found and the Soviet debriefing records are closed. `Wilson2012` is now cited for what it genuinely is — a dated, widely read instance of the accretion, asserting the three-officer rule and the lone refusal, neither traceable to Orlov. One item remains: LLoL’s account attributes the radio silence to US broad-band jamming; no source found, and the Orlov scan gives only an inability to reach the General Staff. Do not restore it without a source.
k5	s3	Review the whole BibTeX arrangement for the Matheo series.	LLoL’s call, reserved 2026m08d04. The 2026m08d03–04 work restored the `.. bibliography::` block to OOv3 (with the non-capturing `:filter: docname in docnames`) and moved `Kemp2025` to `references.bib` under the promotion rule — but the wider question of how the series should cite itself is not settled. Related: POST AA b19 items 1 and 3 (seven pages still use the capturing `:filter: cited and True`; the inline display style for :cite:`Matheo-bNN` is still undecided).
k4	s3	Order, review, and consolidate the SIA numbering.	LLoL 2026m08d04. Headings are currently sequential (SIA-1..SIA-14, then SIA-X1..SIA-X7) so nothing is broken, but the two-series split (numeric vs X) needs a deliberate decision: what earns an X, whether the X-series should fold into the main sequence, and whether the in-text cross-references still point where they should. Editorial, LLoL’s judgement. Added 2026m08d07: SIA-X7 (the B-59 account, moved out of Section 1) now sits after SIA-X6, i.e. a narrative appendix after the draft-status bookkeeping one. That is the only placement that keeps the X-series in numeric order today. In the end-pass it wants to sit next to SIA-6 (crisis-rate calibration), which is where the Cuban excursion is counted and where a reader asks “what actually happened?” — and the draft-status appendix wants to be last.
k3	s3	[DONE 2026m08d04] Pre-submission checks (from the 2026m07d17 session, item #17).	All four HELD against their sources on 2026m08d04 — details in the “Done” section below. One residual, k3 s3: the Rome/AI words are the Rome Declaration’s, quoted via the Bulletin article (`DiazMaurin2026`). Citing the reporting article is defensible, but the primary source — Humanity at the Threshold: A Declaration on Artificial Intelligence and Nuclear Weapons — would be stronger. Decide whether to add it as a second bib entry.
k5	s3	Finalize publication via supersede-don’t-overwrite (mmv6 pilot).	From the 2026m07d17 session. Needs LLoL’s sign-off on the DD b15 §2 amendment (per-paper rolling marker). How-to in `AHA/matheo-naming-and-versioning.md` -> “Updating a floored paper”. After LLoL finishes reading.
k4	s3	[DONE 2026m08d05] Decide the empire-collapse paragraph’s citation.	§1 was reworded 2026m08d04 so the paragraph is explicitly non-load-bearing (“nothing in the forecast below depends on it”), which lowers the risk of citing loosely. On 2026m08d05 LLoL challenged the 2026m08d04 CITE-SCOPE WARNING and was right: elite capture running to brittleness is endogenous self-destruction, so the warning was one axis too wide. Narrowed: the prohibition applies to the attributed mechanism (epistemic vs structural), not to the self-destruction framing. Both proposals from the Kemp placement llog were applied on LLoL’s instruction: `Kemp2025` now supports the closing endogenous claim of the empire paragraph (§1, ~lines 231–247) instead of the middle clause, and one sentence was added to the burden-of-proof paragraph (~line 258) so that “collapse can be invisible until after it has run” answers the “surely we would notice” rebuttal from outside the paper.
k5	s3	[DONE 2026m08d05] Do not quote Kemp — paraphrase only, until a copy is read.	Everything known about Goliath’s Curse comes from secondary sources; no copy has been read. The bib `abstract` field is publisher marketing copy quoted verbatim, not Kemp’s prose — it is a non-rendered archive field (checked 2026m08d05 against built HTML: `abstract` does not appear in rendered bibliographies). Enforced in the applied §1 text: paraphrase only, no quotation marks, no page numbers. Now written into the bib `extranote` as a standing restriction on future edits, which is where it belongs — it binds anyone citing this source, not just b16. Lift only after someone reads a copy.
k4	s3	[DONE 2026m08d05] Amend the `Kemp2025` bib `extranote`.	Rewritten: the CITE-SCOPE section now supersedes the 2026m08d04 warning and separates the two axes (framing = safe, mechanism = still off-limits); the do-not-quote restriction is recorded; what b16 actually cites Kemp for is recorded so future editors need not re-derive it. The nuclear-war claim was confirmed, not removed — the final third argues we inhabit a single Global Goliath threatened by nuclear war, climate breakdown, AI and pandemics (two independent 2025 reviews). The “324 cases” figure was confirmed against Wikipedia. The “Added 2026m08d03” date was wrong and is corrected to 2026m08d04 against git.
k4	s2	[DONE 2026m08d05] Cite Kemp for b16’s own subject, not just the analogy.	Opened by the 2026m08d05 source check: Kemp’s final third treats nuclear war as a threat to the present global order. LLoL approved, with the constraint “only in a general sense — I don’t have the book and can’t check the precise quote.” Applied to SIA-10 Related work as a short paragraph placed after the Baum findings and before the Hellman block, so the section still closes on “this forecast is not an outlier”. It states the convergence, then the limit: that literature names the threat without estimating a rate, so it corroborates the subject and none of the numbers, and it leaves open the same how-often gap Baum names as undone and SIA-9 finds in the insurance record. No quotation, no page numbers.
k2	s1	[DONE 2026m08d05] Add the missing Sphinx-extension entries to `scripts/rstcheck.cfg`.	Bigger than first assessed. Roles: `cite`. Directives: `dropdown`, `bibliography`, `graphviz`. Repo-wide effect: 1907 findings \|rarr\| 615, i.e. 1292 of them — roughly two-thirds of the entire lint output — were false positives from four missing config entries. The b16 OOv3 paper went from 40 findings (39 `cite` + 1 `bibliography`) to zero. Earlier counts of “36” and “37” in this file and in session replies were low; the measured figure is 40. Trade-off accepted and stated: rstcheck no longer inspects `bibliography` directive bodies, and it never could resolve citation keys — `make check-bib` is what catches a bad key, not rstcheck.
k3	s2	Work down the 615 real rstcheck findings now that the noise is gone.	Top classes: 134 short title underlines, 85 inconsistent title levels, 72 unexpected indentations, 62 block quotes ending without a blank line. Mostly in llogs — see the POST AA b21 row on the 154 CRITICAL llog errors, which is the same debt seen from the Sphinx side. Not the same tool: rstcheck findings and Sphinx warnings are separate lists, and clearing one does not clear the other.
k3	s2	Decide the internal-link half of the §1 `TODO REF`.	The marker asks for two things; only the Kemp citation is answered. Natural target for “why societies self-destruct” is `source/jubileesystem/index.rst` (lines 86–93: debt accumulation, incumbent capture, persisting arsenals). That file has no RST label, so it would be a `:doc:` link. Open question for LLoL: should a paper under expert review link its introduction to the project’s own advocacy page?
k3	s3	[DONE 2026m08d04] Fix the stale display text in the restored Provenance paragraph.	Reads “What Changed Between MMv5 and OOv1” while targeting `b16-changelog-mmv5-to-oov2`. Ported verbatim from OOv2; the label drifted there.
k6	s9	[RAISED from k3/s2 on 2026m09d04] Resolve the two remaining ``TODO`` markers in §1 — they RENDER AS VISIBLE BODY TEXT.	Raised because they were checked in the built PDF rather than assumed. Both appear in the text layer of `b16-riskymad-waiting-time_oov3_2026m07d18_BalospeCom_2026m09d04_17h02m14.pdf` — they are not RST comments and a referee reads them, inside the first ~500 words, before reaching any arithmetic. Current source lines: `(TODO REF)` at 474, `(TODO LINK Jacobsen, 2024)` at 534 (the older ~219/~278 line numbers below predate the 2026m09d04 additions). Nothing else in the paper leaks a marker: the `--CONTINUE HERE---` markers and the EPIEIKEIA / P1–P3 alternatives blocks are comments and do not print — checked, not assumed. Nuclear-winter mechanism (`TODO REF`) and Jacobsen 2024 (`TODO LINK`). Toby Ord was resolved 2026m08d05 — `Ord2020` added to `references.bib` and cited at ~line 314 for the wider catalogue of existential risks the single-switch abstraction excludes; the stray `)` and missing full stop on that line were fixed with it. The societies-self-destruct marker was narrowed — its citation half is resolved by `Kemp2025`; only the internal-link half survives, tracked in its own row above.
k5	s3	[DONE 2026m08d05] Complete the §1 close (the two ``–CONTINUE HERE—`` markers).	The tail was doubled, not unfinished. Deleted on LLoL’s instruction: “The life being priced” (duplicated the better Actuarial mapping paragraph, which also states the under-estimate) and “The historical record” (duplicated the earlier Cuban/Able-Archer/Petrov/Arkhipov passage, which numbers the two sides (0)/(1) per the ordering convention). Both were also indented 12 spaces and would have rendered as nested block quotes. Fixed: the severed sentence in the Why this paper exists admonition (it ended `SIA-X1;` then three blank lines, orphaning “nothing in the argument depends on it”) and “More details … is described”. Final order: the Parkinson paragraph, the die, the question, the limit, the Nobel corroboration ending on `#AuditTheMath`, the admonition, then the roadmap.
k4	s3	[DONE 2026m08d05] Add the closing roadmap paragraph to §1.	Two paragraphs: one clause per section 2–11, then a general statement that depth lives in the appendices. Constraint honoured: the OOv3 restructure llog requires the Mary Douglas ring to stay latent — no `A/B/C/C'/B'/A'` headers, no forced mirror-pairings — so the roadmap names only the centre (Section 5, the decision, as the paper’s own metadata already does) and groups the rest by function. Per LLoL: no precise SIA survey, since the numbering is still to be consolidated. Placed after `#AuditTheMath` and ending on the hand-off to the appendices, so §1 still closes on the checking task rather than on a table of contents.
k3	s3	[DONE 2026m08d05] Resolve the parked “Decision to look” comment.	Woven back in as “The difficulty is not only personal”, with LLoL’s `((Parkinson, bike-shedding...))` placeholder resolved to Parkinson’s law of triviality (`Parkinson1957`, now in `references.bib`). The reactor-versus-bicycle-shed illustration is unusually apt for this paper’s subject; the text says plainly that the committee is fictional and the essay a satire, so it reads as an image and not as evidence. The verbatim original is kept in the comment with a dated header saying it has been integrated — so the wording stays checkable and nobody integrates it twice.
k3	s2	LLoL to check two claims the new roadmap makes.	“Every parameter on the table” (Section 3) and “the record needed to reproduce every figure” (SIA-12) are both checkable promises a reviewer can test. They were written from the section openings, not from an audit of SIA-12. Also for LLoL’s eye: the one-clause description of Section 5, whose voice is explicitly LLoL’s to set.
k3	s2	Two ``–CONTINUE HERE—`` markers remain in the OOv3 draft.	Lines ~290 and ~293 at 2026m08d03. LLoL was editing §1 when the session broke.
k4	s2	§2 restructure (2026m07d17 item #14) — LLoL’s structural thinking.	Deferred; no action until LLoL decides.
k5	s3	Pour the b16 floor and send the outreach (2026m07d17 item #18).	Baum + 10 drafts prepared, none sent. After LLoL’s review. This is the step that makes the revised paper public.
k5	s3	Sweep the UNSENT outreach drafts in ``balospe-launch`` for the 1-in-40 / car-crash formulation.	Propose-not-edit — LLoL’s public voice. Ten drafts carry it and none has been sent (confirmed 2026m07d16, even Schneier): `email/bruce-schneier_2026m06d23.txt` (5 hits), `Gideon300dev/letters/first-touches-3_2026m06d23.txt` (10), `.../rovelli-carlo_2026m06d25.txt` (8), `.../robock-alan_2026m06d24.txt` (2), `.../copp-tara_2026m06d25.txt`, `email/move-to-launch/{former-colleagues-science,-COLD,coming-out-to-colleagues,forwardable-science-chain}`. GOOD NEWS FIRST: the drafts are NOT wrong, and they vindicate LLoL. Every one uses “worse than a 1-in-40 chance per year” — a floor, not a constant — and none claims scenario-independence. The floor survives exactly: the base case is 1 in 34.5, which is worse than 1 in 40. The uniform-1-in-40 claim LLoL worried he had made exists only in the Claude-drafted paper prose, not in his outreach. He owes no correction there. TWO things to fix, both upgrades rather than repairs. (1) They can now say something stronger: 1 in 34, computed exactly from the first-passage law, rather than sampled from 40 runs. (2) The real defect: the sentence “likelier to die in an accidental nuclear winter than in a car crash — worse than a 1-in-40 chance per year” puts the two clauses in apposition, so the second reads as the basis for the first. It is not. 1-in-40 is $P(\text{onset})$; the car-crash comparison requires $P(\text{onset}) \times q$. That is precisely the conflation b16 §2.6 warns of (“would overstate the comparison by a factor of $1/q$”). Alan Robock will notice within seconds — it is his field (his and Xia’s 2022 calorie-loss estimates are what b16 now cites for $q$). Fix by naming the $q$ step, or by dropping the em-dash clause and citing the 58x comparison directly. Also stale in those drafts: “median 19 years” → 21.1 exact (a refinement, not an error); the US road-death baseline → global (80x → 58x). The newer `email/seth-baum_2026m07d16.txt` is already written on the corrected numbers.
k4	s2	Sweep the public site pages once the paper’s numbers settle.	`crisis/wagers.rst` (5+ hits on “worse than 1 in 40”), `crisis/science.rst`, `crisis/see-the-problem.rst`, `buy-in/campaign/intro-gofundme.rst`, and the b16 intro companion. The base case supports “worse than 1 in 40”; what needs softening is the “regardless of scenario” framing, since the optimistic corner is ~1 in 113. Deferred by LLoL: “let’s fix the paper first.”
k4	s3	Read Lewis et al. (2014) and Tertrais (2017) properly.	PDFs are in the FF review-prep folder; bib entries pending. The worst-case table currently rests on Claude’s inference about what a Lewis-style reading would include, drawn from Baum’s characterisation rather than from Lewis et al.’s own text. Until read, the table must not claim their authority. Tertrais’s scope note (he excludes accidents, unauthorized launches, and terrorism) is load-bearing for the prior-art section and should be checked against his own words, not the abstract.
k3	s3	Ask A. W. McCoy which four near-misses he counted.	McCoy (UW–Madison historian) supplied the count “4 in 40 years” by personal communication — not a list. LLoL’s instruction (2026m07d15): keep McCoy out of the paper for now and do not attribute a composition to him. If his four differ from the paper’s reconstruction from Baum, that difference is itself publishable evidence on how fluid the A/B boundary is among people who know the history.
k3	s1	Watch the c104 re-pour trigger.	bug c104 predicts a series-wide `mmv6` re-pour becomes cheaper than explaining the decoupling once three or more papers stand at OOv1+ while the floor still reads `mmv5`. Count after this revision: two (b19 PPv1, b16 OOv1). Record the count when it moves; if it passes three and a re-pour still does not pay, c104 is wrong and that must be recorded.
k2	s2	Migrate deprecated tokens carried by the b16 floor.	The floor still carries `[Matheo-2]`-style references and old `h*`-era tokens; the paper flags this itself. Executed as part of the wider site sweep (see “Larger than b16” below), but listed here because it is b16 output that a reader sees.
k4	s2	Expression restructure — fold audience in, facet (2026m07d17 item #10).	De-risked plan recorded in that session + `AHA/matheo-naming-and-versioning.md`. Recommendation from that session: after b16 publishes, not before.
k3	s2	Registry-doc scope + the subfolder question (2026m07d17 item #15, ref #18).	LLoL’s call.
k3	s1	FOOTER-FORM vs content-VVN collision (2026m07d17 item #16).	Also tracked as AA b15.
k6	s9	[DONE 2026m09d04] Set the stake band $E$ in Table 10.	LLoL’s ruling: $90,000 / $900,000 / $9,000,000, keeping the hundredfold spread. Applied throughout — Table 10, Table 11’s ceiling column, SIA-15’s worked example and cross-check comparison, the Abstract, both Broader Significance blocks and Section 11. What this bought. The ceiling now rounds down from the DOT guidance value of $9.1 M rather than up from it, so no rounding anywhere in Table 10 runs in this paper’s favour — which was the one exception flagged in the adversarial pass. The round ceilings survive the change: exact thresholds are 1 in 326 / 3,263 / 32,628, stated as no more than 1 in 300 / 3,000 / 30,000; the hundredfold column is 30.6 / 3.1 / 0.31 per cent, stated as no more than 35 / 3.5 / 0.35. LLoL’s rationale for the hundredfold spread is now IN the paper, and it is stronger than the one the draft had. The draft justified the width as memorability. LLoL’s reason is defensive: what a life is valued at differs enormously between countries, so any single figure invites the objection that it is somebody else’s — the band is set wide on purpose to span that international spread, and the conclusion holds at every point in it. Added, together with the observation that a single per-person figure is meaningful here at all only because the hazard is global and non-excludable: everyone is in the same boat, which the arithmetic assumes and nothing more. On LLoL’s question of moving $K$ from $8 to $9 for a pleasant ratio — recommended AGAINST, and it is a close call worth recording. The attraction is real: with $K = 9$ and $E = 9 \times 10^{n}$ the nines cancel exactly, giving $p^{} = 1/(r \times 10^{n})$ — thresholds of 1 in 290 / 2,900 / 29,000, read straight off the annual hazard. Two costs outweigh it. (1) It loses the clean ceilings.* At 1 in 290 the round figure 1 in 300 would understate the bar (1/300 is smaller than 1/290), so the prudent ceilings become 1 in 250 / 2,500 / 25,000 — uglier than what $8 already gives. (2) $K$ is derived, not chosen: $64 bn a year over eight billion people is $8, and adjusting it to make an output round is a calibration move. Unlike the $10 M ceiling, its direction would at least be safe — over-stating a cost is prudent — so this is a much milder case than the $720,000 one, and it is recorded rather than dismissed. If LLoL wants the cancellation anyway, the honest route is to change the ResearchCity budget to $72 bn and let $9 fall out of it, not to round $8 up for the arithmetic’s sake.
k6	s9	[DECIDED 2026m09d04 — one follow-on] How openly the paper reports that the sent $E = 720{,}000$ was fitted backwards.	Decided: not in the paper at all. LLoL’s ruling, on being shown the draft passage: “I don’t understand why $720,000 is in the paper at all … The paragraph you showed sounds quite performative.” Correct. A general reader has never seen the German letter and neither has a referee, so the passage explained a parameter the paper does not use, from a document the reader does not have, to disclose a defect in something b16 never claimed. It also broke this project’s own rule — the MMv5 → OOv2 changelog opens by saying “This page exists so that the paper does not have to carry its own history in its margins.” All three drafting-history passages are now out of the paper, each keeping its substance: the factor-table guard is stated forward ($\eta = 2$, $L = 50\%$ returns exactly 2 “for that pairing and no other”); the provenance box is replaced by “Why this appendix reports a band and not a number”, explaining the paper’s own two routes rather than the letters; and SIA-16’s review-versus-funding distinction stands on its own merits. Grep confirms zero occurrences of `720`, `German-language` or `enclosure` in the paper. FOLLOW-ON, still open. The fitted-$E$ finding and the four-versus-five-percent resolution are real and must survive somewhere a reader can reach. They are in the session llog (`b16-erlangen-decision-arithmetic-import-llog_2026m09d04_11h30`, Finding 1). When this revision is versioned, a changelog entry is needed — the OOv3 → next changelog, in the series of `b16-changelog-mmv5-to-oov2` and `b16-changelog-oov2-to-oov3`. That is the right home: provenance, in full, outside the paper. Tie this to the supersede-don’t-overwrite row.
k5	s8	[DONE 2026m09d04] Source the two scale figures used to make $64 bn legible.	Opened and closed the same day. Both now rest on one source: `SIPRI2025Milex` gives world military expenditure of $2,718 bn in 2024 and a global military burden of 2.5 percent of world output, which fixes world output near $109 tn without a second citation. So $64 bn is about one fortieth of military spending and about six parts in ten thousand of world output, both rounded so as to over-state the cost. The remembered ~$2.4 tn figure the launch project flagged was close but stale; 2024 is $2,718 bn. See the new row below — the SIPRI entry’s author list is still unconfirmed.
k6	s9	[DONE 2026m09d04] Confirm the new decision-arithmetic bib entries.	Nine entries added to `source/_bib/b16-nuclear-risk.bib` for SIA-15 and SIA-16; `check-bib` clean, all nine resolve in the built HTML and PDF. All closed the same day. Two carried real errors, both caught only because LLoL asked; both are recorded in the entries’ `extranote` rather than silently overwritten. ``USDOT2021VSL`` — was wrong twice, now read from the document. The first version cited a supposed 2024 update at $13.2 M in 2023 dollars, taken from secondary sources because transportation.gov returns HTTP 403 to every automated request (tested via WebFetch and via curl with a browser user-agent, and on the direct PDF link too). It also asserted DOT “revises this every year”. Both were wrong. LLoL supplied the actual PDF (`balospe-launch/CC/DOT VSL Guidance - 2021 Update.pdf`): the document is the March 2021 update, Departmental Guidance: Treatment of the Value of Preventing Fatalities and Injuries in Preparing Economic Analyses, 11 pp. DOT reviews the guidance periodically since 1993, not annually. What the base year is actually for: the guidance sets a $9.1 million anchor in 2012 dollars (2013 update, average of nine studies) and has the analyst escalate it by `VSL_T = VSL_0 x (P_T/P_0) x (I_T/I_0)^e` — CPI-U for prices, BLS Median Usual Weekly Earnings for real incomes, income elasticity e = 1.0. That is why escalated values for various base years circulate in secondary sources, and why a quoted VSL is meaningless without its base year. b16 now cites the anchor, which does not age, and Table 10’s top row reads “the order of the US Department of Transportation’s own guidance value of $9.1 million in 2012 dollars”. ``SIPRI2025Milex`` — author list was a guess and was wrong. Entered as four names led by Tian; the fact sheet has six led by Liang: Xiao Liang, Nan Tian, Diego Lopes da Silva, Lorenzo Scarazzato, Zubaida A. Karim, Jade Guiberteau Ricard. Corrected. Month, the $2,718 bn total and the 2.5 per cent military burden all confirmed on the same page. Closed against the Crossref API: `Evans2005` and `Howard1966` fully confirmed; `Pratt1964` confirmed except the end page (Crossref deposits only the start page 122); `GroomMaddison2019` confirmed on title, volume, issue and pages, with two deposit artefacts recorded (Crossref year 2018 for online-first against the 2019 issue; “Maddison Pr., David” is a deposit error for David Maddison). Two judgement calls, decided with LLoL’s approval 2026m09d04. `Arrow1965` → ``Arrow1971``, Essays in the Theory of Risk-Bearing (Markham, Chicago, 278 pp, checked against OpenLibrary) — it reprints the 1965 Yrjö Jahnsson Lectures and a referee can actually obtain it. `NeumannMorgenstern1944` pagecount 625 \|rarr\| 641, per the OpenLibrary record for the 1944 Princeton edition; editions differ (1980: 648), so it is edition-specific and flagged as such. Recorded so a deliberate omission is not later mistaken for an oversight: the HM Treasury Green Book was NOT cited. It would have supported “$\eta$ = 1 in official guidance”, but a 2022 and a 2026 edition both exist and it could not be confirmed which carries that value. `Evans2005` and `GroomMaddison2019` already carry the claim, so the unverifiable source was cut rather than cited loosely.
k4	s7	Consider whether Section 9’s new second half should carry a figure.	Opened 2026m09d04. The break-even is a one-line inequality with a two-order-of-magnitude robustness claim attached, and the paper’s other quantitative claims all carry a figure. A single plot of $p^*$ against $r$ on log axes, with the base rate, the hundredfold discount and the stake band drawn on, would make the invariance visible at a glance. Not blocking — Table 10 carries the content — but Section 9 is now the only heavily numeric ring section without one.
k3	s3	Decide whether the public `crisis/insurance` page is superseded by SIA-15.	Opened 2026m09d04. `balospe-launch/AAA-to-publish-at-Balospe.com/DecisionArithmetic/` holds a port-ready page targeted at `source/crisis/insurance.rst` that carries the expected-utility route (the coin-flip inversion, the sqrt utility, the 1-in-2,600 and the factor-100 table at 4.6 %). It has not been ported. Now that SIA-15 carries the same route as a cross-check and reports the band rather than the point, porting it unchanged would put a second, sharper set of numbers on the public site — exactly the drift this session was called in to stop. Either re-sync it to the band or leave it unported. Cross-project; belongs to whoever moves it.

Larger than b16 — tracked elsewhere, listed here only so they are not lost#

Per the placement rule, these do not belong in this file. Pointers only:

Migrate the deprecated ``Matheo-N`` citations (2026m07d17 item #11). 116 occurrences across 22 files; renders as empty [] on 8 built pages including the published study/matheo/b15/b15-math-deadlock-mmv5. b16 OOv3 is clean and not affected. Tracked as POST AA b19.
Cross-page bibliography capture. Seven pages still use the capturing :filter: cited and True. Same entry, item 1. This is what had b16’s citations landing in b15’s reference list until 2026m08d04.
Build-warning gating and link guards. AA b18.

Done (kept, not deleted — evidence of what was attempted)#

✅ 2026m08d04 — Pre-submission source checks: all four HELD.
- Nobel 2025, “in an afternoon” — HELD verbatim. The Chicago Declaration for the Prevention of Nuclear War (p.1) reads: “But there is only one challenge that could end civilization in an afternoon.” The paper quotes only the fragment "in an afternoon", so its framing verb (“can”) sits outside the quotation marks and misquotes nothing — though could would track the source more closely if the sentence is ever reworked.
- Rome/Vatican 2026, AI and human judgement — HELD verbatim: “AI built into nuclear systems leaves little time for, or even replaces, human judgement in a crisis.” Note the words belong to the Rome Declaration (Humanity at the Threshold), reported by Diaz-Maurin and Mecklin, Bulletin of the Atomic Scientists, 2026m07d17.
- Guinness, “deadliest place to travel by road” — HELD. Record statement: “The World Health Organisation (WHO) named Libya as the nation with the highest rate of road traffic deaths, with 73.4 per 100,000 people in 2013.” Guinness cites WHO Global status report on road safety 2015.
- Libya data-year — already pinned and correct. The paper reads “~73.4 per 100,000 (2013)” at the stress-test baseline, and both bib entries record 2013 data from the WHO 2015 report. Independently confirmed via Guinness. No change needed.
✅ 2026m08d04 — Remapped the ported Declarations to OOv3 structure. The verbatim port carried OOv2 numbering that OOv3 does not have. Mapping applied: Section 2.5 (Evolvix simulations) → SIA-1; Section 4 (ResearchCity) → Section 6; “headline table” (sceptical critic’s rate) → Table 3 in SIA-5, with Tertrais’s corner worked through in Section 10; Section 2.6 (every soft input pushed against the conclusion) → Table 7 in SIA-9; Section 4.0a (reverse exposure) → SIA-X5; Section 2.0a (headline figure) → Section 4. The Provenance link text now reads “MMv5 and OOv2” (matching its target) and a second link to the OOv2 → OOv3 changelog was added.
✅ 2026m08d04 — Restored the closing Declarations + ``.. bibliography::`` to OOv3. The OOv2 -> OOv3 port had dropped both. Consequence while missing: all 31 :cite: calls resolved into another paper’s reference list (study/matheo/b15/b15-math-deadlock-mmv5), and OOv3 rendered no references of its own.
✅ 2026m08d04 — Converted the one ``:cite:`Matheo-b19``` to ``:doc:``, matching the 15 sibling :doc:`Matheo-bNN` refs already in the paper. Provisional, without prejudice to the BibTeX review above.
✅ 2026m08d04 — §1 empire-collapse paragraph reworded to shift from an unprovable assertion (“the quiet since WWII is merely the calm before a firestorm”) to a burden-of-proof argument built on Arkhipov’s asymmetry-of-consequences logic, with the forecast explicitly decoupled from it.
✅ 2026m08d04 — ``Kemp2025`` added and promoted to references.bib.
✅ 2026m07d17 session — see AHA/aa-b16-session-2026m07d17.md for that session’s DONE list (Nobel refs, GAO-09-39, Libya 73.4, FRBR, link-in-note robustness, supersede-don’t- overwrite decision).