Note

LLog: b16 Nuclear-Risk Bibliography Entries (2026m07d15). Audit trail for adding the six prior-work nuclear-war risk-analysis references supplied by LLoL to a new paper-local bib file for b16 (RiskyMAD). LLog by Claude Opus 4.8, max effort (b16-bib-dv_ClaOp48Max_MMv1r0p0_2026m07d15_12h52).

LLog: b16 Nuclear-Risk Bibliography Entries — 2026m07d15#

VVN: b16-bib-dv_ClaOp48Max_MMv1r0p0_2026m07d15_12h52
Date: 2026m07d15_12h52
Effort: Max (from .claude/effort-level, not confirmed by LLoL this session)
Mode: EDEN (from .claude/mode, not confirmed by LLoL this session)
Session: Ingest six supplied PDFs as formal bib entries for the b16 revision

Prompts (Verbatim)#

Prompt 1 — initial task#

You are working with LLoL to finalize an upgraded revision of the Matheo-b16 paper to be ready for revision. – Here are some PDFs that I downloaded (with source links in the email); I want you to include them properly in the bibtex file for being able to properly cite them in the b16 Matheo paper. I will later ask you to integrate these in whatever b16 lists as “previous work”, but for now simply integrate all these as Journal articles from websites etc (but not Misc) along with all proper citation info (checked on the web): /Users/llol/LLoL-Repos/SethGitHubSetup/__Balospe-com-FF/FF_2026-07-13 b16-Review-prep/sgs21barrett–2013.pdf /Users/llol/LLoL-Repos/SethGitHubSetup/__Balospe-com-FF/FF_2026-07-13 b16-Review-prep/sgs21barrett_app–2013.pdf /Users/llol/LLoL-Repos/SethGitHubSetup/__Balospe-com-FF/FF_2026-07-13 b16-Review-prep/P-nuke–ssrn-3137081–2018.pdf /Users/llol/LLoL-Repos/SethGitHubSetup/__Balospe-com-FF/FF_2026-07-13 b16-Review-prep/Hellman-2008—Bent–74.pdf /Users/llol/LLoL-Repos/SethGitHubSetup/__Balospe-com-FF/FF_2026-07-13 b16-Review-prep/Ch4_Hellman–2021.pdf /Users/llol/LLoL-Repos/SethGitHubSetup/__Balospe-com-FF/FF_2026-07-13 b16-Review-prep/Baum2019_ReflectionsontheRiskAnalysisofNuclearWar.pdf /Users/llol/LLoL-Repos/SethGitHubSetup/__Balospe-com-FF/FF_2026-07-13 b16-Review-prep/2nd—-Textandpdflinksforb16.txt

Prompt 2 — ResearchGate citation export#

and here is the link you were not allowed to access: the citation from researchgate is https://www.researchgate.net/publication/277283323_Risk_Analysis_of_Nuclear_Deterrence here at /Users/llol/LLoL-Repos/SethGitHubSetup/__Balospe-com-FF/FF_2026-07-13 b16-Review-prep/citation-277283323.bib

Prompt 3 — page range confirmation#

The pages are p.14-22

Prompt 4 — build instruction#

Please don’t run make dev any more; there seems to be something wrong with that (as before, super long running times that never end); I’ll rather run it by hand; let me know when you think it needs running and what I should look for.

Files Read#

Supplied source material (all read):

  1. FF_2026-07-13  b16-Review-prep/2nd----Text and pdf links for b16.txt — LLoL’s email to himself carrying the source links and a framing note.

  2. sgs21barrett--2013.pdf — title page (citation block, abstract).

  3. sgs21barrett_app--2013.pdf — title page.

  4. P-nuke--ssrn-3137081--2018.pdf — pp. 1–2 (title page with “Cite as” line, abstract).

  5. Hellman-2008---Bent--74.pdf — cover, pp. 14–15, pp. 21–22 (references).

  6. Ch4_Hellman--2021.pdf — pp. 85–86, 102–104, 124–126; full text extracted via pdftotext to locate the headline risk estimate.

  7. Baum2019_ReflectionsontheRiskAnalysisofNuclearWar.pdf — title page.

  8. citation-277283323.bib — ResearchGate export for Hellman 2008.

Project conventions:

  1. .claude/CLAUDE.md; .claude/effort-level; .claude/mode.

  2. AHA/bibliography-management.md — Rules A / B / 0 / C / D / N, entry-type traps.

  3. AHA/vvn-composition.md — VVN format.

  4. source/_bib/b19-epidemiology.bib — paper-local bib precedent.

  5. source/_bib/matheo-refs.bib; source/conf.py (bibtex auto-discovery).

What Was Done#

Created source/_bib/b16-nuclear-risk.bib — a paper-local bib file holding six new entries, following the b19-epidemiology.bib precedent and its promotion rule. conf.py auto-discovers every .bib in source/_bib/, so no conf.py edit was needed.

The six entries#

Key

Type

Work

Notes

Barrett2013

@article

Barrett, Baum & Hostetler (2013), Science & Global Security 21(2), 106–133

DOI 10.1080/08929882.2013.798984; all fields checked against Crossref

Barrett2013b

@article

The online appendix and supplement to the above

No DOI; visible note marks it as the appendix

Baum2018

@techreport

Baum, de Neufville & Barrett (2018), GCRI Working Paper 18-1

The 60-incident dataset lives here

Baum2018b

@incollection

Baum (2018), in Garrick (ed.), Proceedings of the Workshop on Quantifying Global Catastrophic Risks, 19–50

2018, not 2019 — see Traps below

Hellman2008

@article

Hellman (2008), The Bent of Tau Beta Pi 99(2), 14–22

Pages confirmed by LLoL and against the PDF

Hellman2021

@incollection

Hellman (2021), ch. 4 in Scouras (ed.), On Assessing the Risk of Nuclear War, 85–126

Publisher: JHU APL, Laurel, MD

No @misc was used, per LLoL’s instruction. Every entry carries url, abstract, pagecount (Rules A / B / 0), with internal annotation in extranote (Rule N).

Checking done (not assumed)#

  • Crossref API confirmed Barrett2013’s author list, title, volume 21, issue 2, pages 106–133, and year for the DOI.

  • All eight candidate URLs were fetched and their status codes recorded. The Taylor & Francis DOI target and both SSRN links return 403 to curl (anti-bot), but the DOI itself resolves (302 to tandfonline) and the SSRN links resolve (301 to www.ssrn.com); all other URLs return 200.

  • pybtex parse + APA render test was run directly on the new file. This caught four real defects that inspection alone would have missed — see Testing Found below.

  • Duplicate-key check across all four bib files: 125 keys total, no duplicates.

  • ``make dev``: build succeeded, exit 0, 850 warnings (the site’s pre-existing warning load). Per Prompt 4, LLoL runs builds by hand from now on.

Testing Found (four defects, all fixed)#

Rendering the entries through pybtex-apa-style before declaring the work done exposed:

  1. ``url`` is silently dropped for ``@techreport`` and ``@incollection``. pybtex-apa-style only emits url for @article. Three of six entries would have rendered with no link at all — defeating the purpose of Rule A while still passing a Rule-A field audit. Fixed by repeating the URL inside a visible note, which is exactly what the existing Ferguson2020 entry in b19-epidemiology.bib already does. Confirmed against built HTML that \url{...} inside note renders as a clean clickable link (the \url leakage seen in pybtex’s plain-text backend does not occur on the Sphinx path).

  2. ``month = {mar}`` rendered as (2018 , mar) — stray space, unexpanded month. Removed the month fields.

  3. The quoted inner title in ``Barrett2013b`` was being lowercased to “appendix and supplement to “analyzing and reducing…””. Fixed by bracing the inner title.

  4. Underscores in URLs are safe (sgs21barrett_app.pdf renders correctly), so they were left unescaped. This contradicts AHA/bibliography-management.md, which says to escape _ as \_ in URLs. The one existing precedent in the repo (Kurt_G\%C3\%B6del_... in references.bib) also leaves _ bare while escaping %. See Open Questions.

Item 1 is the important one: it is precisely the “silent non-compliance — form looks right but substance has drifted” pattern CLAUDE.md warns about, and it is invisible without a render test.

Traps recorded in the entries#

  • Baum2018b year. The supplied PDF is named Baum2019_... and says “This version 22 January 2019”, but the workshop proceedings are December 2018 and Baum’s own citation line reads “Seth D. Baum, 2018”. The entry is 2018; the trap is recorded in its extranote.

  • Baum2018b pagecount. 32 pp. is the published extent (19–50); the free author copy is reformatted to 20 pp. pagecount follows the published extent per Rule 0.

  • Hellman2008 has no abstract (magazine feature) and Barrett2013b has no abstract (supplementary material). Both use abstract = {} per Rule B, with contents described in extranote so a reference-checker is not left blind.

ZION / BABL Analysis#

Green Meadow #1 — the bibliographic task itself#

The task is well-posed and the repo already carries a tested rule set (AHA/bibliography-management.md) plus a working precedent (b19-epidemiology.bib). The rules constrain the result without uniquely determining it, and every candidate path is checkable by a render test rather than by authority. guess = many (roughly three independent choices — file placement, entry type, URL preference — across six entries). Three diverse alternatives that also stay in ZION:

  1. Put all six into references.bib instead of a paper-local file. Violates no rule; rejected because the b19 precedent plus the promotion rule (“move it when a second paper cites it”) keep the shared file uncluttered until sharing is a fact rather than a guess.

  2. Use the SSRN links as the primary url for the two GCRI items (that is where LLoL got the PDFs). Rejected because Rule A(2) ranks the publisher’s landing page above a repository, and gcri.org returns 200. Both SSRN links are preserved in extranote, so nothing is lost.

  3. Type the appendix as @misc. Excluded by LLoL’s instruction, and independently the weaker choice — @misc is where the known url-not-rendering problem is worst.

Knife Edge #1 — what Hellman actually estimates#

Here there is exactly one honest path, and the surrounding paths are all BABL.

The email in Prompt 1 states: “Hellman’s 2008 model puts the probability of this class of nuclear war at 2×10⁻⁴ to 5×10⁻³ per year — his upper bound is half the ~1% you land on. So your revised figure is still a higher-risk claim than the literature, just a defensible one now. The spec tells you where you have to argue for the gap.”

Reading the supplied PDFs shows this characterisation does not hold:

  • The 2008 interval (2E-4, 5E-3) is Hellman’s estimate for lambda_CMTC — the failure rate via one mechanism, a Cuban-Missile-Type Crisis. Hellman says so on p. 21 of the Bent article: the analysis “neglects other trigger mechanisms such as command-and-control malfunctions and nuclear terrorism” and therefore “underestimates the threat”. It was never his estimate of total risk.

  • Hellman 2021 (chapter 4, the PDF LLoL supplied) lands on ~1 percent per year for a major nuclear war, and says so in the chapter abstract: “the risk of nuclear deterrence failing currently appears to be on the order of 1 percent per year.” Pages 97–99 give the range: a third of a percent to 3 percent per year, upper-bounded at 10 percent per year and lower-bounded at 0.1 percent per year. The concluding remarks repeat it.

So b16’s ~1 percent does not exceed the literature; it coincides with the most recent published estimate by the very author the email treats as the ceiling. There is no gap to argue for.

Why the alternatives are BABL:

  • Follow the spec as written — build a defensive argument for a “gap” that does not exist. This is over-Complicating (machinery to defend an unnecessary claim) resting on an over-Simplification (“the literature” = one 2008 number). It would make b16 weaker and less accurate while looking more rigorous. A LIE that would breed SINs in every later section that leans on it.

  • Cite only Hellman2008 because a lower prior number makes b16 look bolder — over-Reach by cherry-picking, and the reference-checker who opens Hellman2021 finds it in one minute.

  • Overstate the agreement — claim Hellman “confirms” b16. He does not: he reaches ~1 percent by a different route (order-of-magnitude bounding from crisis frequency), not via RiskyMAD’s three-state Markov chain. Convergence of two independent methods is worth something precisely because it is not identity.

The one path in ZION: cite both, state plainly that the 2008 figure is a single-mechanism lower bound that its own author flags as an underestimate, and let b16’s ~1 percent rest on the independent convergence with Hellman 2021. That is a stronger position than the spec assumes, and it is the one the sources actually support.

EDEN classification#

  • Green Meadow #1 for the bibliographic task (guess = many; three alternatives given above).

  • Knife Edge #1 for the treatment of Hellman’s numbers in b16’s previous-work section.

Both hold on evidence quoted verbatim above from the supplied PDFs, with page numbers, so they are offered not to be believed, but to be checked.

BABL Danger — the revision spec’s premise#

BABL Danger: the revision spec at FF_2026-07-13  b16-Review-prep/1st----b16-revision-spec.rst reportedly tells b16 “where you have to argue for the gap” between its ~1 percent and the literature. That gap is an artifact of reading Hellman 2008’s single-mechanism figure as if it were his total-risk estimate, and of not reading Hellman 2021 — which was downloaded in the same batch and states ~1 percent per year in its own abstract. Acting on the spec unchecked would write a false premise into the paper.

What LLoL can do: before the previous-work integration, decide explicitly whether b16 claims agreement-by-independent-convergence with Hellman 2021 (supported) or a higher-risk claim than the literature (not supported by these six sources). The spec’s section on “arguing for the gap” should be revisited rather than implemented as written.

This llog does not modify b16 or the spec. Nothing beyond the new bib file was written.

Summary#

  1. Done: source/_bib/b16-nuclear-risk.bib created with six entries (Barrett2013, Barrett2013b, Baum2018, Baum2018b, Hellman2008, Hellman2021). No @misc. Rules A / B / 0 / N followed. Auto-discovered by conf.py; no other file touched.

  2. Tested, not assumed: Crossref check, eight URL fetches, pybtex APA render test, duplicate-key check across 125 keys, and a full make dev (succeeded, exit 0). The render test caught four defects, including three entries that would have rendered with no link at all.

  3. Key finding (Knife Edge #1): Hellman 2021 estimates ~1 percent per year — the same figure b16 reaches. The email’s claim that Hellman’s upper bound is “half the ~1% you land on” compares b16 against a single-mechanism 2008 figure that Hellman himself flags as an underestimate. b16’s number agrees with the current literature rather than exceeding it.

  4. Two year/extent traps recorded in extranote: Baum2018b is 2018 despite the Baum2019_... filename, and its 32 pp. is the published extent (author copy: 20 pp.).

Recommendations#

  • k5 s3 — Revisit the spec’s “argue for the gap” section before the previous-work integration. See BABL Danger above. This changes what b16 argues, so it should be settled first.

  • k4 s3 — Add Hellman’s March 2021 *Bulletin of the Atomic Scientists* paper to the bib. It is the actual source Hellman credits for the 1 percent figure (chapter 4, notes 40–41); it was not among the six supplied PDFs. Needed if b16 leans on the 1 percent convergence.

  • k3 s2 — Fix or soften the ``_``-escaping advice in AHA/bibliography-management.md. Testing shows bare _ in URLs renders correctly and the doc’s advice to escape it as \_ is not what the repo actually practises. See Open Questions.

  • k3 s2 — Add a Rule to ``AHA/bibliography-management.md`` recording that pybtex-apa-style drops url for @techreport / @incollection / @misc, and that the fix is to repeat the URL in a visible note. This is currently tribal knowledge living in one Ferguson2020 entry and one memory note.

  • k2 s1 — Update the assessor registry in AHA/vvn-composition.md: it lists ClaOp47Max as “current default”, but the model is now Opus 4.8. This llog uses ClaOp48Max by mechanical extension of the documented pattern; LLoL to confirm.

Open Questions for LLoL#

  1. Bib file placement. Paper-local b16-nuclear-risk.bib (chosen, per b19 precedent) or straight into references.bib? Barrett2013 / Hellman2021 are plausibly cited by b18’s call to action too, which would trigger the promotion rule immediately.

  2. ``_`` escaping in URLs. The AHA doc says escape as \_; testing and the one existing precedent both say bare _ is fine. Which is authoritative — should the doc change?

  3. VVN assessor nickname ClaOp48Max — confirm, and update the registry?

  4. Effort / Mode were read from file, not confirmed this session: Effort: max. Mode: EDEN. Confirm or change?

Build Instruction (per Prompt 4)#

No build is needed for this change right now. The six entries are not yet cited by any page, so nothing renders them; the bib file’s parse was already tested directly with pybtex, and the completed make dev (exit 0) confirmed it does not break the build.

A build will be needed when the previous-work integration adds :cite: calls and a .. bibliography:: directive to b16. At that point:

  • Run make clean && make html, not an incremental build — Rule D: sphinxcontrib-bibtex caches citation targets in the doctree, and an incremental build can keep serving a stale, wrong-page link even after a filter is fixed.

  • What to look for: (a) each [Author Year] link on the b16 page lands in b16’s own References section, not another study’s — the Rule D cross-page capture bug; (b) all six entries show a clickable URL (the three non-@article entries get theirs from the note field, so a missing link there is the regression to watch); (c) no “unbalanced braces” error, which is how an unescaped % announces itself without naming the entry.

  • b16’s bibliography directive must use :filter: docname in docnames, never cited and True.

On this page
Recent Posts
Tags