Late Lessons, Jensen Huang and AI

Revision log: 01-late-lessons-analysis.md#

Revision of 26 September 2026, in response to the fidelity, balance and completeness reviews#

Each issue was checked against the working files in working/late-lessons/ and, where needed, the source chunks in working/text/chunks/ (with a phrase finder that normalises quotes, dashes and line breaks). Outcomes: fixed, fixed (partly) or rejected, with reasons. No issue was rejected outright; partial adoptions are explained. The document grew from about 29,000 to about 45,000 words; nothing of substance was removed except text the reviews showed to be wrong.

Fidelity review (fidelity.md)#

# Issue Verified? Outcome
1 Lead-death rise attributed to modelling cardiovascular effects Yes (hindsight LL2-23: rise came from changed exposure models; 5.5m CVD estimate is separate) Fixed in §4.5, §5.4 (LL2-23), §5.5 item 2 and Appendix A (LL2-23)
2 Derivation of the twelve lessons omits the Preface’s “distilled” account; “cases illustrate rather than test” stated as fact Yes (LL1-16, p. 168; LL1-00, p. 3) Fixed in §3.2, §4.10, §5.6 (labelled Analysis) and Appendix A (LL1-16)
3 “Chapter 28’s three statistics do not reproduce” overstates Yes (hindsight LL2-28: one misread, one unverifiable with plausible 1–2%, one never re-analysed) Fixed in §3.5, §5.4 (LL2-28), §5.5 item 2 and Appendix A (LL2-28)
4 Northern cod reference point “lowered about 40%” contested across working files Yes (hindsight LL1-02 arithmetic implies ~60% cut; DFO 2024 wording implies ~29%) Fixed in K5 and Appendix A (LL2-17): percentage replaced by DFO’s wording and a 30–60% range. Not fixed: themes/T07 (l. 219) and hindsight LL2-17 (l. 41) still say “about 40%”; outside this task’s scope, flagged for correction
5 “Moving target” quoted as the report’s term Yes (phrase absent from all chunks) Fixed in §3.4, §4.7, K11 and Appendix A (LL2-28); now “what this document calls the moving-target problem” with the real quotation
6 Mobile-phone verdict applies symmetry to one side only Yes (hindsight LL2-21: abstract-only access; ICNIRP and study-author roles; persisting signals) Fixed in §5.4 (LL2-21) and Appendix A (LL2-21); verdict now tied to independent cohorts and incidence data
7 “Exceptions” to “no Part A company or regulator voice” are not exceptions Yes (LL2-16 is Part B; Guidotti and Castaño are not company or regulator voices) Fixed in §5.1 item 2. Not fixed: T09 l. 176 (outside scope)
8 “made causation undeniable” is not in the source Yes (LL2-08, p. 189: “undisputed”) Fixed in §4.1
9 Eight misquotations or quoted paraphrases Yes, all eight Fixed: “when no evidence was actually being sought” (§4.1); “paralysis by analysis” / “prudent and careful evaluation” (§3.2 note 12); “costs of failure” (§3.2 note 9, which also now includes the GMO reference on p. 188); “no longer … a viable industry” (§4.5); “has not been inconvenient” (§4.7, Appendix A LL1-12); “the jury is still out” (§3.4, §5.1, §5.2, §5.4, §6.2, Appendix A); “solely responsible, at national level, for all” replaces “sole cause, everywhere” (§4.2, K2); “not enough was known” (§4.2)
10 Minamata mother did not precede “the experts” Yes (LL2-05, p. 105: Kitamura 1959) Fixed in §4.2 and Appendix A (LL2-05). Digest LL2-05 l. 26 not changed (outside scope)
11 Regulator’s “competitiveness” definitional clause used as evidence of intent to change rules Yes (LL2-22, p. 540: text of the Commission Recommendation) Fixed: removed from the “Strong on intent” evidence in §4.3 and I4; retained in §4.3 as a cost qualifier in a definition, rated suggestive. T03 l. 248 not changed (outside scope)
12 Sugar-beet yield loss attributed to the ban without its stated cause Yes (hindsight LL2-16: 2019→2020, virus yellows) Fixed in §4.4, C7 and Appendix A (LL2-16)
13 Cyclamate counted as a false positive Yes (LL2-02 classes it “the jury is still out”) Fixed in §5.2
14 NNI share compares two budget categories Yes (hindsight LL2-22) Fixed in §3.6 and §4.4, with the non-comparability caveat and the National Academies’ contrary assessment
15 Table 27.2 presented as the EEA’s rule for acting on weak evidence Yes (table titled “some examples and illustrations”; definition says “appropriate strength”, p. 649) Fixed in §4.1, §5.2 and the new §3.4 tools table
16 Knowledge-to-ignorance ratio judged as a harm predictor Yes (LL2-27, p. 655) Fixed in §4.1
17 Hammitt et al. called “confirmed” Yes (modest post-1990 European shift; diverse trends) Fixed in §3.1
18 Hospital dose variation cited to LL1-16 with LL1-03’s figure Yes (LL1-16, p. 175: “a factor of 100”) Fixed in §3.2 note 5 and K9
19 CFC-12 figures labelled as all CFCs Yes (LL1-07, p. 82) Fixed in §4.7
20 Shizuoka precedent described as “similar evidence” Yes (LL2-05, p. 98) Fixed in §4.2 and Appendix A (LL2-05)
21 PCB lag rounded to “over 100” Yes (Table A2.1: c. 100) Fixed in §4.2
22 Proportionality attributed to Chapter 1’s own view Yes (LL1-00, p. 13: element of the German Vorsorgeprinzip) Fixed in §3.7 item 5
23 Chapter 17 does not drop “illustrate”; it upgrades it Yes (LL1-17, p. 193) Fixed in §3.3 and §5.1 item 4
24 Michaels’s role ambiguous Yes (LL2-06, fn 1, p. 131) Fixed in §2.1
25 EPRS “4 of 88” citation points to the wrong file Yes (critiques §5.3; hindsight LL2-00) Fixed in §2.6

Balance review (balance.md)#

# Issue Verified? Outcome
1 Lens built to find missed harm, not misplaced alarm Yes (T08 and T09 asked for question-by-question symmetry; synthesis compressed it away) Fixed: §6 intro now names warners, campaigners and restricters; every one of the 72 entries has a Mirror line; I3 asks about funding of harm-side studies; I6’s Mirror asks about warners’ litigation, funding and institutional stakes; new W7 Warning quality and W8 The alarm trap (the reviewer’s “W3b”); the symmetry checks moved to the head of §6 as rule 0 and made self-contained
2 Strength ratings measure documentation in a failure sample; C7 and S4 under-rated Yes Fixed: C7 re-rated (strong on existence of material costs; moderate on size); S4 re-rated (strong existence; moderate predictability) with the after-the-fact note moved to Limits and applied as a general point; §1.4 now says what “Strong” does not mean; §5.8 glosses “High”. Partly: diagnosticity notes added where the files show them (K1, K4, W1, I2, M2) rather than to every entry, because the files do not document diagnosticity for most patterns
3 Project lead’s chapter and authors’ standpoints treated unevenly Yes (digest, notes and hindsight LL2-22; digests LL2-03, LL2-10, LL2-13, LL2-26) Fixed: a Standpoint clause added to all 47 Appendix A entries; LL2-22’s clause gives self-citation, Poland et al. 2008, Maynard et al. 2006, no dissenting panel, and the “preliminary” to “rapidly increasing evidence of risks” escalation; LL2-22 added to the §5.6 protagonist bullet (noting its vindicated MWCNT warning); §3.6 and §4.4 funding figures rewritten with caveats and the National Academies’ view; §1.5 rewritten. Unresolved: fix 4 (re-check of LL2-22 entries by a reader with no connection to the chapter) cannot be done by this reviser; recommended in §1.5
4 “Failures cluster where conviction and self-citation were highest” is selective Yes (digests show protagonist authorship in chapters that held) Fixed in §5.5 item 4 (reworded, rated moderate) and the §4.8 “reports’ own mindsets” row
5 Lessons from prevention failures rated for use under uncertainty Yes (T09 §12.4) Fixed: new §5.1 item 6; case-type tags [K]/[U]/[F] defined in §6.2 with a case list; every entry’s Strength line now gives support by case type; rule 9 and §5.8 weight by [U]/[F]. Tagged at entry level rather than per evidence item, to keep entries readable
6 Interests behind restriction and alarm, including protectionism, missing Yes (LL1-14; critiques §3.3, §9.4; LL2-03, p. 60; LL2-20, p. 499; hindsight LL1-07) Fixed: new §4.3 row; new lens entry I9; I7 Limits; §5.7 item 11
7 Fukushima 2,351 deaths attributed wholly to evacuation and used as flagship cost of precaution Yes (hindsight LL2-18 l. 147–148) Fixed in §4.1 table, §4.5, §5.2 (rebuttal now led by swine flu), T4, C7 Limits and Appendix A (LL2-18)
8 Findings favouring the false-alarm review under-credited Yes (T09 §5 rating; hindsight LL2-28 verdict) Fixed: (a) the moderate–strong finding restored in §5.2; (b) rarity rating reworded as “unmeasured”, with hindsight’s verdict; (c) “do not reproduce” replaced; (d) T09 §12.3 added as §5.1 item 7
9 External-critiques table asymmetric in three directions Yes (critiques l. 70, 81, 150–158) Fixed: paralysis row now “met in wording, not practice” with residual force; access column added and Cox’s rating attributed to this document’s own reading; a standpoints paragraph states plainly that critics’ standpoints were not examined to the same standard
10 Smaller over-corrections (four) Yes (T08 l. 30; hindsight LL2-26; notes LL2-27 l. 185, 250; hindsight LL2-00 l. 150) Fixed in §4.8, §4.1, §5.6 and §5.4 (LL2-00)
11 Post-2013 policy change described in one side’s vocabulary Yes (hindsight LL2-27 l. 143–151; hindsight LL1-03) Fixed: §3.6 analysis now neutral, lists rival explanations and evidence-led relaxations, and adopts T09 §12.6; §3.6 table open-access cell; §4.6 counter-evidence now gives defenders’ and critics’ readings of EO 14303; Appendix A LL1-03 now quotes Lambert’s “substantial lobbies” and reports the ICRP 2025 concession
12 C1 “expect delay regardless of evidence”; §5.5 item 2 one-directional Yes Fixed in C1 and §5.5 item 2 (understatements added; “errors ran both ways” labelled analysis)

Completeness review (completeness.md)#

# Issue Verified? Outcome
1 Lesson 5 has no lens entry Yes (quotations checked, e.g. “routine prophylaxis”, p. 86 not pp. 84, 88; “regardless of the presence”, p. 384) Fixed: new K9; rule 5 now lists all six knowledge states
2 No positive repertoire Yes Fixed: new §6.12 response repertoire (16 instruments); graduated-response check added to rule 0
3 Susceptible groups and windows missing Yes Fixed: new K10. Two suggested citations (LL1-12, pp. 127, 131; LL2-10, p. 224) omitted because not verified
4 Harm expansion and “first harm not decisive” missing Yes (LL1-16, p. 171 added as stronger support) Fixed: new K11
5 Cultural layer thin Yes (quotations checked) Fixed: new M5–M8; layer tags on every entry; T08’s “cannot be separated from interest” kept as M7 Limits. “People at risk choosing not to know” (LL2-15, p. 354) omitted: not verified in the source
6 Political-economic layer one-sided; agenda power absent Yes Fixed: I7 Limits; new I9 (merged with balance issue 6) and I10
7 Law and courts missing Yes (Ethyl appeal ruling, LL2-03, p. 60; Conseil d’État, LL2-16, pp. 374, 382; “academy of science”, LL2-17, p. 420) Fixed: new G8
8 Systemic gaps: shared resources, coupled systems, climate Yes Fixed: new S6, S7; G2 extended (formal compliance against real reduction); C5 extended (system inertia)
9 “Limits” missing from 34 entries Yes Fixed: all 72 entries now have Limits (checked by script)
10 Evidence from elsewhere Yes (LL1-16, p. 180 checked: Norway 1976, UK 1985) Fixed: new W9 (renumbered from the reviewer’s W7 to avoid a clash with balance’s W7); K7 Limits
11 Co-drivers and reversibility of reforms Yes Fixed: L6 extended; new G9
12 Hard to apply: no triage, staging, overlaps, crosswalk, template Yes Fixed: layer and stage tags; §6.2 first pass (reviewer’s ten plus W7 and C7, so the first pass is two-sided); overlaps cross-referenced; rule 10 recording template; §6.13 crosswalk of all 149 theme questions and the twelve lessons, with the previously dropped questions mapped; rule-0 checks now self-contained
13 No comparator rule Yes Fixed: rule 7
14 Synthesis tools named but not listed Yes (notes LL2-27, LL2-28, LL2-02, LL1-16; LL2-26 Table 26.5) Fixed: new table in §3.4 (seven barriers, Box 27.4 with critique, Table 27.2 bands, Table 16.1, three opportunities, six findings, seven false-alarm lessons, PATIO); K7 cross-references Box 27.4
15 Panels and internal dissent under-reported Yes Fixed: new §2.7 panels table; Bayer panel and its partial vindication in Appendix A (LL2-16); hindsight LL2-05 lesson 8 cited in G4
16 Some critiques and critiques §9.4 not carried through Yes Fixed: Durodié, Löfstedt, the European Risk Forum irreversibility argument, selective precaution and the philosophers’ defences added to §5.3; rule 8 applies §9.4’s four countervailing questions to every entry
17 Smaller mechanisms to absorb Yes, except as noted Fixed (partly): added to G2, K6, K2, C4, I5, W3, T4 and new C8. Omitted: LL1-06, p. 66 and LL1-12, p. 128 (not verified); T02 §1.1 softened to what it shows (animal data as early warnings), since it does not show they were “discounted as not relevant”
18 Minor points (labels, §4.9/4.10, §2.6, LL2-18, §1.5) Yes Fixed: §1.4 maps qualifier labels onto the scale; §4.9 names T09’s counterweights and §4.10 T10’s P1–P6; §2.6 notes SOER 2020 and Drivers of change are EEA restatements; Appendix A LL2-18 adds the scenario-list and security-remit points; §1.5 lists where LL2-22 carries weight

Unresolved or outside scope#