Late Lessons, Jensen Huang and AI
Markdown version

Review: completeness and breadth of 01-late-lessons-analysis.md#

Strand A critical review, 26 September 2026. Scope: completeness and breadth only. Accuracy of individual quotations and numbers is not re-audited here.

What was compared. The analysis document (1,110 lines) against all ten theme files (T01–T10, including each file’s closing list of fifteen diagnostic questions), the “transferable insights” sections of all 47 digests, the “technology-neutral lessons” and “implications” sections of the hindsight files (read in full for LL2-05, LL2-06, LL2-12, LL2-13, LL2-18 and LL1-17; searched for the rest), targeted passages of the notes (LL2-16, LL2-27, LL2-28), and both external files.

Bottom line. Sections 2–5 are a faithful and broad condensation of the working files; most theme patterns appear somewhere in section 4. The weakness is section 6. It carries forward about a third of the 150 theme questions, is weighted heavily towards the epistemic layer, is almost entirely a catalogue of failure modes, leaves out several of the best-supported patterns in the reports, and leaves “Limits” off 34 of its 55 entries. It is usable as a checklist by a reader who already knows the corpus. It is not yet usable as the stand-alone lens the document says it is (lines 18, 643).

Section 6 by layer (entries counted against the six layers the document promises in §1.1, line 16):

Layer promised in §1.1 Lens entries Count
Epistemic K1–K8, W1–W6, T1–T4 18
Political-economic I1–I8 8
Economic (trajectories and costs) L1–L6, C1–C7 13
Institutional G1–G7 7
Systemic S1–S5 5
Cultural or cognitive M1–M4 4

Section 4 labels each theme by layer (lines 299–459); section 6 drops the labels and reorganises by topic, so the imbalance is not visible to a user.

Issues are ranked by importance: first the missing substance that most weakens the lens, then balance and usability, then the faithful-record gaps in sections 2–5.

1. Lesson 5 (“real-world conditions”), the best-supported lesson, has no lens entry#

Location. §6 (lines 641–976). Compare §3.2, row 5 and note 5 (lines 206, 221).

Problem. Section 3.2 rates lesson 5 strong and calls it “the widest case support of any lesson” (~10 cases). In the lens it survives only as a vocabulary check inside G1 (“controlled use”, “closed systems”, line 884) and as one phrase in M2 (“assumed performance to specification”, line 953). No entry asks how the technology will actually be installed, maintained, used, misused, transformed or enforced, or who other than the operator would detect a departure. Related material in the working files is also missing: LL2’s “indeterminacy” state (unpredictable uses, answered by pre-market benefit assessment), use creep into routine or prophylactic uses with no demonstrated need, and transformation of the product in use.

Evidence. - Lesson 5: LL1-16, pp. 174–175 (tanks leaked, “closed systems” leaked and were dumped, controls unenforced, doses varied tenfold); LL1-05, p. 57 (WTO: “controlled use” cannot be relied on); LL1-15, pp. 160–162 (offal controls designed around commercial convenience; ~48% of abattoirs visited in 1995 failing, p. 162); LL1-11, p. 115 (containment depends on enforcement; digest LL1-11 insight 12). - Transformation in use: LL1-06, p. 67 (the tested product differs from the transformed exposure; digest LL1-06 insight 7). - Indeterminacy and use creep: LL2-27, Table 27.1, p. 656; LL1-03, p. 34 (shoe-shop fluoroscopes under recommendation-only rules); LL1-08, pp. 84, 88 (DES promoted for “routine prophylaxis in all pregnancies”); LL2-16, p. 384 (seed dressings used “regardless of the presence and abundance of pests”); LL2-22, pp. 544–546 (controlled-use assumptions diverge from practice; “solutions looking for a problem”). - Exposure routes the regulated metric ignores: LL2-09, pp. 204–206 (air monitoring missed skin uptake; digest LL2-09 insight 4); hindsight LL2-06, lesson 2 (full-shift average did not control peaks or skin contact). - Misuse and detection by third parties: LL2-A3, p. 737; hindsight LL1-17 (illegal CFC-11 found by atmospheric monitoring); T10 question 6.

Fix. Add an entry to §6.2 or §6.8, for example:

K9. Designed conditions against real use. Appraisals assume containment, maintenance, compliance, intended use and the product as sold. In practice systems leak and degrade, rules go unenforced, products transform, and uses spread beyond those where any benefit was shown. Ask: What does the appraisal assume about containment, maintenance, dose, compliance and purpose? Who, other than the operator, would detect leakage, misuse, unappraised uses or non-compliance? What does the product become in use and at end of life? Are uses spreading into routine, prophylactic or trivial applications with no demonstrated benefit? Evidence: as above. Strength: Strong (~10 cases). Limits: some rules worked quickly once enforced (all-species feed ban; global TBT ban; T06 §8). The reports’ “real practice” claims for emerging technologies are sometimes asserted rather than documented (digest LL2-22, “suggestive”).

Also add “indeterminacy” and “variability” to rule 5 in §6.1 (line 651), which lists only four knowledge states although §2.5 (line 160) records LL2’s six.

2. The lens has no positive repertoire: it lists how things fail, not what worked#

Location. §6 as a whole. Only W5 (line 727) and the “conditions of success” clause in G5 (line 904) describe success. §6.11 has no counterpart question.

Problem. Almost all 55 entries are failure diagnostics. The reports and the hindsight files also document response designs that worked, partly worked, or failed in instructive ways. These are scattered through §4 (critical loads, the Montreal ratchet, nitrites, BSE de-escalation, producer-pays). They never reach the lens, so a user can identify a failure mode but gets no evidence-graded menu of responses. Rule 4 (line 650) separates prevention from precaution but gives no tools for either. The omission also tilts the lens against the reports’ own conclusion that precaution is a framework for broadening responses, not a binary ban (critiques §8; T09 question 15).

Evidence (instrument; where it worked; its limits). - Graduated, exposure-reducing measures: nitrites (LL2-02, p. 25, vindicated in hindsight LL2-02); DDT “centrist” exposure reduction (LL2-11, p. 248); LL2-02 lessons, pp. 34–35. Moderate for nitrites, suggestive otherwise (digest LL2-02, insight 8). - Provisional action plus committed research: the “double reaction” (LL2-28, p. 673); the Swann procedure (LL1-16, pp. 173, 181); held workable in law (hindsight LL1-17, lesson 5). - Emergency or interim powers: DBCP standard in about two months (LL2-09, pp. 206–207). - Measurable intermediate thresholds: critical loads (LL1-10, pp. 106–107); exceedance down to 3.5% by 2024 (hindsight LL1-10). - Jointly produced fact base: EMEP and the joint UK–Scandinavian programme (LL1-10, pp. 103–107; digest LL1-10, insights 7 and 11). Necessary, not sufficient. - Review ratchet plus transition finance: Montreal (LL1-07, pp. 78–81); later adaptations (hindsight LL1-07). - Pre-agreed triggers: LL2-17, p. 423; LL2-12, p. 274. Asserted in the reports; hindsight shows triggers get re-specified downwards (hindsight LL2-17). - Supply choke-point controls: LL2-12, p. 273; hindsight LL2-12, lesson 5. - Class- or function-based restriction: LL1-13, pp. 141–142; hindsight LL2-10 (2024 EU bisphenols rule), hindsight LL2-13, lesson 2. - Prior justification plus optimisation (radiation protection): LL1-03, pp. 34–35; rare and only partly effective (LL1-16, p. 176). - Independent outside re-analysis: LL2-17, pp. 412–413; LL1-02, p. 21 (digest LL2-17, insight 4). - Open, costed review for de-escalation: hindsight LL1-15 (the Over Thirty Months rule replaced after a review put its cost at about £2bn per death prevented). - Producer-pays at source: hindsight LL2-13, lesson 3 (which also shows the contest moving to cost attribution). - Surveillance built alongside restriction: hindsight LL1-09 (DANMAP and Svarm produced the evidence that later judged the bans). - Acting while the window is open: LL2-20, p. 498 (Caulerpa in California after 17 days). - Small jurisdictions funding their own evidence: LL2-12, pp. 271–272 (Bermuda).

Fix. Add §6.12 “Response repertoire”: a table with columns instrument; what it does; cases where it worked; cases where it failed or was diluted; strength, built from the list above. Add one symmetry check to §6.11: “Has the full range of graduated, provisional and reversible responses been considered, or only allow-or-ban?”

3. Susceptible groups, life stages and critical windows are missing#

Location. §6.2. K2’s “Ask” mentions “averages” in passing (line 663). No entry is devoted to susceptibility.

Problem. One of the most consistent patterns in both volumes is that reference subjects, average exposures and adult or acute endpoints hide the most sensitive groups and life stages (foetus, neonates, children, low-baseline subgroups), and that timing of exposure can matter as much as dose. Hindsight rates LL2’s persistence-and-vulnerable-window warning its best-vindicated introductory claim (Appendix A records this at line 1022, but the lens does not use it). T07 question 6 and T01 question 7 ask exactly this; neither survives into §6.

Evidence. LL1-14, pp. 150, 152–153 (low-baseline prepubertal children missed; hindsight LL1-14 confirms children’s levels were overestimated); LL2-26, Table 26.3, p. 630 (adult male as reference subject) and pp. 638–639 (averages hide concentrated harm); LL2-27, p. 650 (“more the timing of the dose, rather than the dose itself”); LL2-10, pp. 219, 224 (“the time makes the poison”; neonates least visible); LL2-05, pp. 105–107 (placenta assumed protective; congenital disease recognised first by a mother); LL1-08, pp. 84–88; LL1-12, pp. 127, 131; LL2-11, pp. 247–248 (evidence thinnest where exposure highest); LL2-28, p. 674 (variable susceptibility; “tipping point exposures” differ); LL2-00, pp. 7–8; hindsight LL2-00; hindsight LL2-26 (child vaccine-response data behind PFAS limits); digest LL1-03, insight 9 (risk estimates drawn from one atypical high-dose cohort).

Fix. Add:

K10. Who is most sensitive, and when? Ask: Which groups, life stages or parts of the system are most sensitive, and are they represented in tests, reference subjects and averages? Is the timing of exposure considered as well as its size? Where is exposure highest and evidence thinnest? Strength: Strong (reference subjects and averages hiding sensitive subgroups; developmental windows, strengthened in hindsight). Limits: non-monotonic dose–response at environmental doses did not hold up (hindsight LL2-10); the evidence is densest for endocrine and neurodevelopmental agents.

4. “Harm expansion” and “the first harm is not the decisive one” have no lens entry#

Location. §6. Harm expansion appears only as a caveat in K3’s “Limits” (line 672). It is described in §3.4 (line 252), §4.7 (line 447) and §5.5 (line 593).

Problem. LL2’s editors present harm expansion as one of their main 2013 findings, and hindsight strengthened it for every named agent. The related mechanism, that controlling the first and most visible harm breeds false confidence about slower or different harms, is rated strong in T02 §1.4 and in several digests. The lens mentions it only as a reason to discount apparent expansion. It never asks the forward question: what else might this agent or failure mode do, at lower doses or to other endpoints? The “moving target” defence (“today’s technology is now safe”, LL2-28, p. 672), with its long asbestos history of “problem already solved”, is also missing.

Evidence. LL2-28, p. 672; LL2-A3, p. 717; LL2-A2, pp. 704–714 (digest LL2-A2, insight 3: “strong”); LL1-16, pp. 171, 173–174; LL1-03, p. 33 (1925 tolerance dose about 700 mSv a year against 20 mSv now); LL2-06, pp. 133–134 (acute beryllium disease controlled, chronic disease below the limit); LL1-05, pp. 54, 56 (asbestosis-only rules); LL2-24, pp. 590–593 (liability for a class of harm when on notice of a lesser one; digest LL2-24, insight 3); hindsight LL2-28 (limits cut after 2013 for lead, asbestos, PM2.5, PFAS and BPA). Counter-evidence: selection and detection effects (hindsight LL2-A3); counterexamples in nickel (2020) and perchlorate (2026) (hindsight LL2-26); DES daughters’ breast cancer now contested (hindsight LL1-08).

Fix. Add:

K11. The first harm is rarely the last. Ask: Is oversight calibrated to the first, most visible harm? Has controlling it created a sense that the hazard is handled? What other endpoints, lower exposures or populations could the same agent or failure mode affect? Are claims that observed harms belong to superseded versions being tested? Strength: Strong for confirmed hazards; moderate as a prior for suspected ones. Limits: as above.

5. The cultural and cognitive layer is thin (4 entries), and §4.8’s patterns were dropped#

Location. §6.10 (lines 945–966) against §4.8 (lines 459–475) and T08 (Patterns A–K).

Problem. §4.8 lists nine patterns. The lens keeps four. Dropped or reduced to a phrase: - enthusiasm, hype and the premium on modernity; - who counts as an expert, and borrowed credibility; - seeing the public as prone to panic (reduced to a word list in M4); - the reports’ own mindsets (moved to §6.11).

Several cultural mechanisms in the digests and hindsight files appear nowhere in the document: - the ideology that profit-seeking itself serves society; - cultures of denial built by “good people”, and “ethical blindness”; - professional and disciplinary cultures that fix what counts as harm; - national-identity and competitiveness narratives, and strategic designation; - the “safety myth” as an institutional culture; - media practice that sustains apparent controversy; - focusing events filtered through national political culture; - public campaigns and elections as triggers; - lay criteria for judging hazards that differ from experts’ criteria; - people at risk choosing not to know.

For a lens that claims to span “cultural or cognitive” layers (line 16), four entries, two of them about individual cognition, are too narrow.

Evidence. - Enthusiasm and modernity: LL1-03, p. 31; LL1-08, p. 88; LL2-03, p. 53; LL2-13, p. 280; LL2-22, pp. 545–546 (“nano-fever”; “should” versus “could”); LL1-16, p. 176. - Ideology and cultures of denial: LL2-25, pp. 613–616; LL2-28, p. 678 (“self-serving bias”, “ethical blindness”). - Expertise and borrowed credibility: LL2-06, pp. 135–136; LL1-15, pp. 160–165 (curated advisers; value-laden advice presented “as if it was purely scientific”, p. 165); LL2-05, p. 103 (“Centre” against “hick” universities); LL1-09, p. 98; LL2-16, pp. 379, 386–389 (T08 Pattern E). - Professional culture: LL1-16, p. 174 (clinicians’ acute focus); LL2-22, pp. 543, 545, 550 (an agency “rooted in chemistry”; developers’ mental models resisting safety). - National and strategic framing: LL2-03, p. 53 (“survive among the nations”); LL2-05, pp. 96, 99 (company town; “Never stop it!”); hindsight LL2-06, lesson 9 (strategic designation turns policy from reducing use to securing supply); hindsight LL2-18 (“safety myth”). - Media and focusing events: LL2-07, p. 166 (digest LL2-07, insight 10); LL1-05, pp. 56–57; LL1-10, pp. 102–103, 104–105; LL2-03, Haigh panel, p. 63 (UK unleaded accepted “within half an hour” before an election); LL2-18, pp. 442–445 and hindsight LL2-18, lesson 8. - Lay criteria and chosen ignorance: LL2-20, p. 500; LL2-15, p. 354.

Fix. Expand §6.10 to at least seven entries: keep M1–M4; add M5 Enthusiasm and the premium on novelty or modernity; M6 Who counts as an expert (composition, borrowed credibility, disciplinary culture, T08 question 6); M7 Organisational and national cultures (denial cultures, ideology, strategic designation, “safety myth”); M8 Salience (media balance, focusing events, campaigns, elections). Tag each with its layer, and keep §4.8’s “cannot be separated from interest” caveat (T08 §14, item 7) as a “Limits” line.

6. The political-economic layer examines only one side: capture of precaution, selective precaution and agenda power are absent#

Location. §6.5 (I1–I8) and I7 (line 799).

Problem. The I-entries examine developers’ and producers’ interests. They never ask whose interests a restriction serves. The working files document precaution captured for protectionism or competitive advantage, rival firms positioning around substitutes, and the “selective precaution” finding that the comparative literature says is the phenomenon to explain. I7 presents countervailing interests only as an enabler of action (“action often waited … for an organised interest”), with no limit noting that the same interests can capture precaution. Also missing is the reports’ own power diagnosis: decisions “made by a few people on behalf of many”, the “democratic deficit”, and control over how problems and “innovation” are defined. §5.7 lists “no analysis of power” as a limit of the reports (line 623), and T10 P1 says the lens should give these mechanisms more weight than the reports did. The lens does the opposite.

Evidence. - Protectionism and rival interests: LL1-14, pp. 150, 153–154 and hindsight LL1-14 (dispute settled by beef quotas at third-country exporters’ expense); critiques §3.3 (Majone: misuse for protectionism, double standards, regressive effects) and §9.4 (item 2: ask whether a response “can be captured to protect incumbents or markets”); LL2-20, pp. 498–499 (firms favouring legislation so competitors cannot undercut them); hindsight LL1-07 (DuPont’s shift partly commercial positioning); LL2-03, p. 60; LL1-07, p. 80 (first-mover penalty); T03 P10. - Selective precaution: critiques §3.4 and §9.4 (item 5); LL1-16, p. 168; Hammitt et al. 2005 (already cited at line 190). - Agenda power: LL2-28, pp. 671, 678; LL2-05, pp. 92, 114; LL2-03, pp. 52–53 (a seven-hour conference at which “No ‘innovation’ other than TEL was discussed”); LL2-19, p. 461 and digest LL2-19, insight 3 (defining “innovation” is a distributive choice); LL2-17, p. 419 (knowledge institutions survive their failures; costs fall on the least powerful). - Economic centrality: LL2-05, pp. 96, 99 (digest LL2-05, insight 6).

Fix. Add a “Limits” line to I7 on interest-aligned precaution, and two entries:

I9. Whose interests does the restriction serve? Ask: Who gains commercially or politically from a proposed restriction, standard or substitute? Is it applied to this risk but not to comparable ones, and why? Would it survive the same scrutiny applied to the technology? Strength: Moderate (hormones, horticulture, substitute positioning; critics’ strongest distributive point).

I10. Who decides, and who frames the problem? Ask: How many people take the pathway decision, and who is absent? Who defines the problem and what counts as “innovation” or “safe”? How economically central is the activity to the jurisdiction deciding on it? Strength: Moderate (vivid cases, no comparison set; T04 P6f).

Location. §6.8 (G1–G7). “Court” appears nowhere in §6. Litigation appears only as I1’s observability caveat and T1’s Pfizer citation.

Problem. T06 §11 rates “courts cut both ways; the legal standard decides the outcome” strong (15 or more episodes), and §4.6 includes it (line 429). The lens never asks which legal standard will govern: a precautionary statute with reasoned discretion, a quantified “significant risk” test, individual foreseeability, or sole-cause causation. Nor does it ask how much deference reviewers will give, or whether liability can attach to a class of harm. Hindsight adds that legal accountability runs on a narrower foreseeability test than official inquiries use.

Evidence. LL1-04, p. 40 (1980 Benzene); LL2-08, p. 187, Box 8.1 (upheld “on the frontiers of scientific knowledge”); LL2-03, p. 60 (Ethyl v EPA); LL1-05, p. 57 (WTO on asbestos); LL1-14, p. 153 (WTO on hormones); LL2-16, pp. 378–380 (Conseil d’État; Bayer’s defamation suits lost); LL2-17, p. 420 (Nunavut: a court would not “become an academy of science”); LL2-24, pp. 586–593; LL2-21, p. 525; LL2-A3, pp. 736–737 (burden, deference, independence); hindsight LL1-09 (Pfizer), LL2-16 (PAN Europe), LL2-18, lesson 9 (foreseeability narrower than inquiries; acquittals final 2025), LL2-24 (Milward; FRE 702 tightened 2023).

Fix. Add G8. The legal standard decides. Ask: What standard of proof and causation will courts and trade tribunals apply? How much deference do they give to a reasoned regulator? Can responsibility attach to a class of harm when a lesser harm was known? Is litigation the only route by which internal knowledge will surface? Strength: strong (two-way role); moderate (deterrence). Limits: courts’ precaution is narrower than the reports’ (Pfizer: risk must be “adequately backed up”, not “purely hypothetical”; T06 §13, item 7).

8. Systemic layer gaps: shared resources, tightly coupled failure, and a climate case that is almost unused#

Location. §6.9 (S1–S5). LL2-14, a 38-page systems case, is cited once in the whole lens (G5, line 905).

Problem. Three strands of systemic evidence are missing: - Shared, depletable resources. Antibiotic efficacy, groundwater, fish stocks and public assets are cases where each local use is a system-wide cost, and “loss of use” is harm even without toxicity. T07 question 8 asks this; L5 covers only resistance treadmills. - Tightly coupled systems and extreme events. LL2-18 and LL2-15 are the corpus’s only cases of acute catastrophic failure, as opposed to chronic diffuse harm. They show safety cases built on scenario lists and independence assumptions; published estimates of extremes not reaching design bases; confidence built on “no accident yet”; monitoring failing in the extreme it exists to observe; design codes built on short records; and emergency responses causing harm themselves. Only fragments reach K2’s evidence and S4. - The climate case’s institutional and systemic lessons. Framework action comes fast and effective action slowly; formal compliance diverges from real reduction (offsets, windfalls); system inertia locks in harm before it is observed; precaution is easiest when those who pay also benefit; authoritative assessment bodies are single points of failure; structural uncertainty is under-estimated.

Evidence. - Shared resources: LL1-09, pp. 94, 96–97 (digest LL1-09, insight 9); LL2-A3, pp. 731–732 (digest LL2-A3, insight 9); LL1-11, pp. 112, 114, 119 (digest LL1-11, insight 5); LL1-02, pp. 19–20, 22, 26; LL2-24, p. 602. - Coupled systems: LL2-18, pp. 432, 438–439, 444–448 (digest LL2-18, insights 1, 3, 4, 12); LL2-15, pp. 351–355, 357–360; hindsight LL2-18, lessons 1, 3, 10; hindsight LL2-15 (Ahr 2021; Valencia 2024); T07 §3.14. - Climate: LL2-14, pp. 309, 311, 314, 321–328, 330, 332–334, 337 (digest LL2-14, insights 1–12); hindsight LL2-14.

Fix. Add S6. Shared resources and loss of use, and S7. Tightly coupled systems and extremes (Ask: What cascades or common-cause failures lie outside the scenario list? Does the design basis use short records, and have published estimates of rare extremes reached it? Will monitoring and warning survive the event? What would the emergency response itself cost? Strength: moderate–strong, resting on official inquiries; the chapter’s own probability arithmetic is unreliable). Extend G2 with “formal compliance against real reduction” (LL2-14, pp. 324–326), and C5 with “system inertia” (LL2-14, pp. 314, 337).

9. “Limits” is missing from 34 of 55 entries, including every C, G and S entry#

Location. §6. The introduction (line 643) promises that every entry gives “its known limits”.

Problem. Entries without a “Limits” line: W2, W4, W5, T2, T4, I6, I7, I8, L2, L5, L6, K8, C1–C7, G1–G7, S1–S5, M1, M3, M4. The gap falls on the institutional, economic and systemic layers, where the working files hold explicit counter-evidence. That makes these entries read as stronger than the evidence allows, which breaks rule 2 (symmetry).

Evidence (counter-evidence available for each). - C1: fast action despite concentrated costs (vinyl chloride, DBCP, DES; T05 §3.1). - C2: estimates can also be too high (MTBE’s upper-end estimate, the Dutch asbestos counterfactual; T05 §3.2). - C3: where the protected and the exposed are the same people (LL2-11, pp. 246–249). - C4: two courts accepted the state’s criteria (hindsight LL2-05). - C5: Deepwater Horizon is a weak test of the cap argument, and compensation tables would have paid for a probably non-existent harm (hindsight LL2-24, Claims 7 and 9). - C6: producer-pays triggered litigation over cost attribution (hindsight LL2-13, lesson 3). - G1: the reports relabel too (“precautionary prevention”, LL1-00, pp. 12–15). - G2: some rules worked fast once enforced (BSE cohort cases 1,248 to 111; TBT; SO2; T06 §8). - G3: replacing an anchor creates the next anchor (hindsight LL2-06, lesson 1). - G4: harmonised measurement narrows factual disputes but not normative ones (hindsight LL2-05, lesson 8); CLARITY-BPA (hindsight LL2-10). - G5: small jurisdictions sometimes lead (LL2-12, pp. 271–272). - G6: value-driven outcomes can rest on factual error (Brent Spar; hindsight LL1-17, lesson 7). - G7: the “homo-illogical cycle” is extrapolated from floods (T06 §10). - S1: stocks can become resources, and some recovery was faster than feared (halon bank; Great Lakes delistings; MTBE attenuation; T05 §3.13). - S2: much improvement came from structural change, not governance (T07 §5, item 4). - I6: the exit-route thesis is equally explained by interest alignment (hindsight LL2-06, lesson 6). - L2: benefits were real and large in several cases (DDT; sugar beet; PCB fire safety; T04 P2). - M3: organisations reversed where they had less sunk commitment (§4.8, line 470).

Fix. Add a “Limits” line to every entry, drawing on the cited counter-evidence and each theme’s “Counter-evidence” section.

10. Evidence from elsewhere: discounted when it warns, relied on when it reassures#

Location. §6.3. W4 mentions “harm was elsewhere” as a cost-distribution point (line 722). There is no entry on cross-context evidence.

Problem. Several cases show a warning discounted because the harm had appeared in another place, population or country. Minamata’s cause was identified, yet the poisoning recurred at Niigata. Other cases show the mirror error: “no harm elsewhere” used as reassurance where conditions differed. Hindsight adds that countries ban asbestos once mesothelioma appears in their own data. That finding is cited in §3.1 (line 192) but never turned into a lens question. For self-propagating agents, the best predictor was track record elsewhere, not intrinsic properties. That qualifies K7’s property screening. T02 question 14 asks this, and it was dropped.

Evidence. LL1-12, p. 126 (Florida eagle declines published 1952–58; Great Lakes research only after local loss); LL2-05, pp. 102, 105 (digest LL2-05, insight 10); LL1-16, p. 180 (UK took nine years to accept Scandinavian acid-rain science); LL1-10, p. 105 (outsiders’ evidence counted once the sceptic’s own institutions confirmed it; digest LL1-10, insight 11); LL1-05, p. 54 (Germany 1938/1943 against the UK); LL2-20, pp. 490, 500–501 (invasiveness elsewhere is the best predictor, degrading as conditions change); LL2-27, Box 27.4, criterion 8 (analogy), p. 653; hindsight LL1-00 (odds of a ban 14.1 times higher).

Fix. Add W7. Evidence from elsewhere (Ask: Is harm seen in another place, population or setting being discounted because it has not appeared locally? Is “no harm elsewhere” being relied on where conditions differ? What analogous track record exists, and who would have to confirm it for it to count here? Strength: moderate). Add a “Limits” line to K7: property proxies are chemical-specific; for adaptive or self-propagating agents, track record elsewhere predicted better.

11. Co-drivers, obsolescence and the reversibility of protective reforms are not lens questions#

Location. §6.6 (L6, line 839) and §6.8 (G7, line 913). The pattern is described in §5.5 item 7 (line 599) and §4.4 (line 389).

Problem. Many exits came through unrelated co-drivers: catalytic converters, “pure chance” alignment with forest concerns, production ending as “no longer necessary”, economic collapse, resistance. This makes those exits contingent, and it makes credit for improvement hard to assign. Separately, protective reforms proved politically reversible while incumbent capital persisted for decades. The document states this as a finding (§5.5, item 7) but gives the lens no question for it. G7 covers only fading vigilance. T04 question 15 and hindsight LL2-18 lessons 2 and 8 were dropped.

Evidence. LL2-03, pp. 60, 63–65 (digest LL2-03, insight 9: “final elimination often needs an independent co-driver, which makes it fragile”); LL2-05, p. 105 (digest LL2-05, insight 7: acknowledgement waited for obsolescence); LL1-10, pp. 104–106 and hindsight LL1-10 (structural change “did much of the work”); hindsight LL2-11 (DDT’s exit through resistance and new tools); LL2-21, fn 9, p. 515; LL1-12, pp. 128, 132 (commitments erode with political change); LL1-16, p. 180 (draft rules withdrawn in 1979 as “an unnecessary burden”); hindsight LL2-18, lessons 2 and 8; hindsight LL2-24 (French alert commission abolished 2026); hindsight LL2-13, lesson 1 (deferral instruments: delay and dilution are separable).

Fix. Add to L6: “What unrelated developments would change the trajectory anyway, and how durable would an exit be if they reversed? When harm falls, how much is due to the measure?” Add G9. Protective reforms are reversible; incumbent capital is not. Ask: How durable is the coalition behind a reform? Can it be deferred, derogated or re-specified? Is delay being tracked separately from dilution? Strength: moderate, strengthened in hindsight.

12. The lens is hard to apply: no triage, no staging, overlapping entries, no record of dropped questions#

Location. §6.1 and the structure of §6. Appendix B, line 1109.

Problem. A user faces 55 entries, 11 symmetry and usage rules, and no guidance on order or priority. - No lifecycle staging. The reports’ own structure is temporal: deployment, first signal, response, “effective action” as a process, then legacy (LL2-A2, p. 702; T10 P5). The lens is not organised or tagged by stage. It has no question about the governance window closing as commitment grows. Neither report names the Collingridge dilemma, but both analyse it (line 170). Also missing: “intervene at design, before lock-in” and “reform windows are scarce” (LL2-22, pp. 539–540, 547, 550–551; digest LL2-22). - Overlaps. Examples: K2, G4 and M2; W3 and M3; C1 and W4; L3 and S2; K4 and K7. These are not cross-referenced. - No layer tags (see the table above). - No traceability. Appendix B says §6 “consolidates” the 150 theme questions, but there is no crosswalk showing which were merged and which were dropped. This review found these dropped: T01 Q7; T02 Q5, Q14; T04 Q15; T05 Q15; T07 Q6, Q8; T08 Q6; T10 Q6. These survive only in part: T06 Q14; T07 Q12; T09 Q7, Q14, Q15. - No recording template. Rule 1 says a pattern is “a reason to look harder”, but nothing says how to record an application (present, absent or unknown; evidence; confidence) or how to avoid adding up hits into a verdict. - Not fully stand-alone. §6.11 relies on §4.3, §5.6 and unexplained “hindsight” references, although line 18 says §6 “can be used on its own”.

Fix. 1. Tag each entry by layer and by stage (pre-deployment / scaling / first signals / contested / after restriction / legacy). 2. Add a short “first pass” of about ten entries, chosen by strength and breadth of support (for example K1, K2, K9, K10, W2, T1, I1, I5, L3, G2). 3. Add a crosswalk table (entry → theme questions and the twelve lessons), listing dropped theme questions with a reason. 4. Merge or cross-reference the overlapping entries. 5. Add a one-line recording template. 6. Replace internal cross-references in §6.11 with the substance.

13. Add a comparator rule: which comparable actors or jurisdictions chose differently?#

Location. §6.1 (rules) and §6.11 (symmetry checks).

Problem. The reports’ most persuasive evidence often comes from natural comparisons. Shizuoka acted on similar evidence in 1950 while the national ministry refused in 1957. Sweden and Denmark acted on growth promoters before the EU. California eradicated Caulerpa where France did not. The US acted on DES before Europe. Bermuda acted on boosters before the EU. Firms within the same sector differed. T05 question 15 names within-sector and cross-jurisdiction variation as “the best check on claims that action is unaffordable or inaction inevitable”. The lens has no such rule. That leaves users with the selection-biased case record as their only reference class, against rule 1.

Evidence. LL2-05, pp. 98–99; LL1-09, pp. 95–96; LL2-20, p. 498; LL1-08, p. 89; LL2-12, pp. 271–273; hindsight LL2-25, Claim 1 (a rival refiner declined MTBE; Materion co-drafted a tighter limit); LL2-17, p. 414 (Norway, with the hindsight caveats in LL2-17); critiques §3.4.

Fix. Add rule 7 to §6.1: “Look for comparators. Who, facing similar evidence, acted differently, and what happened to them? Treat heterogeneity among firms and jurisdictions as evidence.”

14. The reports’ own synthesis tools are named but not listed#

Location. §3.4 (lines 248–252) and §2.5 (line 159).

Problem. For a “faithful record” (line 13), several central LL2 instruments are referred to only by name: - Seven barriers (LL2-27, p. 645): corporate opposition; definitional confusion; complex systems; the clash between evidence standards for causation and for policy; skewed cost analysis and unrealistic prices; short-termism; failure to engage the public. The notes record that the chapter groups barriers 1, 6 and 7 as “political and economic power”. - Box 27.4’s twelve criteria for action (p. 653): toxicity; novelty; persistence; bioaccumulation; spatial range; seriousness; irreversibility; analogy; inequitable distribution; feasible alternatives; potential to stimulate innovation; potential for future learning. This is directly usable in K7, yet only some items appear there. The notes’ critique, that the box mixes hazard evidence with policy considerations and gives no weights (notes LL2-27), belongs beside it. - The bands of Table 27.2 (p. 658) and the rungs of LL1’s Table 16.1 (p. 184). - LL2-28’s three “opportunities” (pp. 671–672) and six headed findings (pp. 672–680), of which §3.4 gives only parts of the first. - LL2-02’s seven lessons (pp. 34–35), of which five are given. - The PATIO elements (LL2-26, p. 638).

Fix. Enumerate each list in §3.4 (a compact table is enough), with the notes’ critique beside Box 27.4. Cross-reference Box 27.4 from K7.

15. Panels and internal dissent are under-reported, and the only industry panel’s partial vindication is missing#

Location. §5.1, item 2 (line 492); §3.5 (line 268); Appendix A entries for LL2-03, LL2-05, LL2-08, LL2-14, LL2-16, LL2-17.

Problem. The document notes that panels exist but rarely says what they argue or whether they complicate their chapter. Four are the main internal counterweights in LL2: - Bayer panel (LL2-16, pp. 401–402). It argued that colony losses are multifactorial and that large-scale monitoring found no correlation with seed dressings. Hindsight finds the multifactorial framing held up, as did the narrower claim that honeybee colonies often show no measurable harm in the field. The authors’ reply did not engage the monitoring studies (digest LL2-16; hindsight LL2-16, Claim 9). §4.1 and Appendix A report “honeybee a poor sentinel” without linking it to this dissent. - Castaño against Grandjean (LL2-05, pp. 122–124, 130). They read the same exposure data through different thresholds. Hindsight: harmonised measurement narrowed the factual dispute but not the normative one (hindsight LL2-05, lesson 8). - Panel 14.1 (LL2-14, pp. 332–335). It says structural uncertainty is systematically under-estimated, while the main text says uncertainty is resolved (digest LL2-14, “internal tension”). - Panel 17.2 (LL2-17, pp. 421–422). It treats scientific advice as a “fixed boundary condition”, while the main text critiques science’s entrenched authority (digest LL2-17).

Also absent: - the LL2-03 panels (Kovarik, Haigh, von Storch, Millstone), which “draw out” the contingency the chapter’s lessons omit (digest LL2-03); - Huff’s Panel 8.2 on animal bioassays, which presents contested claims without the counter-view (digest LL2-08; hindsight LL2-08, Claim 6); - Selin’s panel on framings outliving warnings (LL2-05, pp. 125–128); - Onasch’s thin substitution economics (LL2-04, p. 87); - the LL2-00 observation that four commentaries were left off the panel list (p. 16).

Fix. Add a short “Panels and internal dissent” table to §2.3 or §5.1: panel, author and standpoint, argument, whether it complicates the chapter, and hindsight. Add one line on the Bayer panel to Appendix A (LL2-16), and cite hindsight LL2-05, lesson 8, in G4.

16. Some external critiques and the critiques file’s lens recommendations are not carried through#

Location. §5.3 (lines 522–535); §6.

Problem. Several critics are missing from §5.3: - Durodié on fear, the limits of a “right to know” and the social costs of precaution (critiques §3.5); - Löfstedt on how alarms are handled and communicated (acrylamide), which the 88-case framework does not assess (§3.3); - the European Risk Forum’s argument that precaution is effectively irreversible because investment stops (§3.3). This bears directly on T3 and T4; - the philosophers’ defences (Sandin et al.; Steel) (§8).

More importantly, critiques §9.4 gives explicit instructions for using the reports as a lens. Pair every lesson with the critics’ countervailing mechanisms: substitute risks, forgone benefits for the least advantaged, capture to protect incumbents or markets, and reversibility. Treat selective precaution as the thing to explain. The lens implements only some of this, in C7, S4 and T3, and not entry by entry.

Fix. Add the missing critics to §5.3 with one-line assessments. In §6, either add a standing “countervailing check” line to each entry, or state in §6.1 that the four critics’ questions apply to every entry, and implement issue 6.

17. Smaller mechanisms from the digests and hindsight files that the lens should absorb#

Location. Various §6 entries.

Problem and fix (each is one clause added to an existing entry): - G2. Early controls hit the tractable segment, not the largest source: small boats before ships; cheap uses before costly ones; restricting individuals rather than sources; lowest commercial stakes first (LL1-13, pp. 138–139; LL1-06, pp. 66, 72; LL1-12, p. 128; LL2-16, p. 382; LL2-13, pp. 289–290). Rules keyed to stated purpose let the same agent continue under another label (LL1-09, pp. 93, 95; hindsight LL1-09, “prevention” labels). - G2 and G7. A statutory duty to investigate can be met in form while deferred in substance (hindsight LL2-05, lesson 3). A flagship case’s early choices become templates (hindsight LL2-13, lesson 9). - K6. Control achieved by the lead producer does not travel down a dispersed supply chain (hindsight LL2-06, lesson 4). Concentrated production can mean global exposure through supply chains (LL2-19, p. 468). - K2. Screens built around the last hazard reassure falsely about the next (hindsight LL2-12, lesson 3). Validation lag discounts newer, more sensitive methods (LL2-10, pp. 222, 229). Signals from test systems and surrogates are discounted as not relevant: early animal data for PCBs, DBCP, vinyl chloride and DES (T02 §1.1). - C4. When effects fall below what epidemiology can detect, “the number of victims” becomes a choice of method (hindsight LL2-18, lesson 4). - I5. Strategic or “critical” designation turns policy from reducing use to securing supply (hindsight LL2-06, lesson 9). - W3. Much of a late bill can go on buying back credibility and market access rather than reducing risk (hindsight LL1-15; T05 §3.9). Strategies that depend on controlling information fail abruptly (LL1-15, p. 164; digest LL1-15, insight 10). - T4. When a precautionary step is cheap, a lower evidence threshold is proportionate, a point accepted across the divide in the mobile-phone case (LL2-21, pp. 515, 518, 520; digest LL2-21, insight 7). - C (new line). The cost of delay includes unwinding lock-in and repairing credibility, not only physical harm (§4.5 row at line 405 has no lens counterpart; T05 §3.9).

18. Minor completeness and consistency points#