T10 — The canonical lessons: what the Late Lessons reports concluded, how the lessons changed 2001–2013, and how well they hold#

Strand A thematic synthesis. Written 26 September 2026.

Sources. Digests of all 47 sections; full notes for LL1-00, LL1-16, LL1-17, LL2-00, LL2-26, LL2-27, LL2-28, LL2-A2, LL2-A3 (and parts of LL2-22); hindsight files for those sections and LL2-02; external/critiques.md; external/context.md. Key quotations were checked against working/text/chunks/.

Conventions. - LL1 is the 2001 report; report page = PDF page. LL2 is the 2013 report; report page = PDF page − 2. Citations take the form “LL1-16, p. 171”. - “Hindsight LL1-17” means the hindsight file for that section. - LL1 is “All rights reserved”, so its wording appears only as short phrases. LL2 authorises reproduction with acknowledgement (LL2-00, p. 2), and the twelve lessons are given in full from its reprint (LL2-00, p. 11). - Labels: Reports say (what the text claims); Evidence (what the cases and later evidence show); Analysis (my inference). - Ratings: strong (several cases, consistent with later evidence); moderate (several examples, but gaps, selection or hindsight risk); suggestive (plausible, thin or single-case support); asserted (normative, or stated without evidence).

0. Summary of findings#

1. The twelve lessons are best read as a checklist of failure modes, not as findings from the cases. The lessons, and the synthesis around them, are mainly about how knowledge is produced and used. The editors say the framework came first: it was a pre-existing EU-funded appraisal scheme (the ESTO project). The fourteen cases were used to “test or elaborate” it (LL1-16, p. 168). The editors call the lessons “illustrative, rather than definitive” (LL1-16, p. 169).
2. The lessons are unevenly supported. Six rest on several cases and have been strengthened by later evidence: absence of search is not absence of harm; real-world conditions depart from assumed ones; blind spots sit inside the discipline that owns the problem; dependence on the regulated party’s own evidence; monitoring; and evidence thresholds as choices about who bears the cost of error. The claims about innovation, technological diversity, public intuition and trust are thinly evidenced and partly contested.
3. The 2013 volume kept the twelve lessons unchanged and called them “highly pertinent” (LL2-00, pp. 9, 11). It surrounded them with new material (section 3): a knowledge–power thesis, a false-positive review, a working definition, a six-way typology, a banded evidence scale, “criteria for action”, seven barriers, “harm expansion”, and a shift from regulating hazards to governing innovation. Its tone is more certain and more accusatory than 2001’s, and some 2001 hedges were dropped.
4. The diagnosis does not match the remedy, and the reports say so. The 2001 Preface ranks political will above trusted information (LL1-00, p. 4), yet the conclusions frame most lessons as better information (LL1-17, p. 194). In 2013, unequal power is named but put “well beyond the scope” of the report (LL2-28, p. 672).
5. Later uptake followed a gradient (analysis). Lessons that ask little of incumbents were adopted most readily: monitoring, uncertainty characterisation, transparency. Lessons that would re-price risk or shift the burden of producing evidence moved least: alternatives, polluter pays, pre-funded compensation, independent generation of data. No institution adopted the twelve lessons as a package.
6. The main weaknesses of the canon are structural. Every 2001 case was selected because harm occurred; the 2013 false-positive count (4 of 88) is fragile; risk-for-risk trade-offs are under-counted; and several 2013 “emerging” warnings have since moved toward reassurance. The mechanisms travel well. The frequencies and the promises about innovation do not.

1. How the lessons were built, and what the 2001 synthesis claims#

1.1 Method#

Reports say. Fourteen “well-known” hazards were chosen, “where sufficient is now known about their impacts”, and each author answered four questions: first credible warning; actions and inactions; costs and benefits and their distribution; lessons (LL1-00, p. 11). Authors were mostly “active participants” in their histories, and every case is a false negative: industry was invited to supply false positives, but “no suitable examples emerged” (LL1-00, pp. 12–13). The editorial team “distilled” the authors’ lessons into twelve under the EEA Scientific Committee’s guidance (LL1-00, p. 3). The ESTO project on technological risk (Stirling 1999) “provided the initial framing” (LL1-16, p. 168); section 16.3 deliberately goes beyond the cases (p. 169); and the editors grade their own lessons as empirically grounded, general, balanced and practical (p. 182).

Analysis. The lessons are a framework illustrated by cases, not an inductive typology. No coding method, counter-case search or tally is reported (notes LL1-16, bias check). Four of the seven editors also wrote case chapters: fisheries, asbestos, PCBs and MTBE (LL1-16 notes).

1.2 The headline diagnosis (LL1-16, p. 168; LL1-00, pp. 3–4)#

Reports say. - The gap between warning and effective action was “many years or decades, and in some cases over a century”. - Unequivocal precautionary action stayed “relatively scarce” even after the principle was articulated in the 1970s and 1980s. - In “many” cases, adequate information existed but did not reach decision-makers, or was discounted. - In “some” cases (asbestos, PCBs, the Great Lakes, sulphur dioxide), warnings were ignored because of “short-term economic and political interactions”. - Acceptance of precaution varies between institutions on both sides of the Atlantic, rather than between continents (LL1-16, p. 168). - The Preface adds that misplaced “certainty” about the absence of harm delayed action in “most” cases (LL1-00, p. 4); that prevention costs are tangible and short-term while the costs of inaction are diffuse and long-term (pp. 3–4); and that political will “seems to be an even more important factor” than trusted information (p. 4).

Evidence (hindsight LL1-16, Claims 1–3; hindsight LL1-00, Claims 3–4). - The long gap holds for the confirmed-hazard cases. For several agents it ran well past 2001: the US chrysotile asbestos ban came in 2024, 126 years after the 1898 warning. - The US–EU pattern is confirmed. A random-sample comparison found no significant average difference in precaution between the US and Europe, 1970–2004, and a “mixed and diverse pattern” (Hammitt et al. 2005). - The information strand is strengthened for PCBs, and for persistent chemicals generally, by documents disclosed in litigation. - A qualification. The Phillips Inquiry rejected producer bias in the UK agriculture ministry’s (MAFF’s) BSE policy decisions, so that case supports non-delivery of information more than interest-driven discounting. - Political will over information gains support from the stalled Rotterdam listing of chrysotile, but was never tested comparatively. Countries tend to ban asbestos once mesothelioma shows up in their own data, which suggests will follows locally owned evidence.

Rating. - Long gaps for confirmed hazards: strong. - The frequency claims (“most”): moderate, unverified case by case. - Political will ranked above information: suggestive as a ranking, with good case support.

1.3 The 2001 conclusions (LL1-17, pp. 192–194)#

Reports say. - Regulation must balance the costs of being too restrictive against the costs of being too permissive. Inaction led to costly consequences that were not, and “sometimes could not have been”, foreseen (p. 192). - Three kinds of failure recur: warnings ignored, including “loud and late” ones; appraisal scoped too narrowly; and actions taken without weighing alternatives or real-world implementation (p. 192). - The principle is defined as a broad “framework of thinking” governing foresight under uncertainty and ignorance (p. 192). - Table 17.1 pairs risk with prevention, uncertainty with “precautionary prevention”, and ignorance with “precaution” (p. 192). The “precaution” row does not depend on knowing the specific harm. It screens substances on persistence and bioaccumulation, casts the information net wide, monitors over the long term, and favours diverse, adaptable technologies with fewer “monopolies”. - The level of proof is “a key political decision with profound ethical implications”. It moves “the size, nature and distribution of the costs of being wrong”. It should depend on the harm, the claimed benefits, the alternatives, and error costs in both directions (p. 193). - Most lessons improve information. None removes the underlying dilemma, but together they “would at least increase the chances” of anticipating harm (p. 194). - Precaution “can also” stimulate innovation and better science. Over-precaution “can also be expensive” (p. 194). - The central conclusion: maximising innovation while minimising hazard is “ultimately a matter of political discourse”, and would go better if it embraced the twelve lessons (p. 194).

Evidence (hindsight LL1-17, all claims). - Failure diagnosis: held up, and was strengthened by cases the EEA did not select (PFAS knowledge by 1970; BPA and PFAS endpoints outside earlier appraisals; substitution from CFCs to HFCs). - Table 17.1: partly held up. It was taken up in scholarship and in European Parliament research-service guidance, but not in law. Its main successor adds “ambiguity” (Stirling 2007, 2010), a category the 2001 table lacks. - Level of proof as political: strengthened analytically. In Pfizer v Council (2002), the court said the scientific committee had no democratic legitimacy and the Council could depart from its advice. EU hazard cut-offs and the 2016 US chemicals-law amendments followed. Explicit two-way weighing of error costs remains rare. - Property screening and monitoring: strengthened (Stockholm Convention listings rising from 12 to 37; detection of illegal CFC-11 emissions). Diversity as insurance: untested. - Innovation: contested. The meta-analytic evidence centres on no significant effect. - The twelve lessons as a package: “unclear”. Nobody adopted them as a package, so there is nothing to test.

Rating. - Level of proof as a distributive choice: strong. - Property screening plus monitoring: strong as policy, moderate as proof that surprises get smaller. - The three failure types: moderate–strong within the cases. - Innovation, trust and “package” claims: asserted.

Analysis. Chapter 17 compresses Chapter 16 and drops its caveats: “illustrative, rather than definitive”, appraisal scaled to the stakes, no over-reliance on one set of prescriptions (LL1-16, pp. 169, 183), and the lowest rung of the evidence ladder, “scientific suspicion of risk” (p. 184). Without that rung, the Commission’s “reasonable grounds for concern” looks like the floor, though the editors’ own table puts a rung below it (notes LL1-17, §7).

2. The twelve lessons and their evidence base#

The wording below is as reprinted in LL2-00, p. 11. LL1-16 (pp. 168–169) and LL1-17 (pp. 193–194) use the same list; Chapter 16’s section headings word some lessons differently. The number of 2001 cases each lesson draws on is counted from the illustrations Chapter 16 names. Those illustrations are second-hand from the case chapters.

Overview table#

Lesson 1. “Acknowledge and respond to ignorance, as well as uncertainty and risk, in technology appraisal and public policymaking.”#

Reports say. - Every activity in the cases had some risk assessment. What was neglected was the near-certainty that some factors lay outside its scope (LL1-16, p. 169). - The illustrations are CFCs, DES, TBT and antimicrobials (pp. 169–170). - Where the harm cannot be predicted, certain properties can stand in for it: novelty, persistence, dispersal, irreversibility, and global scale with “only one ‘experimental’ model” (pp. 170–171; halocarbons, PCBs, MTBE, asbestos, TBT, fisheries, antimicrobial resistance). - The chapter separates two kinds of ignorance (p. 171): - institutional ignorance: knowledge exists in society but not at the point of decision. It features in “most” cases, and the remedy is better communication. - societal ignorance: nobody knows. It features in “many” cases, including BSE, and the remedies are research, diversity and flexibility. - 2013 adds ambiguity, variability and indeterminacy (Table 27.1, LL2-27, p. 656) and a qualitative “knowledge-to-ignorance ratio”, low for nanotechnology, GM food and mobile phones (pp. 654–655).

Evidence. - The distinction became standard in precaution scholarship and advisory work, mainly through one of the 2001 editors (hindsight LL1-17, Claim 2). - Property screening was institutionalised: Stockholm Convention listings grew from 12 to 37; persistent-bioaccumulative-toxic (PBT/vPvB) criteria entered REACH; and CLP added persistent-mobile-toxic (PMT/vPvM) classes in 2023 (hindsight LL1-16, Claim 4). - Bioaccumulation proved too narrow a screen for mobile PFAS. At least one acutely toxic transformation product fell outside persistence screening entirely. Nobody has tested whether screening made surprises “smaller” (hindsight LL1-16, Claim 4). - Critics turn the point around. Marchant argues that mesothelioma, the ozone hole and DES cancers were not foreseeable at deployment, so precaution could not have prevented them; the stronger charge is slow response once evidence appeared (external/critiques.md §4).

Rating. Strong as a conceptual lesson, and in the form “screen on properties that raise the cost of being wrong”. Moderate as an account of why the historical failures happened.

Lesson 2. “Provide adequate long-term environmental and health monitoring and research into early warnings.”#

Reports say. - Asbestos, benzene and PCBs had no systematic monitoring. The editors infer an assumption that harm would reveal itself in time (LL1-16, p. 172). - For BSE, reassurance cited absence of evidence “when no evidence was actually being sought” (p. 172). - Swann’s concerns about antimicrobials were not followed up until the 1990s. TBT baselines were “frequently under-rated” (p. 172). - Case authors make the same recommendation independently: Lambert (radiation, LL1-03, p. 36), Gee and Greenberg (asbestos, LL1-05, p. 60), TBT (LL1-13, p. 142), and Farman, whose ozone discovery came from long-term measurement (LL1-07, p. 82). - The chapter also concedes that research can “compound uncertainty”. A fisheries model grew less predictable as data were added, and Great Lakes research amplified uncertainty (pp. 172–173). - 2013 additions: underpowered monitoring can miss a halving of a population (LL2-26, p. 634); “surprise-sensitive” indicators with citizen science (LL2-28, p. 675); monitoring without pre-set action thresholds becomes “academic” (LL2-12, p. 274); vigilance fades after a crisis, the “homo-illogical cycle” (LL2-28, p. 680).

Evidence. - Strongly vindicated: EU active BSE testing from 2001 found about 7,000 cases among about 50 million cattle, confirming that clinical surveillance had missed many (hindsight LL1-16, Claim 10); atmospheric monitoring exposed illegal CFC-11 production; and wildlife sampling revealed that PFOS was globally distributed (hindsight LL1-17, Claim 3). - Radiation surveillance databases proved their worth, and some units were closed when no need was perceived (hindsight LL1-03). - The fading of vigilance after a crisis is supported for political direction (nuclear policy), but not for Japan’s post-Fukushima regulator (hindsight LL2-28, Claim 10).

Rating. Strong. Many cases and independent later confirmation, with the reports’ own caveat: monitoring must be powered, reported and tied to decisions.

Lesson 3. “Identify and work to reduce ‘blind spots’ and gaps in scientific knowledge.”#

Reports say. Blind spots sit “at the heart of” the disciplines in charge (LL1-16, p. 174): ozone satellite data “set aside”; Swann’s advice marginalised; MTBE’s persistence under-questioned; wrong assumptions about TBT degradation; low-oestrogen children missed in the hormones case; radiation risk models built on high-dose bomb survivors (p. 173). The “problem already solved” belief recurs: asbestos from 1906, and substitutes with similar fibres (pp. 173–174). The 2013 volume turns this into a list of nine initial assumptions later shown to be wrong, among them “safe” or natural doses, the adult male as reference subject, acute effects standing in for chronic ones, protective barriers, averages, monotonic dose-response, and dismissal of animal data (LL2-26, p. 630, Table 26.3; LL2-28, p. 677).

Evidence. - Since 2001, the endpoints that finally drove action lay outside earlier appraisals: immune (Th17) effects for BPA, vaccine antibody response for PFAS (hindsight LL1-17, Claim 1). - Farman’s own account supports “flagged and re-examined” rather than “set aside”, a small editorial overstatement (LL1-07 digest).

Rating. Strong. Many cases, and later confirmation outside the EEA’s selection. Table 26.3 is compiled with hindsight: it lists only assumptions that failed.

Lesson 4. “Identify and reduce interdisciplinary obstacles to learning.”#

Reports say. The discipline that first sees effects can hold appraisal “captive” (LL1-16, p. 174): clinicians’ acute focus (asbestos, radiation); engine and air-quality expertise for MTBE, with water effects “essentially disregarded”; a health-built sulphur dioxide regime that struggled with ecology; a human-health focus for hormones; veterinary framing for antimicrobials and BSE. Swann’s single cross-sectoral advisory process was never set up (p. 174). Chapter 17 asks for health and environment to be integrated (LL1-17, p. 193). In 2013 the editors reported that “bureaucratic silos” had not disappeared (LL2-28, p. 670). They also criticised risk committees dominated by one discipline (bees, lead, BPA, nuclear; LL2-28, p. 677).

Evidence. Slow structural change. The EU’s “one substance, one assessment” package came into force only in 2026 (hindsight LL2-00, Claim 10). In the one prospective audit inside the reports, this lesson was noted in four of eleven expert reports and in none of the regulations (LL2-22, p. 548).

Rating. Moderate. Seven illustrative cases; by the reports’ own 2013 judgement, little changed.

Lesson 5. “Ensure that real world conditions are adequately accounted for in regulatory appraisal.”#

Reports say. Technologies are assumed to perform to specification (LL1-16, pp. 174–175): MTBE tanks leaked; PCB “closed systems” leaked and were dumped; halocarbon containment failed; benzene reached people by uncontrolled routes; radiation doses for the same examination varied 100-fold between hospitals; hormones were assessed only for authorised, single-substance use; and there was non-compliance in PCB disposal, fisheries, BSE slaughterhouses and halocarbon smuggling, plus neglected sensitive subgroups. The WTO Appellate Body (2001) found that “controlled use” of asbestos could not be relied on (p. 175; LL1-05, p. 57).

Evidence. - The Phillips Inquiry found the specified-offal ban for animal feed “unenforceable and widely disregarded”. - Illegal CFC-11 production was detected by monitoring, not by reporting (hindsight LL1-16, Claims 7 and 10).

Rating. Strong. The widest case support (about ten cases), repeated by later events.

Lesson 6. “Systematically scrutinise the claimed justifications and benefits alongside the potential risks.”#

Reports say. - Partial appraisal of fixes: tall stacks, MTBE, hormones (LL1-16, p. 176). - DES trial data from 1953 showed it did not work, yet it was prescribed for another 20–30 years (p. 176; LL1-08, pp. 86, 90). - The radiation “justification principle” exists, but many X-rays remain of doubtful use (p. 176). - Unpriced harms gave asbestos, halocarbons and PCBs an “unjustifiable advantage” (pp. 176–177). - 2013 additions: - acceptable risk belongs to a use, not to a substance (LL2-04, pp. 80, 83); - the benefits of technologies are “demonstrated” only by market success, while harms face formal assessment (von Schomberg, LL2-27, Box 27.5, p. 661); - “Indeterminacy” should be met with pre-market assessment of benefits (Table 27.1, p. 656).

Evidence. - The US federal efficacy review in 1971 still rated DES “possibly effective”, 18 years after the negative trial. Action came only when a vivid harm appeared (hindsight LL1-08). - Growth in CT scanning overwhelmed per-examination dose reductions (hindsight LL1-03). - The claim that growth promoters had little benefit held as policy, but the “net loss” economics were not generalised (hindsight LL2-A3, Claim 8).

Rating. Moderate. DES is compelling, and radiation shows a mechanism that only partly works. The other illustrations concern narrow appraisal rather than benefit scrutiny.

Lesson 7. “Evaluate a range of alternative options for meeting needs alongside the option under appraisal, and promote more robust, diverse and adaptable technologies so as to minimise the costs of surprises and maximise the benefits of innovation.”#

Reports say. - Once committed, a technology is reinforced “even if markedly inferior” (LL1-16, p. 177). Lock-in can be arbitrary, through “chance and first-leader advantage” (p. 187). - The examples: MTBE adopted without scrutiny of alternatives; TBT’s “chemical for chemical” substitution; X-ray alternatives underused; second-generation CFC substitutes “unduly tolerated”; husbandry without antimicrobials not promoted (p. 177). - Diversity is insurance: the surprise is smaller with several technologies than with one “near monopoly” (asbestos, halocarbons, PCBs; p. 187). - 2013 additions: - late action “consolidated technological monopolies” at “unrealistically low prices” (LL2-27, p. 659; LL2-28, p. 673); - regrettable substitution: PCE, lead, DDT, booster biocides, CFCs to HFCs, MTBE (LL2-28, p. 676); - “feasible alternatives” is one of the twelve criteria for action (Box 27.4, p. 653).

Evidence. - Alternatives assessment was institutionalised: a US National Academies framework (2014), EU “safe and sustainable by design” (2022), and a General Court annulment of an authorisation for failing to assess alternatives (hindsight LL2-02, Claim 10; hindsight LL2-00, Claim 8). - Regulators moved to class-wide restrictions (PFAS group proposal; 2024 bisphenol rule) because substitution kept recurring within families, e.g. CFCs → HCFCs → HFCs → HFOs, whose breakdown product TFA persists (hindsight LL1-16, Claim 8; hindsight LL2-28, Claim 2). - In South Africa, withdrawing DDT before a substitute was proven backfired (LL2-11, pp. 243, 250). - Nobody has tested diversity as insurance (hindsight LL1-17, Claim 3). - In the prospective nanotechnology audit, the alternatives row was completely blank (LL2-22, p. 548).

Rating. Moderate for the alternatives and lock-in half. Suggestive for diversity as insurance.

Lesson 8. “Ensure use of ‘lay’ and local knowledge, as well as relevant specialist expertise in the appraisal.”#

Reports say. - Workers (asbestos, PCBs), communities (Love Canal), fishers, Swedish farmers and BSE slaughterhouse workers often knew first (LL1-16, pp. 177–178). - Lay knowledge needs the same scrutiny and has its own fallacies, such as the “pensioners’ party” error of taking healthy retirees as proof of safety (p. 178); fishers can be less precautionary (p. 177). - 2013 adds more early warners: a Minamata mother, the wives of DBCP workers, DES mothers, beekeepers, clinicians and factory inspectors (LL2-28, p. 675).

Evidence. - The MMR alarm shows lay concern can be wrong and costly (hindsight LL2-02, Claim 9). - The Keats report in the fisheries case worked through a commissioned scientific re-analysis, not raw observation (LL1-02 digest). - Harada, the Minamata researcher, at first dismissed the mother’s hypothesis himself (notes LL2-28).

Rating. Moderate. Several cases, and the 2001 chapter is notably balanced. The 2013 claim that non-scientists “reliably” gave early warnings is not quantified. Nobody examines how often lay warnings were false.

Lesson 9. “Take full account of the assumptions and values of different social groups.”#

Reports say. - Public intuitions “may sometimes prove quite robust”. Aversion to the unfamiliar is defensible as a rational response to uncertainty (LL1-16, p. 178). - The examples: revulsion at feeding offal to cattle (BSE), misgivings about antimicrobials, and the value choices embedded in fisheries objectives. - Recent controversies (GMOs, disposal of an oil storage installation) turn on values (LL1-17, p. 193). - 2013 adds “participatory precaution” for ambiguity (LL2-27, p. 656) and a proposed institutional place to analyse value conflicts (LL2-28, p. 680).

Evidence. - Values were written into law: member-state opt-outs on GMO cultivation (Directive 2015/412) and the OSPAR ban on dumping offshore installations (1998). - Brent Spar: the outcome rested partly on a campaigner’s oil estimate that was wrong and later withdrawn. - Participation and legitimacy: participation raises perceived legitimacy mainly when it changes outcomes (hindsight LL1-17, Claims 6–7).

Analysis. The 2001 examples were selected because public intuition proved right. Public rejection of irradiated food and GMOs appears only as a “cost of failure” of traditional approaches (LL1-16, p. 188). Whether it was well founded is not asked.

Rating. Suggestive as an epistemic claim. Moderate as a claim about legitimacy.

Lesson 10. “Maintain the regulatory independence of interested parties while retaining an inclusive approach to information and opinion gathering.”#

The wording is ambiguous. Chapter 16’s heading clarifies it as independence from economic and political special interests (LL1-16, p. 178).

Reports say. - Appraisal “frequently fails” because it depends on information “produced and owned” by the parties being assessed (LL1-16, p. 179). - The examples: progress on benzene, asbestos and PCBs only in the 1960s–70s; MAFF “responsible first to the industry”; weak “refutations” sustained (PCBs, acid rain); suppression (Great Lakes); dismissal of state scientists (Californian sardines); and the Southwood committee trimming its advice to what was politically “realistic” (p. 179). - The chapter itself hedges: “Not all of these cases demonstrate” the distorting effect (p. 179). - 2013 makes this central: “the value of independent scientific research and risk assessments” is one of three common themes (LL2-00, p. 10); product defence and manufactured doubt (LL2-27, pp. 645–646); eliminate financial ties rather than only disclose them (LL2-26, p. 636).

Evidence. - EU law accepted the diagnosis but deliberately kept applicant-generated data. The 2019 Transparency Regulation added study notification, disclosure and verification studies. - The Court of Justice in Blaise (2019) told authorities not to give applicant studies automatic “preponderant weight”. - An independent re-analysis of industry chlorpyrifos data found effects the summary had not reported (hindsight LL1-16, Claim 5). - Complications: independent public assessors diverge on the same evidence, by up to 1,000-fold on BPA (hindsight LL2-00, Claim 2); an industry–union joint draft helped produce a stricter beryllium standard (hindsight LL2-06); and Phillips rejected producer bias in MAFF’s policy decisions.

Rating. Strong that dependence on interested parties’ evidence is a structural weakness. Moderate that independence alone yields timely or agreed decisions.

Lesson 11. “Identify and reduce institutional obstacles to learning and action.”#

Reports say. - Short political and business cycles (asbestos, benzene, PCBs). - Changes of government that reversed protective measures: UK rendering standards withdrawn in 1979; the Californian sardine programme. - Friction between departments and between levels of government: MAFF told the Department of Health about BSE only after about 17 months; the federal fisheries bureau opposed state precaution. - Agencies defending their past decisions. - National differences in accepting the same science: Norway in 1976 versus the UK in 1985 on sulphur dioxide; the US and UK on scrapie (LL1-16, pp. 180–181). - 2013 names this as the lesson with “less progress” (LL2-28, p. 670), and adds the “moving target” (by the time harm is confirmed, the technology has changed and “today’s technology is now safe”), sunk-investment lock-in, and scale that overwhelms monitoring (p. 672).

Evidence. - Protective rules are fragile under changes of administration (US chemical and PFAS rollbacks in 2025–26). - The EU switched to competitiveness framing (hindsight LL2-27, Claims 1 and 10). - The US scrapie decision could not be verified independently (hindsight LL1-16, Claim 2).

Rating. Moderate. The illustrations are well chosen but not analysed systematically. The “moving target” is suggestive: it is asserted without a worked example (notes LL2-28).

Lesson 12. “Avoid ‘paralysis by analysis’ by acting to reduce potential harm when there are reasonable grounds for concern.”#

Reports say. - Paralysis comes from “information overload or lack of political will” (LL1-16, p. 181). The examples: the US Supreme Court’s benzene “straitjacket”, and Swann’s refusal to let “the cry for more research” hold up action. - The editors acknowledge the tension with “know more” (p. 181) and offer two resolutions: state whether hazard reduction should precede the research (p. 173); and distinguish prospective appraisal, where broadening is precautionary, from retrospective appraisal, where it can weaken the case against any single agent (p. 181). - Whether a demand for more evidence counts as “prudence” or “paralysis” depends on who bears the costs (p. 182). - 2013 additions: - once compelling human evidence appeared, four to seven cases were enough to trigger rapid action on DES, vinyl chloride and DBCP (LL2-27, p. 645); - nano governments used calls for information “as a substitute for action” (LL2-22, p. 547); - “good enough” epidemiology in Minamata (LL2-05, pp. 111–112).

Evidence. - Pfizer (T-13/99, 2002) upheld action on “reliable” but incomplete data, not on “mere conjecture”. - The Court of Justice upheld the neonicotinoid restrictions under the precautionary principle (C-499/18 P, 2021). - “Reasonable grounds for concern” remains “poorly defined” in EU legislation (Garnett and Parsons 2017). - US policy in 2025 moved the other way (hindsight LL2-00, Claim 10; hindsight LL2-27, Claim 1).

Counter-cases within the reports. - Hormones. The EU ban was “in reality, a political risk assessment” made against both expert committees, with no demonstrated health benefit and trade sanctions to show for it, though “probably” justified by later research (LL1-14, pp. 153–154). The dispute was settled by quota, not by agreement on the science (hindsight LL2-A2, Claim 6). - Warnings later weakened. The 2013 volume’s own mobile-phone and GM-health warnings have since moved toward reassurance (hindsight LL2-28, Claims 1 and 5).

Rating. Moderate that demands for more evidence work as delay. Asserted as a general decision rule, because the trigger is undefined and the reports give no method for telling signals apart in advance.

3. What the 2013 volume added (Introduction; Part E; Annexes 2–3)#

3.1 Introduction and Preface (LL2-00)#

Reports say. - The frame. The subtitle becomes “science, precaution, innovation”; the Preface centres “the relationship between knowledge and power” (LL2-00, p. 7). A second volume was needed for more cases, false positives, emerging technologies and systemic challenges (p. 9). - What the nine new health cases show. Three themes: “more than sufficient evidence for much earlier action”; slow and sometimes obstructive business; the value of independent science (p. 10). - Four reasons for delay: the novelty of the issues; poorly evaluated information; opposition by “corporate and scientific establishments”; status-quo institutions (pp. 10–11). - Who caused the harm. Harms were “for the most part” caused by “irresponsible corporations” (p. 11). - False positives are “few and far between” (p. 10, from LL2-02).

Evidence (hindsight LL2-00). - The persistence and vulnerable-window claims were strengthened. - “Irresponsible corporations” is weakened by harms driven by public authorities: Flint, Camp Lejeune, and state liability for Minamata. - The trust claim is weakened by a 68-country survey (Cologna et al. 2025). - “4 of 88” circulates in policy documents without independent re-testing.

Rating. Persistence and developmental windows: strong. The three themes: moderate (seven of nine cases strengthened; BPA contested; DDT a real trade-off). “For the most part … irresponsible corporations”: asserted, and weakened.

3.2 Chapter 26, Science for precautionary decision-making (Grandjean)#

Reports say. - Research concentrates on well-studied substances (top 20 took 12% of chemical links; 13 data-poor priority chemicals attracted almost no work; LL2-26, pp. 625–628). Standard design tilts toward false negatives (Table 26.4: ten features versus three; p. 635). - Ask “how large an effect can the study have overlooked?” Use upper confidence limits as a plausible worst case (pp. 633–635). - A new research ethos, “PATIO”: participatory, accessible, transparent, inventive, open-minded (p. 638). - Science’s task is to judge whether evidence justifies starting “transparent and democratic procedures” (p. 638). - Four recommendations to researchers (p. 639). The verdict: science “does not have a good track record” (p. 640), though footnote 2 concedes political will may matter more (p. 624).

Evidence (hindsight LL2-26). - The flagship examples (PFAS, BPA) were vindicated in the EU. - The one-directional error claim is the most weakened: the replication crisis, the ocean-acidification decline effect, and the report’s own mobile-phone chapter all show false positives at research level. - “Ask what could have been missed” is now mainstream. - The warning that openness can be turned against research was borne out.

Rating. Strong on absence of proof and missed effect sizes. Moderate on research inertia. Two-sided, not one-sided, on error.

3.3 Chapter 27, More or less precaution? (Gee)#

Reports say. - Two roles. The principle justifies timely action, and it triggers debate about which innovation pathways a society wants (LL2-27, p. 644). - Seven barriers (p. 645): corporate opposition; confusion over definitions; complexity; the gap between the high evidence bar for causation and the lower bar for policy; skewed cost analysis; short-termism; lack of public engagement. - EEA working definition (p. 649). The principle justifies action under complexity, uncertainty and ignorance, against potentially serious or irreversible threats, “using an appropriate strength of scientific evidence” and weighing the pros and cons of action and inaction “and their distribution”. - Analytical tools: Bradford Hill’s causal features reappraised for multicausality, so that their absence is weak evidence of no effect (pp. 651–654); twelve “criteria for action” (Box 27.4, p. 653); and an IPCC-derived evidence scale with probability bands (Table 27.2, p. 658), in which Sweden’s 1973 law appears at both ends (manufacturers must show safety at “very strong”; regulators may act at “weak”). - Communication points. “Not established” judgements seldom say who bears the error, “risk takers or risk makers” (p. 658). Risk assessors should consider consequences (p. 658). - Participation. A six-stage participatory risk-analysis cycle (Fig. 27.2, p. 660). Five of 34 cases are said to describe precautionary actions (pp. 644–645); two are contested (hormones, EU GMO rules).

Evidence (hindsight LL2-27). - The barriers held up. - The fight over the standard of proof became explicit policy. The EU moved toward graded evidence; US Executive Order 14303 (2025) disfavours “overly precautionary assumptions”. - The Interphone example is selective. - The “emerging hazards” diverged: chemicals moved toward the chapter’s view, GM food and mobile phones away from it. - The innovation claim is contested.

Rating. Strong for the procedural tools. Asserted for the net-benefit prescription.

3.4 Chapter 28, In conclusion (unsigned)#

Reports say. - Common features of the cases: a few decide for many; no mechanisms to respond to warnings; misleading prices; poor accounting across capitals (LL2-28, p. 671). - Three opportunities: rebalance capitals through precaution, prevention and polluter pays; broaden evidence and engagement; build adaptive governance (pp. 671–672). - Six findings. The first is to reduce delays. It includes: - the moving target; - sunk-cost lock-in; - “harm expansion”: confirmed hazards prove more diverse and harmful at lower doses (p. 672); - an asymmetry argument under irreversibility: tip policy “towards avoiding harm, even at the cost of more false alarms” (p. 673). - Power. Unequal power is “well beyond the scope” of the report (p. 672). - The close. Everything proposed remains “good intentions” until institutionalised (p. 680).

Evidence (hindsight LL2-28). Harm expansion is strengthened for the named agents: limits for lead, asbestos, PM2.5, PFAS and BPA were cut after 2013. The headline statistics are fragile: “4 of 88”; the 1% hazard-research figure, which conflicts with Chapter 27’s 3%; and “half of all articles”, which misreads the underlying study. The GM human-health sentence is weakened, and two quotations attributed to the Fukushima inquiry are misattributed.

Rating. Harm expansion: strong for confirmed hazards, moderate as a general prior (selection). The irreversibility asymmetry: moderate, a sound conditional whose premises (reversible measures, continued research, no irreversible costs of acting) are not tested. The moving target and sunk-cost lock-in: suggestive, with no worked example. The headline statistics: not reliable.

3.5 Annexes 2 and 3#

Reports say. - Annex 2. Table A2.1 (p. 702, from Gee 2009) gives “years of substantial inaction” per case: asbestos 101, PCBs about 100, benzene 81, radiation 65, down to BSE 10–17 and TBT 5–30. - Annex 3 updates nine 2001 cases. Its editors state that harm “expands over time”, often at exposures once thought “safe” (LL2-A3, p. 717).

Evidence. - Annex 2’s lag figures follow inconsistent dating rules. The PCB and Great Lakes figures do not follow from the table’s own entries. In the analysis notes, lags shrink with the era of the first warning. That could reflect improvement, or truncation (notes LL2-A2). - “Effective action” was a sequence over decades, not a date. EU asbestos limits were cut again in 2023, the US banned chrysotile in 2024, and the EU benzene limit fell to 0.2 ppm from 2026 (hindsight LL2-A2, Claim 1). - Annex 3’s warnings were reliable on direction but weak on magnitude and mechanism. The claims that failed mostly rested on contributors’ own unpublished work (hindsight LL2-A3). Its own ozone update is a story of recovery and does not illustrate “harm expands”.

Rating. Decades-long lags: moderate (true in direction, not a consistent measure). “Harm expands”: strong for established hazards, with counter-cases (MTBE, farmed fish, hormone residues).

4. How the lessons evolved, 2001 → 2013#

Analysis. The 2013 volume adds real tools: the working definition, the evidence scale, the ambiguity row, and the risk-taker/risk-maker distinction. It also expresses more certainty than its evidence base had gained: the innovation and false-positive claims firmed up on evidence from the same network, and Chapter 28’s three most-cited statistics cannot be reproduced.

Some 2001 candour survived: the business chapter warns against blaming “with hindsight” (LL2-25, pp. 607, 616); Chapter 26’s author calls himself “part of the inertia” (LL2-26, p. 628, fn 3); the Bayer panel and the authors’ reply are printed side by side (LL2-16); and Chapter 28 admits it leaves power out of scope (p. 672).

5. Recommendations: what, to whom, and with what result#

Analysis. The addressees shift across the two volumes: - 2001 mostly addresses regulators and scientists, and frames remedies as information and institutional design. - 2013 extends to business, finance, citizens and innovation policy. It proposes the most re-distributive instruments, such as pre-funded compensation and tax shifts, and these are the ones that moved least.

6. Cross-cutting patterns#

P1. The reports diagnose power but prescribe information. (Analysis; moderate.) - The 2001 Preface ranks political will above information (LL1-00, p. 4), yet the conclusions frame most lessons as information (LL1-17, p. 194). Chapter 26 cites that Preface judgement, then addresses the information side (LL2-26, p. 624, fn 2). Chapter 28 names power asymmetry and rules it out of scope (LL2-28, p. 672). This fits the EEA’s mandate as an information agency (external/context.md §1.1). - For a lens, the institutional and political-economy mechanisms (lessons 10–11, Chapters 25 and 27) deserve more weight than the reports’ own summaries give them.

P2. The evidence threshold is the fulcrum, and it became openly political. (Strong.) - The same move runs through every synthesis chapter: Box 16.1 and Table 16.1 (LL1-16, pp. 170, 184), LL1-17 (p. 193), Table 27.2 (LL2-27, p. 658), LL2-28 (p. 673). - Courts and legislatures have since treated the threshold as a choice: Pfizer (the scientific committee lacked democratic legitimacy), EU hazard cut-offs, and the 2016 US chemicals-law reform excluding costs from risk findings. In 2025 the US reversed direction. - The reports never specify who should set the threshold or how its factors should be weighed (notes LL1-17, §7).

P3. Uptake followed a gradient from low-cost to redistributive lessons. (Analysis; moderate.) - Adopted furthest: monitoring, uncertainty characterisation, transparency, graded hazard classes. Adopted in part: independence (applicants still generate the evidence; disclosure was added on top). Barely adopted: alternatives assessment as a default, diversity, polluter-pays tax shifts, pre-funded compensation. - The reports’ own prospective nanotechnology audit shows the same pattern: knowledge lessons widely noted; alternatives, independence, lay knowledge and institutional obstacles rarely (LL2-22, p. 548). A co-author of that chapter commissioned this analysis.

P4. Warnings are more reliable about direction than about magnitude. (Strong.) - Annex 3 was right about which way gasoline, asbestos, BPA, growth promoters and ozone would move, and wrong about several magnitudes and mechanisms (hindsight LL2-A3). - “Harm expansion” is a strong prior for confirmed hazards and a weak one for suspected hazards (hindsight LL2-28, Claim 3).

P5. “Effective action” is a process, and adopting a rule is not the same as reducing risk. (Strong.) - Table A2.1’s single dates hide decades of follow-on measures (hindsight LL2-A2, Claim 1). - The offal ban and CFC-11 show rules without verified compliance (lesson 5).

P6. The move from hazard control to innovation governance invited an organised counter-frame. (Moderate.) - The 2013 innovation claims were answered by an industry “innovation principle”, written into Horizon Europe (2021) and echoed in the Draghi report (2024). - As Chapter 17 itself says, the balance between the two is “ultimately a matter of political discourse” (LL1-17, p. 194). Both sides argue from thin evidence (external/critiques.md §7).

7. Counter-evidence, complications and critiques#

1. Selection on the outcome, and hindsight. - All fourteen 2001 cases were chosen because harm occurred (LL1-00, p. 11), and outcome knowledge inflates judgements of foreseeability (Fischhoff 1975, in hindsight LL1-16, Claim 1). The reports therefore support mechanisms, not frequencies; the missing denominator is critics’ strongest structural objection (external/critiques.md §4). - The reports’ own standard (“the spirit of the times”, LL1-00, p. 11) is applied to excuse two of the four false positives (LL2-02, pp. 31–32), but not symmetrically to the false negatives.
2. The false-positive record is weak on both sides. - In 2001 the report called false positives the “smaller” risk while having found none (LL1-00, p. 16). - In 2013, “4 of 88” relies on four choices that each lower the count: a “high confidence” bar before a case counts as a false positive; no stated bar for calling a risk “real”; exclusion of market effects and unregulated alarms; and a sample drawn from critics’ lists. - Cox (2007) disputed the criteria. Neither of the 2001 candidates (North Sea sludge, Y2K) was assessed (hindsight LL1-00, Claim 6). - Hindsight shows false positives can be long-lived: saccharin labelling lasted 23 years, and irradiation approvals stalled for about 15. Unregulated alarms can be costly: MMR (hindsight LL2-02, Claims 7, 9).
3. Risk-for-risk trade-offs are under-counted. The reports’ own cases include protective fixes that went wrong: MTBE adopted to solve lead, tall stacks that exported harm (LL1-17, p. 192). Therapeutic antibiotic use rose after the growth-promoter bans (hindsight LL1-17, Claim 9), and the Stockholm DDT exemption embodies a trade-off with malaria. Yet LL2-02 files trade-offs as “mistaken” false positives, keeping them out of the ledger, a point Graham and Wiener, and Sunstein, press (external/critiques.md §3.2).
4. The hormones case inside LL1. The one 2001 case about a contested precautionary restriction goes unused in the conclusions. It shows an action against expert advice, driven by public concern, with trade costs and no demonstrated health benefit (LL1-14, pp. 153–154; notes LL1-17, limitation 2). In 2013 it is listed among the precautionary actions that, the chapter says, “illustrate the value” of the principle (LL2-27, pp. 644–645).
5. Internal tensions the editors acknowledge. - “Know more” versus “don’t wait” (LL1-16, p. 181). - Independence (lesson 10) versus inclusion (lessons 8–9). - Lifting a restriction needs research that “genuinely reveals” a concern is unfounded (p. 173), while keeping it needs only unresolved uncertainty (p. 181), which amounts to a ratchet.
6. Operational vagueness. “Reasonable grounds”, “appropriate strength of evidence” and “adequate” have no criteria (notes LL1-17, limitation 6). A review of 15 EU laws found the decision to invoke precaution “poorly defined” (hindsight LL2-00, Claim 1). The incoherence critique is largely met by the two-sided EEA definition, but the operational one is not (external/critiques.md §9).
7. Claims weakened by later evidence. Innovation holds only in the weak Porter sense (hindsight LL1-17, Claim 8). “Very low” trust in scientists was not supported even by 2001 Eurobarometer data (hindsight LL1-00, Claim 10). The one-directional false-negative claim is contradicted by the replication crisis (hindsight LL2-26, Claim 4). “For the most part … irresponsible corporations” is weakened by public-authority harms (hindsight LL2-00, Claim 4). So are Chapter 28’s statistics and GM-health sentence (hindsight LL2-28).
8. Factual slips in the synthesis chapters. Swann is dated 1967, 1968 and 1969; “PCB-induced chloracne” is placed in 1899, before PCBs were made; the 1974 US DES reversal, a court ruling, is attributed to the farm lobby; halocarbon “regulatory neglect” ignores the 1977 aerosol bans; BSE maternal transmission is overstated (hindsight LL1-16, Minor checks); Table 1.1 dates the EU DES ban 1987 instead of 1981; and the loose Peto asbestos range is repeated in 2013 as “some 400 000” (hindsight LL1-00).
9. Standpoint. Both volumes disclose participant authorship (LL1-00, p. 12; LL2-00, pp. 9–10). The editors wrote the false-positive chapter and the concluding chapters. Supportive follow-up (Hansen and Tickner; Hansen and Gee 2014; Stirling’s later typology) comes largely from the same network. Where this matters, the 2013 confirmations of 2001 claims are not independent.

8. Open items and verification notes#

• Annex numbering. The local “(05-2013)” file has both Annex 2 (overview, pp. 701–716) and Annex 3 (updates, pp. 717–737); its contents page lists them 30 pages later (notes LL2-00). The DTU-hosted copy uses that later pagination and has no Annex 3 (hindsight LL1-00). Inference: it is a different, probably earlier, version. Cite the version used.
• “Misplaced certainty … in most of the case studies” (LL1-00, p. 4) still needs a case-by-case check. The digests suggest it fits BSE, asbestos, radiation, DES, antimicrobials and sulphur dioxide (dilution judged harmless). It fits less well for hormones (a contested restriction), MTBE (the threat was “never considered”, closer to ignorance) and fisheries (a regime precautionary in intent).
• The trigger in lesson 12 has no operational definition in either volume. Any lens built on it must supply one.

9. Technology-neutral diagnostic questions#

1. For each sub-question, is the state of knowledge risk, uncertainty, ambiguity or ignorance, and is the appraisal tool matched to it? (Lesson 1; Box 16.1, LL1-16, p. 170; Table 27.1, LL2-27, p. 656; hindsight LL1-17, Claim 2.)
2. Which properties would make being wrong expensive (persistence, irreversibility, ubiquity, mobility, reach, novelty)? Are they used as triggers even without a named harm? (Table 17.1, LL1-17, p. 192; Box 27.4, LL2-27, p. 653; hindsight LL1-16, Claim 4.)
3. When someone says “no evidence of harm”, was the search actually done, how long was follow-up relative to latency, and how large an effect could it have missed? (LL1-16, pp. 172, 184; LL2-26, pp. 633–635; LL2-27, p. 658; BSE active testing.)
4. Who monitors what, with enough power to detect what, and who pays for monitoring through quiet periods? (Lesson 2; LL2-26, p. 634; LL2-28, pp. 675, 680; CFC-11 detection.)
5. Which discipline or institution owns the appraisal, and which endpoints, populations or media fall outside its frame? (Lessons 3–4; LL1-16, pp. 173–174; Table 26.3, LL2-26, p. 630; BPA and PFAS endpoints.)
6. Does the appraisal assume ideal performance, containment and compliance? How would misuse, leakage or non-compliance be detected by someone other than the operator? (Lesson 5; LL1-16, pp. 174–175; BSE offal; CFC-11.)
7. Are the claimed benefits and efficacy tested as rigorously as the risks, and by whom? (Lesson 6; DES, LL1-08, pp. 86, 90; Box 27.5, LL2-27, p. 661.)
8. What alternative ways of meeting the need were assessed? Would a restriction push users toward close analogues that share the problematic property? (Lesson 7; LL1-16, p. 177; LL2-28, p. 676; hindsight LL2-28, Claim 2.)
9. Who generates and controls the evidence? Are all studies registered in advance, raw data accessible, and independent verification funded? (Lesson 10; LL1-16, p. 179; hindsight LL1-16, Claim 5.)
10. What level of evidence triggers what action, who set it, was it set openly or by default, and who bears the cost if it is too high or too low (“risk takers or risk makers”)? (Lesson 12; LL1-17, p. 193; Table 27.2 and p. 658, LL2-27; hindsight LL1-17, Claim 5.)
11. Does the same body both promote and regulate? What has it already publicly committed to that would make changing course look like an admission of error? (Lesson 11; LL1-15, pp. 161–162; LL2-06, pp. 148–150.)
12. Whose operational and local knowledge has a channel into appraisal, and is it scrutinised as rigorously as expert claims? (Lesson 8; LL1-16, pp. 177–178; LL2-28, p. 675.)
13. Which value choices are embedded, and does participation actually change decisions? (Lesson 9; LL1-17, p. 193; hindsight LL1-17, Claims 6–7.)
14. How would a false alarm be recognised and reversed, and what do the precautionary measures themselves cost: forgone benefits, substitute risks, and who bears them? (LL1-17, pp. 192, 194; LL2-28, p. 673; hindsight LL2-02, Claims 7, 9.)
15. Is “action” a planned sequence with verified reductions in harm, or a single rule taken as the endpoint? (Annex 2, LL2-A2, p. 702; hindsight LL2-A2, Claim 1.)