T01 Knowledge, uncertainty, ignorance and surprise#

Cross-cutting thematic synthesis, Strand A. Written 26 September 2026 from the section digests, audited notes and hindsight files in working/late-lessons/, plus external/critiques.md and external/context.md. Notes read in full for this theme: LL1-16, LL1-17 (Table 17.1 and levels of proof), LL2-26 (sections 26.3–26.10) and LL2-27 (sections 27.3–27.8), plus the two external files. Other notes and hindsight files were searched by term and read where relevant.

Conventions - LL1 is EEA 2001 (report page = PDF page); LL2 is EEA 2013 (printed page = PDF page − 2). Citations give section id and report pages, e.g. (LL1-15, pp. 163–164). - Each pattern separates Reports (what the reports say), Evidence (how well the reports’ own cases support it, and what the hindsight files show since) and Analysis (my interpretation; my inferences are labelled). - Ratings: strong = four or more cases across both volumes, some primary or independent support, not weakened later; moderate = several cases but with protagonist sourcing, hindsight risk or real counter-cases; suggestive = one or two cases, or an inference; asserted = no case evidence (normative or theoretical). - The hindsight files were compiled largely by direct retrieval without web search; their access caveats carry over.

1. Main findings in brief#

1. The typology is sound and is the reports’ most distinctive conceptual move. “Uncertainty” hides distinct states (risk, uncertainty, ignorance; LL2 adds ambiguity, variability, indeterminacy) that need different responses (LL1-16, p. 170; LL1-17, p. 192; LL2-27, p. 656). It became standard in scholarship, not in law; its most influential successor adds ambiguity.
2. The evidence is much stronger on unused knowledge than on genuine ignorance. Ignorance at deployment usually gave way to a longer phase in which relevant knowledge existed but was not generated, assembled or acted on (the editors’ “institutional ignorance” in “most” cases, LL1-16, p. 171). Hindsight moved several “surprises” into that column (MTBE, PCBs).
3. Best evidenced: “no evidence of harm” is produced by the search (not looking, low power, short follow-up, detection limits, narrow endpoints). Fourteen or more cases; strengthened by hindsight. Its missing counterpart: well-designed nulls, followed long enough, can cap risk.
4. Framing decides what can be found: endpoint, scope, admission rules, legacy categories, averages (about 16 cases). Hindsight adds that indicator organisms and reference points can flip verdicts.
5. Monitoring catches what models and inventories miss; self-referential estimates mislead during rapid change. Strongly confirmed after 2001, but monitoring is underpowered, hard to fund and needs triggers.
6. Surprise recurs and is usually found by systems not built to find it. Property screening and monitoring have a track record; diversity is untested; novelty alone predicted poorly; several recent surprises came from remedies.
7. Lay observers often gave the first signal (about 12 cases), best as a detection channel feeding structured inquiry.
8. Weakest: frequency and one-directional bias claims, and chapters that applied these lessons asymmetrically.

2. The reports’ conceptual apparatus#

2.1 Risk, uncertainty, ignorance, and LL2’s additions#

Reports. - LL1 Box 16.1 (LL1-16, p. 170) defines three states: - risk: outcomes and probabilities known. Risk assessment is “valid” here and “can save lives”; - uncertainty: no adequate basis for probabilities. Conventional assessment is “too narrow”; - ignorance: some possible outcomes are themselves unknown, a “continual prospect of surprise” that needs humility and institutional capacity for reflection. - Table 17.1 (LL1-17, p. 192) pairs the states with prevention, “precautionary prevention” and precaution. It dates its examples: asbestos as a risk “1965–present”; antibiotics in feed as uncertainty from 1969; CFCs before 1974 and mesothelioma before 1959 as ignorance. Its precaution row lists strategies that do not require knowing the specific harm: - screening on persistence and bioaccumulation; - the broadest possible sources of information, including long-term monitoring; - robust, diverse and adaptable technologies, with fewer “monopolies”. - Further dimensions. LL1 notes complexity, indeterminacy, ambiguity and “the nature of disagreement” (p. 184) but does not formalise them. - LL2 Table 27.1 (LL2-27, p. 656) adds three states: - ambiguity, answered by “participatory precaution”; - variability; - indeterminacy (unpredictable uses, such as X-rays in shoe shops), answered by pre-market benefit assessment. - Figure 27.1 (p. 654) shows the boundary with ignorance growing as knowledge grows. - The knowledge-to-ignorance ratio (pp. 654–655) is qualitative, high for asbestos, radiation and tobacco and low for emerging technologies. It comes with a rule of thumb: a low ratio calls for precaution plus novel research.

Evidence. - Internal inconsistencies. LL1 pairs uncertainty with “precautionary prevention” and dates asbestos risk from 1965. LL2 pairs uncertainty with “precaution” and dates asbestos from 1930 (context.md §1.4). The 1965 date understates the report’s own record: asbestosis in 1930 and lung cancer in 1955 (LL1-05 Table 5.1, p. 61; LL1-17 notes). - Uptake (LL1-17 hindsight, Claim 2). The typology was taken up in the European Parliament research service’s guidance and in the risk-governance literature. It was not adopted in EU law, and the Commission’s 2000 Communication remains unrevised. Its most influential descendant, Stirling’s four-way scheme (2007, 2010), adds ambiguity. - The knowledge-to-ignorance ratio “proved hard to judge in advance”. LL2 grouped nanotechnology, GM food and mobile phones as low-ratio fields. GM food and mobile phones later moved toward reassurance on human health, and nanotechnology produced specific rather than general hazards (LL2-27 hindsight).

Analysis. Use the four-way form (risk, uncertainty, ambiguity, ignorance), and assign states to sub-questions, not to whole technologies. One technology can sit in “risk” for a known endpoint, “uncertainty” for a second and “ignorance” for unknown failure modes, all at once. The LL1-17 hindsight file makes the same point: knowledge states move, and unevenly within a single issue.

Rating. Strong as a concept. Moderate as an explanation of the historical failures. Suggestive at best for the knowledge-to-ignorance ratio as a predictive tool.

2.2 Institutional versus societal ignorance#

Reports. - LL1-16 (p. 171). - Institutional ignorance: knowledge exists in society but is “not available to the decision-makers”. It is present in “most” cases, and the remedy is communication and social learning. - Societal ignorance: nobody knows. It is present in “many” cases, including BSE, and the remedy is research plus diversity, adaptability and flexibility. - On p. 174 the editors apply “institutional ignorance” to disciplinary capture, where information “was available” but disregarded (MTBE). - LL1 Preface (LL1-00, p. 4). The absence of political will “seems to be an even more important factor” than the availability of trusted information. - LL2 (LL2-00, p. 10). The nine Part A cases show “more than sufficient evidence for much earlier action”. - LL2 Annex 2 (LL2-A2, pp. 705, 709, 712, 714) implicitly separates delays in knowing from delays in responding.

Evidence. - The definition blurs two failure modes. Knowledge that never reached the decision point is one; knowledge that reached it and was discounted is another (LL1-16 notes). - Marchant’s critique (2003). Several flagship cases were ignorance at deployment (mesothelioma, ozone depletion, DES adenocarcinoma), and precaution cannot prevent the genuinely unanticipated. The stronger charge, which he accepts, is slow response once evidence emerged (critiques §4).

Analysis (mine). For use as a lens, split the reports’ two categories into three, because each has a different remedy: 1. Societal ignorance: remedies are broad observation, reversibility and diversity. 2. Distributed knowledge not assembled: remedies are channels, integration and duties to consult. 3. Knowledge delivered but discounted: remedies are evidential standards, independence and accountability.

Rating. Strong for the usefulness of the distinction. Moderate for the reports’ frequency words (“most”, “many”), which are not counted.

2.3 Levels of proof as part of the theory of knowledge#

Reports. Table 16.1 (LL1-16, p. 184) and Table 27.2 (LL2-27, p. 658) run from “scientific suspicion of risk” to “beyond all reasonable doubt”. Sweden’s 1973 law demands strong evidence of safety from manufacturers but only suspicion of risk for regulators to act. The chosen level shifts “the size, nature and distribution of the costs of being wrong” (LL1-17, p. 193). Scientific convention guards against false positives: “not being wrong is more important than being safe” (LL1-16, p. 184). “Not established” rarely says for what purpose, or who bears the error (LL2-27, pp. 657–658).

Evidence. Courts and legislatures now treat the threshold as a choice about who bears the cost of error (Pfizer 2002; EU hazard cut-offs; LL1-17 hindsight). But Table 27.2 mixes legal, statistical and IPCC scales, and neither report gives a method for choosing a rung (LL2-27 digest; LL2-28, p. 676).

Analysis. An evidential standard decides which states of knowledge count as “knowing”, so “we don’t know” is often a statement about the standard as much as the evidence.

Rating. Strong as description; the normative choice is contested.

3. Case map: where the knowledge failures sat#

Compressed from the digests and hindsight files. “First signal” means the first signal that the reports treat as credible. The hindsight column shows where later evidence changed the epistemic picture.

Analysis (mine), based on the table’s 22 rows. - Where first signals came from: people outside formal assessment (workers, inspectors, naturalists, fishers, beekeepers, a mother, clinicians) in about 9 cases; work done for another purpose or by chance in about 7 (PCBs, DES, SO2, TBT, BPA, booster biocides, EE2); targeted theory, research or committees in the rest (ozone mechanism, Swann, the lead warnings, MTBE laboratory studies, BSE, mobile phones). - Two phases. Genuine ignorance at deployment is plausible in roughly half the rows. In nearly every row, the longest delay came after a credible signal. The reports’ evidence therefore bears most on how knowledge is produced, joined up and admitted, and less on ignorance as such.

4. Patterns#

4.1 “No evidence of harm” produced by the search itself#

Reports. - BSE. Reassurance cited absence of evidence “when no evidence was actually being sought”, “a classic example” (LL1-16, p. 172). The national surveillance service was excluded, a key experiment began only in 1996, and there was no random survey (LL1-15, pp. 163–164). - Asbestos. The lung cancer excess became clear only after 25 years of follow-up. “Negative” cancer studies have very low power without 20–30 years (LL1-05, p. 55). - Other cases: no antimicrobial monitoring (LL1-09, pp. 94, 97); Denmark dismissed MTBE because petrol components were “rarely found in groundwater” while nobody monitored for it (LL1-11, pp. 112, 114; the “unmonitored taken as absent” reading is the notes’); no DBCP studies below 5 ppm and no surveillance until workers noticed (LL2-09, pp. 205, 210); beryllium’s lull explained by diagnostic limits (LL2-06, pp. 135, 139–140); detection-based hormone thresholds (LL1-14, pp. 149–150, 152); passive counting at Minamata (LL2-05, pp. 107–111); lead safety asserted, not shown (LL2-03, p. 70); GM assessment (LL2-19, p. 468); “non-positive” read as “negative” (LL2-21, p. 511). - False alarms. About a third of claimed false alarms were cases where “lack of evidence of harm has been misinterpreted as evidence of safety” (LL2-02, p. 21). - Statistical diagnosis: low power and exposure misclassification (LL1-16, p. 184); “the greatest error” is inferring absence of harm from absence of proof (LL2-26, p. 631); treat the upper confidence limit as a plausible worst case and ask “How large an effect can the study have overlooked?” (pp. 633–635); monitoring would miss a 50% decline in most whale stocks and 55% of polar bear and sea otter stocks (p. 634, correctly read); absent Bradford Hill features are not evidence of no causation (LL2-27, pp. 653–654).

Evidence. Fourteen or more cases across both volumes. Hindsight strengthened the pattern: - BSE: EU active testing from 2001 found disease in countries previously thought free, and Germany found BSE only once testing began in November 2000. Back-calculation from testing data found two to four times more infections than earlier estimates (LL1-15 hindsight). - Minamata: the statutory residents’ survey reached only a 32-person pilot by 2026 (LL2-05 hindsight). - Beryllium: surveillance is too sparse to evaluate the 2017 rule (LL2-06 hindsight). - Statistics: the American Statistical Association’s 2016 statement endorsed the critique of significance testing (LL2-26 hindsight).

Cases that cut against. - Mobile phones. The chapter used latency to discount null studies while accepting short-latency positive ones (LL2-21 digest). Later cohorts, national incidence trends (incompatible with a relative risk of 1.08 or more at a 10-year lag) and a WHO-commissioned review capped large risks. The hindsight file concludes that “non-positive is not negative” needs its counterpart (LL2-21 hindsight). - The ratchet. LL1 lifts a restriction only when research “genuinely reveals” a concern is unfounded (p. 173), but keeps one on unresolved uncertainty (p. 181) (LL1-16 notes).

Analysis. The operative variable is the quality of the search, not the presence or absence of findings. Null evidence carries information when there is real exposure contrast, when follow-up is at least as long as plausible latency, and when there is power to detect an effect of the size that matters. The reports supply the right diagnostic (the size of effect that could have been missed) but mostly apply it in one direction.

Rating. Strong. The logic is uncontroversial, the cases are many, and hindsight confirmed them. The symmetric corollary, that well-designed nulls can bound risk, is also strong on hindsight.

4.2 Framing and bounding: the question asked decides what can be found#

Reports. Five sub-mechanisms recur.

1. The first visible harm sets the endpoint. Radiation limits targeted acute effects, with “no realisation” of latent cancer (LL1-03, p. 33); clinicians’ acute focus shaped asbestos and radiation standards (LL1-16, p. 174); TBT’s first UK target of 20 ng/l rested on acute tests while dogwhelks were sterilised at 3–5 ng/l (LL1-13, pp. 136–137, 141); control of acute beryllium disease masked the chronic disease (LL2-06, pp. 133–134); acute and mortality endpoints misled on lead (LL2-03, pp. 52, 59–60). Table 26.3 lists nine “erroneous assumptions”, including “acute effects reflect chronic” (LL2-26, p. 630). Absence of acute signs is no reassurance (BSE, halocarbons; LL1-16, p. 171).
2. Scope set for assessors. JECFA assessed only authorised use, single substances and manufacturers’ data (LL1-14, p. 150). A French committee asked whether Gaucho was “solely responsible, at national level, for all” bee losses, “a question that had never been asked”, using spray-era tools on systemic seed-dressings (LL2-16, pp. 375, 377, 379). Probabilistic safety assessment depends on the scenarios analysts list and on independence assumptions; security was excluded from stress tests; a 2001 paper on a roughly 1,000-year tsunami never reached the design basis (LL2-18, pp. 438, 444, 447–448). Narrow methods produce “effective denial” of risks outside the frame (LL2-28, p. 677).
3. Evidence-admission rules. Excluding non-guideline academic studies moved BPA “safe” levels by orders of magnitude (LL2-10, pp. 220–223). 29 TCE assessments reached four conclusion types, and IARC and ECETOC weighted the same PCE studies differently (Rudén, LL2-04, pp. 84–85). The same MTBE data produced four verdicts (LL1-11, p. 113).
4. Legacy categories. Chemical identifiers and tonnage thresholds made nano forms legally invisible (LL2-22, pp. 537–541); MTBE predated the new-substances cut-off (LL1-11, p. 116); BPA’s known oestrogenicity did not follow it into materials use (LL2-10, pp. 216–217).
5. Averages and reference subjects. Low-oestrogen children were missed (LL1-14, pp. 150, 152–153); neonates were least visible (LL2-10, p. 224); evidence is thinnest where exposure is highest (LL2-11, pp. 247–248); averages dilute harm to vulnerable groups (LL2-26, pp. 638–639).

Evidence. About 16 cases, and hindsight strengthened most of them. - Hormones: the WTO Appellate Body made the assessor-scope point (LL1-14 digest). - Neonicotinoids: EFSA, courts and auditors confirmed the method critique. The 2013 bee guidance was still not implemented in 2026 (LL2-16 hindsight). - BPA: EFSA’s 2023 intake limit rested on an academic, non-guideline study. But the jointly designed CLARITY-BPA reproduced the guideline-versus-academic split, which points to admission rules and “different languages” more than to funding (LL2-10 hindsight). - PCE: classification is still split across five bodies (LL2-04 hindsight). - Probabilistic safety assessment: broadened, not abandoned (LL2-18 hindsight).

Hindsight adds two refinements: - A right indicator with a wrong mechanism. TBT imposex established causation although the steroid mechanism was wrong (RXR activation replaced it), and the wrong mechanism hid vertebrate effects (LL1-13 hindsight). - The choice of indicator organism can decide the verdict. Honeybee colonies were buffered and under-reported harm that fell on wild bees (LL2-16 hindsight). OSPAR’s TBT status depends on which whelk is monitored (LL1-13 hindsight).

Cases that cut against. - Framing can produce false alarms too. The swine-flu warning was over-weighted because it fitted prevailing theory: “Perhaps too much faith was placed on the ability of science to foresee” (LL2-02, p. 31). - The reports show framing effects in their own chapters. The GM chapter held GM claims to strict standards and agroecology claims to lax ones (LL2-19 digest).

Analysis. This is the most transferable mechanism in the theme, because it operates before any evidence is gathered. The recurring structure is an appraisal system built for one mode of harm or exposure, applied to a new mode, whose null result is then read as safety.

Rating. Strong.

4.3 Disciplinary and institutional silos#

Reports. Lesson 4 holds that the discipline whose effects appear first can hold appraisal “captive”, producing institutional ignorance (LL1-16, p. 174). - Air versus water. MTBE was appraised through engine and air-pollution knowledge; water persistence was “essentially disregarded, though the information was available” (LL1-16, p. 174; LL1-11, p. 114). - Veterinary versus human. - A feed additive was assessed by an animal-nutrition committee, and Swann’s single cross-sector committee was never set up (LL1-09, pp. 93–96). - BSE was framed as a veterinary issue, and MAFF told the health department only after 17 months (LL1-15, pp. 159–160). - Drug regulators and the environment. Drug agencies had little interest in environmental effects (hormones; LL1-14, p. 152) and “limited expertise in environmental issues” (EE2; LL2-13, p. 284). - Upstream knowledge, downstream decisions. PCE toxicology “never figured in the water mains product design” (LL2-04, pp. 82, 84). DBCP knowledge did not reach user companies (LL2-09, p. 211). - Pieces of one anomaly. British, Japanese and NASA groups each held part of the ozone anomaly (LL1-07, p. 82). - Other examples: “political, disciplinary and geographic silos” in ecosystem management (LL2-17, pp. 407, 415); an environmental agency “rooted in chemistry” facing nanomaterials (LL2-22, p. 543); invasive-species turf disputes (LL2-20, pp. 496–497); the call to integrate health and environment (LL1-17, p. 193).

Evidence. About 13 cases. The illustrations are clear, but most are second-hand in the synthesis chapters. Hindsight adds three things: - MTBE: litigation records show US regulators and industry flagged the groundwater risk in 1984–88. The failure was institutional, not societal (LL1-11 hindsight). - SO2: the largest benefit (avoided fine-particle deaths) and the largest unintended cost (warming unmasked by removing sulphate) both came through channels outside the frame of the original decision (LL1-10 hindsight). - Integration works. EU integrated surveillance had quantified animal–human resistance links for some combinations by 2024 (LL1-17 hindsight).

Analysis. Silos hide effects on both sides of the ledger: unexpected harms and unexpected benefits. The SO2 hindsight is the clearest example. The causal weight of silos is hard to separate from the weight of interests, because a silo is often where an interest sits.

Rating. Moderate to strong.

4.4 Models versus monitoring#

Reports: how models misled. - Fisheries. Stock estimates converge on the truth only after about five more years of data, giving “a false sense of assurance” during decline; Harris described “rose tinted glasses” (LL1-02, p. 21). The Canadian model was tuned to offshore landings, held recruitment and natural mortality constant, and excluded inshore data as “messy and often anecdotal” (LL2-17, pp. 411, 414); Templeman had warned that catch rates can rise as a stock collapses (p. 413). A multi-species model grew less predictable as data were added (LL1-16, pp. 172–173). - Ozone. Theory expected the first effects high in the tropics, models and observations agreed that little was happening, and software flagged very low values as “suspect” (LL1-07, pp. 80, 82). - Floods. Design floods rest on short records and codes assume “the past is the key to the future” (LL2-15, pp. 351–352, 355). Beven: treating gaps in knowledge as random variability “will produce over-confidence” (pp. 357–358). - Climate and radiation. Expert groups underestimated structural uncertainty (LL2-14, p. 333). Radiation risk rested on one atypical high-dose cohort (LL1-03, p. 35); one canonical dataset can anchor regulation for decades (LL2-A3, p. 721).

Reports: how monitoring found surprises, and its limits. The long Halley Bay record (LL1-07, p. 82); a Swedish precipitation network built to study plant nutrients (LL1-10, p. 102); Keeling’s CO2 record (LL2-17, p. 410); Great Lakes indicator organisms showing success and stalling (LL1-12, p. 129); TBT baselines (LL1-13, p. 142). Lesson 2 (LL1-16, pp. 171–173) and the call to fund surveillance “even when an immediate need is not perceived” (LL1-03, p. 36) follow. The reports also state the limits: monitoring is hard to fund (LL1-07, p. 82); results must be used (LL1-16, p. 172); without action thresholds it becomes an “academic pursuit” (LL2-12, p. 274); it is often underpowered (LL2-26, p. 634); it fails in the extremes it exists to observe (LL2-15, pp. 359–360); shared monitoring did not end resistance for years after the 1977 attribution (LL1-10, pp. 103, 105).

Evidence (hindsight). Both halves were strengthened. - Estimates still overshoot. An ICES workshop found 15–20% of age-structured assessments show significant retrospective patterns, mostly overestimates; the 2025 North Sea cod assessment downscaled the previous year’s spawning stock by 35% (LL1-02 hindsight). Northern cod reached “Healthy” status in 2026 partly through a lowered limit reference point: in DFO’s words, “a downward revision of the LRP, not an increase in the quantity of cod” (LL1-02 hindsight). Reference-point choices can redefine harm or recovery (LL2-A2 hindsight). - Monitoring caught what inventories missed: unreported CFC-11 production (2018), PFOS in wildlife worldwide (2001) (LL1-17 hindsight), HFC-23 emissions above reported levels (LL2-14 hindsight). The 2021 Ahr peaks matched floods of 1804 and 1910 left out of the hazard assessment (LL2-15 hindsight). - Observation is fragile: ground-based ozone instruments are no longer manufactured (LL1-07 hindsight); surveillance units closed in quiet periods (LL1-03 hindsight); observation systems falter under geopolitical rupture (LL2-17 hindsight).

Cases that cut against. Model-based rules rebuilt many stocks, including Barents Sea cod after 1990, so LL1-02’s pessimism was too general (LL1-02 hindsight). Critical loads, a model-based intermediate threshold, made acid-rain action tractable (LL1-10, pp. 106–107), and exceedance fell to 3.5% by 2024 (LL1-10 hindsight). And independent is not automatically right: the independent Fukushima caesium estimate the chapter relied on was about twice the later consensus, while the official figure fell inside it (LL2-18 hindsight).

Analysis. The real contrast is between knowledge systems that can register an unexpected signal and those that filter it through prior expectation. The problem is not models as such but models calibrated to the activity’s own outputs, key variables held constant, anomalies screened against expected ranges, and reference points revised without anything changing in the world.

Rating. Strong that monitoring detects surprises models and inventories miss. Strong that self-referential or hindsight-convergent estimates mislead most during rapid change. Moderate for the reports’ broader scepticism about models.

4.5 Measurement capability sets the horizon of knowledge#

Reports. - Detection limits and analytical methods. - The cause of TBT imposex stayed unknown for about a decade, until analytical chemistry improved. Ignorance of the problem’s extent also reflected missing baselines (LL1-13, pp. 136, 142). - Great Lakes contaminants appeared as methods arrived: PCBs after 1966, hexachlorobenzene in 1972, mirex in 1974 (LL1-12, p. 126). - Optical microscopy could not track asbestos fibrils, and the public limit was pegged to its detection limit (LL1-05, pp. 56–57). Cheap instruments shaped both the evidence and the metric of limits (LL2-A3, p. 725). - Early Gaucho studies could not detect residues below 10 ppb, and public researchers were told not to go below that figure (LL2-16, p. 373). - Routine water tests could not see organic solvents (LL2-04, p. 77). - A diagnostic test transformed the count of beryllium cases (LL2-06, pp. 135, 137). - Reliable CO2 trends came only in the 1960s, and Callendar’s data were distrusted (LL2-14, p. 311). - EE2 had to be measured at parts per trillion (LL2-13, pp. 284, 292, 295). - Convenient proxies used as safety claims. - DBCP’s smell (1.7 ppm, above the 1 ppm limit itself) was accepted as a warning (LL2-09, p. 205). - Vinyl chloride’s odour threshold was 4,000 ppm, not 400 (LL2-08, p. 184). - “Safe” lead levels fell as measurement improved (LL2-03, pp. 61–62).

Evidence. About 13 cases. Hindsight: - Better measurement brought tighter limits. The EU cut its asbestos limit 10–50-fold and moved to electron microscopy (LL2-A3 hindsight). The EU mandated effect-based oestrogen monitoring because chemical measurement at the standard is still extremely hard (LL2-13 hindsight). - Better measurement can also move status without new evidence of harm. New analytical methods triggered the US DES withdrawals of 1972–73, while the FDA disclaimed any health hazard (LL1-14 digest). - Apparent “harm expansion” partly follows where research and detection go (LL2-A3, LL2-28 hindsight). - Limits can rise with method changes. EFSA raised its nickel intake limit in 2020 after a revised benchmark-dose analysis (LL2-26 hindsight).

Analysis. Measurement capability is a hidden parameter in every “no effect” finding. Proxies chosen for convenience, such as odour or visible fibres, can quietly become safety claims.

Rating. Strong.

4.6 Long latency, delayed effects and developmental windows#

Reports. - Asbestos. Harm peaks 50–60 years after peak imports (LL1-05, p. 52). The “latency lacuna” means that by the time evidence on today’s exposure arrives, conditions have changed again (p. 55). Claims since 1906 that the problem was “already solved” kept resetting the clock (LL1-16, pp. 173–174). - Across generations. Latent radiation cancers (LL1-03, p. 33); DES effects in the next generation (LL1-08, pp. 84–88); in-utero PCB effects (LL1-12, pp. 127, 131). - Timing as well as dose. It matters for PCBs (LL1-06, p. 67), and “it seems clear that it is more the timing of the dose, rather than the dose itself” that distinguishes harmful exposure (LL2-27, p. 650). - Delayed physical effects. Skin-cancer excess from ozone loss peaks decades after peak production (LL1-07, p. 76). Less than two-thirds of committed warming was visible (LL2-14, p. 314). - Epidemiology. Latency and transient exposure defeat it for BPA (LL2-10, p. 219). - Invasive species. Lag phases lasting decades erase traceability and liability (LL2-20, p. 497). - Exposure ahead of evidence. For mobile phones, fewer than 10% of Interphone cases had 10 or more years of use (LL2-21, pp. 512, 517). LL2 generalises this as hazards “largely unknown, yet already widespread” (LL2-00, p. 10). - Accountability. Latency plus firms that disappear defeats compensation (LL2-24, pp. 592, 596).

Evidence. Twelve or more cases. Hindsight strengthened the core: - UK male mesothelioma deaths peaked in 2016. The timing was right, though the chapter overstated the peak by 20–35% (LL1-05 hindsight). - Low-dose worker and CT cohorts show radiation risk per unit dose at least as high as the bomb-survivor estimates (LL1-03 hindsight). - Late DES outcomes continued to appear (LL1-08 hindsight). - Appendix surveys found vCJD prion carriage in about 1 in 2,000 people (LL1-15 hindsight). - PCBs still drive most Great Lakes fish advisories (LL1-12 hindsight).

Cases that cut against. - Follow-up cuts both ways. For DES, breast cancer in daughters is now contested, there is no overall cancer excess, and third-generation effects remain unresolved after 25 years. “Harm only expands” needs qualifying (LL1-08 hindsight). - Long follow-up can exonerate. For mobile phones it capped large risks (LL2-21 hindsight).

Analysis. Latency changes the epistemics in three ways: 1. early nulls become uninformative; 2. deployment outruns evidence, so exposure is universal before evidence matures; 3. each incremental improvement creates a new, as yet unevaluable, exposure regime.

Latency is a reason to discount early nulls. It is not a reason to discount later nulls from studies with adequate follow-up. The mobile-phone chapter blurred that line.

Rating. Strong.

4.7 Surprise, and how to act under ignorance#

Reports. - The claim. “Surprise is inevitable”. Every activity had some risk assessment, but the “virtual certainty” that some factors lay outside its scope was neglected (LL1-16, p. 169). - Examples: CFCs in the stratosphere; DES next-generation effects, “a complete surprise” (LL1-16, p. 170); TBT accumulation in top predators, “simply not envisaged” (LL1-13, p. 141); mesothelioma (LL1-17, p. 192; LL2-27, p. 654); PCBs in the environment, “much of a surprise” because they were used in contained systems (LL2-11, p. 242). - Named sources of surprise: the inertness short-term safety seems to demand means persistence, a “deep-seated paradox” (LL1-07, p. 83); behaviour under conditions never considered (LL1-16, p. 169); environmental transformation (LL1-06, p. 67); interactions (LL2-04, p. 81; LL1-05, p. 55); global scale, “only one ‘experimental’ model” (LL1-16, p. 171); unpredictable uses (LL2-27, p. 656); cascading common-cause failure (LL2-18, pp. 447–448). - Detection was often by chance or through other work: PCBs found while analysing DDT (LL1-06, p. 64); the DES cluster, a “fortuitous accident” (LL1-08, p. 86); BPA traced from contaminated labware (LL2-10, p. 217); Irgarol found by a herbicide survey (LL2-12, p. 267); intersex roach found in a routine check (LL2-13, p. 282); acidification found through a nutrient network (LL1-10, p. 102). For ozone, the editors say “essentially serendipitous” (LL1-16, pp. 172–173); Farman stresses systematic long-term measurement (LL1-07, p. 82). - Proposed responses: the precaution row of Table 17.1; proxies for unknown harm such as irreversibility, novelty, persistence, dispersal and scale (LL1-16, pp. 170–171); twelve unweighted “criteria for precautionary action” (LL2-27, p. 653); recognise much sooner “when ignorance has been replaced by understanding, however rudimentary” (LL1-07, p. 83); extend responsibility to unknowns “predictable in principle even though not in specifics” (LL1-16, p. 189). - Farman’s own candour. A conventional 1965 assessment would have found “no known grounds for concern”. It would have noted that nobody knew what happens to CFCs once released, then discounted that gap against 30 years without apparent harm (LL1-07, p. 82).

Evidence (hindsight). - Property-based screening was widely institutionalised (LL1-17 hindsight): - the Stockholm Convention (37 listed entries); - EU bars on persistent, bioaccumulative and toxic pesticides; - the 2023 hazard classes for persistent, mobile and toxic substances. - The Convention’s effectiveness evaluation reports falling levels of regulated substances. Its limits: slow listing, exemptions, regrettable substitution and no compliance mechanism. - Monitoring detected surprises after 2001 (see 4.4). - Technological diversity has not been adopted by any regulatory regime and is untested (LL1-17 hindsight). - Novelty alone was a weak predictor. Persistence, bioaccumulation, irreversibility and wide dispersal did better at picking cases that later warranted action (LL2-27 hindsight). - Several post-2001 surprises came from remedies: - HFO refrigerants degrade to persistent TFA (LL2-A2 hindsight); - SO2 cuts unmasked warming (LL1-10 hindsight); - evacuation after Fukushima caused 2,351 disaster-related deaths, while UNSCEAR finds no documented radiation-caused disease (LL2-18 hindsight); - excluding pregnant women from research became a cause of the ignorance the DES chapter deplored (LL1-08 hindsight); - invasive mussels remobilised PCBs in Lake Ontario (LL1-12 hindsight).

Critique. Marchant argues that precaution cannot prevent the genuinely unanticipated. Wynne’s position is compatible with this: ignorance argues for monitoring, diversity and reversibility more than for prohibition in advance (critiques §4, §8).

Analysis. 1. Surprises were usually registered by systems not designed to find them. Breadth of observation matters as much as targeted testing. 2. Several labelled “surprises” (DES, MTBE, PCBs) turn out on inspection to be knowledge that existed but was never assembled. “Surprise” is partly a statement about institutions. 3. Because important recent surprises came from interventions, responses to ignorance must be applied to remedies, substitutes and protective actions too.

Rating. - Strong that surprise recurs and is often found through other work: about 9 cases, plus post-2001 confirmations. - Moderate for property-based screening as a response: strong within persistent chemicals, untested beyond. - Suggestive for diversity as insurance. - Asserted for enlarged responsibility.

4.8 Sentinels and signature outcomes#

Reports. - Wildlife as early warning. Wildlife harm can act as a “sentinel” for human hazards (LL1-16, p. 171), though Ch 16 names no case. The case chapters supply them: - cormorant eggs, ranch mink and gulls in the Great Lakes (LL1-12, p. 126); - cats and fish die-offs at Minamata (LL2-05, pp. 96–97); - birds rather than seals, because only fishermen watched seals and they regarded them as pests (LL2-11, pp. 242–243); - imposex as a specific biomarker (LL1-13, pp. 135, 142). - Distinctive outcomes speed recognition. - A rare cancer identified DES (7 of 8 cases against 0 of 32 controls; LL1-08, p. 84). - The rarity of angiosarcoma “made causation undeniable” for vinyl chloride (LL2-08, p. 189). - Just 4–7 cases of very rare cancers or of sperm reduction triggered prompt action (DES, vinyl chloride, DBCP), against more than 40 years for tobacco (LL2-27, pp. 645, 647). - The converse. - Harms that add to common diseases may never be settled (LL2-08, p. 189). - The common DES reproductive harms were found only because the rare cancer triggered follow-up (LL1-08, pp. 86–87). - Harm is noticed where observers value what is harmed: the oyster crop, not the pest snail (LL1-13, p. 136; LL2-12, p. 267).

Evidence. - The DES pattern of rare cancer against common harm is now well quantified (LL1-08 hindsight). - TBT imposex worked despite the wrong mechanism (LL1-13 hindsight). - The honeybee inverted as a sentinel: the managed, economically visible species was among the least informative at population level (LL2-16 hindsight).

Analysis. Detection is biased toward outcomes that are distinctive, rare, acute or commercially visible. Diffuse increments to common conditions, and harms to parts of a system nobody values commercially, are the default blind spot. Choosing a sentinel is a design decision, not a given.

Rating. Strong for the signature-outcome effect (five or more cases). Moderate for wildlife as sentinel (several cases, but selected, and hindsight shows a sentinel can mislead).

4.9 Lay, local and frontline knowledge#

Reports. Lesson 8 asks that lay and local knowledge be used, with the same scrutiny as expertise (LL1-16, pp. 177–178). - Examples: - asbestos factory inspectors and workers (LL1-05, p. 53); - the inshore fishers’ Keats report, dismissed as “biased pseudoscience” (LL1-02, p. 21; LL2-17, p. 412); - DBCP workers who compared notes at lunch (LL2-09, p. 204); - the Minamata mother who recognised congenital disease while experts believed the placenta protected the foetus (LL2-05, pp. 105–106); - beekeepers (LL2-16, pp. 372–373); - Love Canal residents (LL1-12, p. 127); - Swedish farmers (LL1-09, p. 95); - abattoir workers “apparently better informed” than officials (LL1-16, p. 178); - warners including “clinicians, factory inspectors, workers, anglers, bee keepers and community members” (LL2-26, p. 631); - harmed third parties as early warners (LL2-25, p. 609). - Admission rules shut such signals out. Inshore data were excluded as “messy and often anecdotal” (LL2-17, p. 414). - The reports’ own counterweights: - the “pensioners’ party fallacy”, where healthy retirees were read as proof of safety (LL1-05, p. 60; LL1-16, p. 178); - fishers can be less precautionary, and dissent tracks interest (LL1-02, pp. 20–21; LL1-16, p. 177). - The stronger claim. Publics make “a basically correct distinction between uncertainty and ignorance” (LL1-16, p. 185). It rests on two GMO focus-group studies, one first-authored by an editor (LL1-16 notes).

Evidence. - About 12 supporting cases, selected because the warnings proved right. - Counter-cases inside the reports: the pensioners, offshore fishers denying decline (LL1-02 notes), and Harada’s own initial dismissal of the Minamata mother (LL2-28 notes). - Hindsight: - the Minamata residents’ account was vindicated, but official counting is still passive (LL2-05); - the practitioners’ indicator species was a poor sentinel (LL2-16); - a lay-amplified alarm (MMR) caused lasting harm (LL2-02 hindsight); - public intuition is credited in the reports only where it proved right (LL1-16, p. 178, against p. 188 on irradiated food, which expert bodies later judged safe).

Analysis. The evidence supports lay knowledge as a detection channel: early, and grounded in real operating conditions. It supports it less as a validation channel. In the reports’ best cases, lay signals worked when turned into structured inquiry: - the Keats report worked through a commissioned re-analysis of official data (LL1-02 digest); - DBCP workers’ suspicion led to union sperm counts and then a NIOSH study (LL2-09, pp. 204, 206); - the Minamata signals led to epidemiology (LL2-05, p. 98).

The characteristic failure is an institutional admission rule, not a lack of lay insight.

Rating. Moderate for the detection role. Suggestive for claims that publics understand uncertainty better than institutions.

4.10 When research deepens, or is used to sustain, uncertainty#

Reports. - More research can “compound uncertainty”; Great Lakes research “amplified the uncertainties” (LL1-16, pp. 172–173). Broadening a cause-to-effects appraisal widens the net, but broadening an effects-to-causes inquiry can weaken the case against any one agent (p. 181). Complexity framings were “not inconvenient” to those resisting costs (LL1-12, p. 129), and multicausality is used to excuse inaction (LL2-16, p. 391); “the absence of consistency between studies does not imply an absence of causality” (LL2-27, pp. 652–654). - Research inertia. Thirteen data-poor chemicals on an EPA priority list had 352 article links in a decade, far fewer than lead (LL2-26, pp. 626–629). Hazard research is a small share of development funding, given inconsistently as “3 %” (LL2-27, p. 646) and “1%” (LL2-28, p. 679), both unsourced. - Deliberate uncertainty. Research as the “antidote” and teams to “keep the controversy alive” (tobacco; LL2-07, p. 154); research refused so industry could “act as if the cause was ‘unknown’” (vinyl chloride; LL2-08, p. 184); shifting rationales with a fixed conclusion (beryllium; LL2-06, pp. 137–138); public calls for “extensive research” alongside private acknowledgement (PCBs; LL1-06, p. 65); “manufactured uncertainty” (LL2-25, p. 610). - Delay dressed as diligence. Researchers’ appetite for data can align with polluters’ interest in delay (Grandjean, LL2-05, pp. 122–123); nobody generated the decisive evidence after divergent decisions (hormones; LL1-14, p. 153); “more information as a substitute for action” (LL2-22, p. 547). The counter-model is Swann: “the cry for more research should not be allowed to hold up our recommendations” (LL1-16, p. 181).

Evidence. Documentary evidence of deliberate doubt-production grew after 2013: US court findings, PFAS industry documents showing knowledge by 1970, and cross-sector studies (LL2-07, LL1-17 hindsight). Research attention can shift fast once controversy arises (LL2-26 hindsight). Of more than 10,000 nanosafety papers, most “do not offer any kind of clear statement” on safety, so low knowledge was partly a problem of quality, not quantity (LL2-27 hindsight).

Analysis. Ignorance is partly produced, by three routes: not doing the decisive study, posing questions that cannot be answered, and generating noise. The reports’ most practical answer is the Swann procedure: state the research question, its duration, funder and independence, and whether action waits for it (LL1-16, p. 173).

Rating. Strong for deliberate cases with internal documents (tobacco, vinyl chloride, beryllium, PCBs). Moderate for research inertia (one author’s bibliometrics; counts, not content). Moderate for “more research can increase uncertainty” (two cases, sound logic).

4.11 Certainty language and the communication of uncertainty#

Reports. - The general charge. “Misplaced ‘certainty’ about the absence of harm” (LL1-00, p. 4; LL2-00, p. 6). - BSE. The expert committee advised that “no risk” could not be stated categorically. A month later the minister told Parliament beef was “perfectly safe”. Categorical reassurance then made further precaution self-incriminating (LL1-15, pp. 161–162). - Fisheries. Scientists gave precise low numbers to avoid ranges being exploited, producing false precision (LL1-02, p. 22). - Other examples: - DBCP safety rested on “authoritative assertion but without evidence” (LL2-09, p. 211); - nuclear regulation speaks a “language of certainty” (LL2-18, p. 448); - a divided consortium’s “oracular” conclusion was reported in opposite ways on the same day (LL2-21, pp. 517–518; LL2-27, p. 657); - calibrated vocabularies can create false precision (LL2-14, pp. 332–334); - science is provisional among peers but expected to deliver certainty to outsiders (LL1-16, p. 185); - hedged wording is exploited by selective quotation (LL2-26, p. 637); - false reassurance costs trust (LL2-27, p. 658).

Evidence. - The BSE record was strengthened (LL1-15 hindsight). The later, more open regime also used candour about uncertainty to relax a measure: the Over Thirty Months rule was replaced by testing after a review put its cost at about £2bn per death prevented. - Claims of collapsing trust in scientists were weakened. EU data did not show “very low” trust even in 2001, and a 2025 68-country study found no widespread lack of trust (LL1-00, LL2-00 hindsight).

Analysis. The failure is not uncertainty itself. It is converting a conditional scientific judgement into an unconditional public claim, which then constrains later learning and makes graded responses look like admissions of error.

Rating. Strong for the BSE mechanism. Moderate as a generalisation. Weakened for the trust-collapse claims.

5. Counter-evidence, complications and critiques#

1. Selection and hindsight. All LL1 cases are confirmed harms, chosen “where sufficient is now known” (LL1-00, p. 11), mostly by participant authors (p. 12). The reports can show that signals were neglected, not how often heeding signals of comparable strength would have been right (critiques §4). Marchant adds an asymmetry of verification: false positives stay provisional, false negatives become settled.
2. “Surprise” and “ignored warning” are coded inconsistently, in both directions. DES: the editors call next-generation effects “a complete surprise” (LL1-16, p. 170); the chapter says warnings were ignored (LL1-08, pp. 88, 90). Ozone: the editors say “serendipitous” and “set aside”, Farman “systematic” and “flagged”; hindsight sides with Farman (LL1-07 hindsight). MTBE “was never considered” (p. 117), yet regulators flagged it in 1984–88 (LL1-11 hindsight). The PCB evidence of 1937 was published and shared (LL1-06 hindsight). Whether pre-1930 asbestos warnings were actionable is disputed (Bartrip; Deane, p. 60). Date each foreseeability claim separately.
3. The absence-of-evidence argument is applied asymmetrically. LL2 allows action on “weak” evidence (10–33%, Table 27.2) but counts a false positive only at “high confidence” (67–95%) of no harm (LL2-02, p. 18; critiques §5.3). Restrictions are lifted only when research “genuinely reveals” a concern is unfounded (LL1-16, pp. 173, 181). The mobile-phone chapter used latency selectively (LL2-21 digest). The GM chapter concedes “it is plausible that there simply are no effects to be found” (LL2-19, p. 468), yet the conclusion asserts a health threat (LL2-28, p. 674), a claim later assessments weakened.
4. One-directional bias claims. Table 26.4 lists ten design features biased toward false negatives against three toward false positives (LL2-26, p. 635); Ch 27 claims “systemic biases towards not finding a causal link” (LL2-27, p. 654). The replication crisis, the “decline effect” in ocean-acidification studies of fish behaviour, and recall-bias findings show false positives are common in published research (LL2-26, LL2-27 hindsight). “Nearly all” limits falling has counter-examples (nickel 2020, perchlorate 2026).
5. “Harm expansion” is partly a selection and detection effect, with counter-cases (MTBE, fish, hormone residues) (LL2-A3, LL2-28 hindsight).
6. “Know more” against “don’t wait”. The editors acknowledge the tension (LL1-16, p. 181) but give no rule for when enough is known; Box 27.4 is unweighted (LL2-27, p. 653); LL2-28 names threshold factors without a method (p. 676).
7. Responses to ignorance have costs and can create their own ignorance (exclusion from research, evacuation harms, unmasked warming, regrettable substitutes; 4.7). The reports rarely apply their epistemic lessons to precautionary actions themselves. Lay knowledge and public intuition are likewise credited selectively (4.9).
8. Independence is not accuracy. Independent Fukushima release estimates overshot (LL2-18 hindsight); the Yablokov Chernobyl figure should not be relied on; Hardell’s odds ratios are incompatible with incidence data (LL2-21 hindsight).
9. Typology gaps. LL1 lacks ambiguity; the knowledge-to-ignorance ratio grouped fields that later diverged; diversity as insurance is untested.
10. The chapters show the failings they diagnose. Ch 26 misstates confidence intervals and misreads the whale and polar-bear data (LL2-26 digest); Ch 18’s accident arithmetic rests on a non-peer-reviewed web source; Ch 21 is written by protagonists. This does not refute the lessons; it shows they apply to the reports too.

6. Overall weighting for use as a lens#

7. Technology-neutral diagnostic questions#

Each question can be asked of any emerging technology and of the people developing it, and each is tied to the evidence above.

1. What state of knowledge applies to each question, as of when and on whose assessment? For each claimed harm pathway, is it risk, uncertainty, ambiguity or ignorance? Is the classification dated and attached to a sub-question rather than to the whole technology? (2.1; LL1-17, p. 192; LL2-27, p. 656; LL1-17 hindsight)
2. What would the reassuring evidence actually have ruled out? Is there real exposure contrast, follow-up as long as plausible latency, and power for the effect size that matters? What is the upper confidence bound? (4.1, 4.6; LL1-05, p. 55; LL2-26, pp. 633–635; LL2-21 hindsight)
3. Is anyone looking where the harm would show? Which questions critical to the current reassurance are unresearched? Who would fund them, when would they report, and does action wait for them? (LL1-16, pp. 172–173; LL1-15, pp. 163–164; LL1-15 hindsight on post-2001 testing)
4. For what mode of harm were the assessment tools built? Does this technology introduce a mode of exposure, failure or interaction those tools cannot see? Is the assessment question answerable? (4.2; LL2-16, pp. 375–379; LL1-13, pp. 136, 141; LL2-18, pp. 438, 447–448)
5. Which rules decide what counts as admissible evidence, and how far would the answer move under the alternatives? (LL2-10, pp. 220–223; LL2-04, pp. 84–85; LL1-11, p. 113)
6. Do existing categories, identifiers or thresholds make new variants invisible to oversight? (LL2-22, pp. 537–541; LL1-11, p. 116; LL2-10, pp. 216–217)
7. Who is not represented by the average or the reference subject? Where is exposure highest and evidence thinnest? (LL1-14, pp. 150–153; LL2-26, pp. 630, 638–639; LL2-11, pp. 247–248)
8. Where does relevant knowledge already sit, and does it reach the decision point? Consider other disciplines, upstream suppliers, downstream users, other agencies and insiders’ internal records. (4.3; LL2-04, pp. 82–84; LL1-15, pp. 159–160; LL1-07, p. 82; LL2-08, pp. 183–186)
9. Are the indicators of safety or success independent of the activity being judged? Could estimates be tuned to the activity’s own outputs, with key variables held constant? Could a reference point be revised so that status improves without any change in the world? (4.4; LL1-02, p. 21; LL2-17, pp. 411–414; LL1-02 and LL2-A2 hindsight)
10. Is there broad, long-term, independently run observation able to register the unexpected? Does it have pre-agreed triggers for action and adequate power? How will it survive quiet periods? (LL1-16, pp. 171–172; LL2-12, p. 274; LL2-17, p. 423; LL2-26, p. 634; LL1-03, p. 36 and hindsight)
11. Are frontline observers’ reports admitted? Is there a route from anecdote to structured investigation, and a check against their own fallacies? (4.9; LL2-17, p. 414; LL2-09, p. 204; LL1-02, p. 21; LL1-05, p. 60)
12. Which intrinsic properties would make being wrong costly whatever the specific harm? Consider persistence, accumulation, mobility, irreversibility, dispersal and scale. Is novelty being used as the signal instead? (LL1-16, pp. 170–171; LL2-27, p. 653; LL1-17 and LL2-27 hindsight)
13. Which harms would be distinctive enough to be noticed, and which would disappear into common conditions or fall on things nobody monitors? Is the chosen sentinel informative? (4.8; LL2-08, p. 189; LL1-08, pp. 86–87; LL2-16 hindsight)
14. How is uncertainty communicated, and does any public claim of safety make later caution look like an admission of error? (4.11; LL1-15, pp. 161–162; LL2-18, p. 448; LL2-27, pp. 657–658)
15. Would the same epistemic standards be applied if the evidence pointed the other way? And what surprises could the proposed remedy, substitute or protective action create? (Section 5; LL1-16, pp. 173, 181; LL2-21 hindsight; LL1-10, LL1-08 and LL2-18 hindsight; LL2-12, p. 273)