T07: Complexity, systems and scale across the two Late Lessons reports#

Strand A thematic synthesis, drawing on the section digests, notes and hindsight files for LL1-00 to LL2-A3, plus external/critiques.md and external/context.md (details in §7).

Page references. “LL1” pages are report pages of EEA Environmental Issue Report 22 (2001). “LL2” pages are report pages of EEA Report 1/2013. Section IDs follow the project convention (e.g. LL2-17 = LL2 Chapter 17).

Three voices, kept apart: - Reports: what the chapters, panels and editors say. - Hindsight: what the project’s post-publication checks (to September 2026) found. These come from the hindsight/ files unless stated. - Assessment: my own analysis. It is labelled as such.

Evidence ratings. - Strong: documented in several cases, with independent corroboration or robust confirmation in hindsight. - Moderate: documented in one to a few cases, or with significant caveats. - Suggestive: illustrated, but inferential, single-case or contested. - Asserted: stated without case evidence, or a normative position.

Case counts are counts of report sections. They are not independent observations: many chapters share authors, editors and professional networks, and every historical case was chosen because harm occurred.

1. Main findings in brief#

1. The most robust content is a set of mechanisms, not complexity theory. These hold across many cases and survive hindsight well: - persistent agents deployed at scale create stocks that outlast control; - fixes that disperse harm relocate it instead of reducing it; - assessments built on acute, high-dose, adult or average endpoints miss chronic, developmental and system-level effects; - product-by-product regulation produces substitution treadmills.
2. “Deployment outruns knowledge” is well documented for the historical persistent chemicals. It holds for CFCs, PCBs, lead, MTBE, asbestos and BPA. The reports’ forward-looking claims about emerging technologies have a mixed record: the mobile-phone warning has largely not been borne out.
3. Irreversibility needs splitting into kinds. Latency “pipelines” and persistence-driven harm are strongly supported. Ecological irreversibility was overclaimed in places (northern cod reopened in 2024) and right in others (western Baltic cod, Caulerpa, converted floodplains). Social and institutional lock-in often proved more durable than the ecological damage.
4. “Threshold” means two different things. Dose thresholds for named hazards receded as study continued (“harm expansion”): strong for the named agents, and partly a selection effect. System thresholds (collapse, tipping points) are real but cannot be located in advance; hindsight supports the reports’ scepticism about generic prediction, and shows that thresholds used as governance reference points get moved.
5. Mixtures and multicausality. Strong: single-agent and “sole cause” framings miss or deny combined effects. Moderate: LL2’s relaxation of causal criteria under multicausality is logically sound but can shield hazard claims from refutation. Some specific synergy claims (pesticide–pathogen) weakened in hindsight.
6. Global problems. An institution’s reach must match the source. Framework agreements come fast and effective measures slowly. The ozone regime succeeded under conditions (concentrated producers, substitutes, finance, a review ratchet, monitoring) that climate lacked.
7. Monitoring. Long-term monitoring found most of the system-level surprises in these reports. It is fragile, and statistically underpowered for rare or precipitous change.
8. Interventions have their own system effects, which the reports rarely examine. Hindsight adds several cases (§5, item 3).
9. Separately regulated problems interact (aerosols and climate, invasive mussels and legacy PCBs, ozone substitutes and climate). This is the clearest “systemic” lesson, and mostly visible only in hindsight (§4).
10. Complexity is a rhetorical resource on both sides. LL2’s editors use it to justify earlier action; LL1’s Great Lakes author says it served those resisting clean-up; regulators and firms used multicausality to demand proof of a sole cause.

2. How the reports frame complexity, systems and scale#

Reports: LL1 (2001). - Framing. The Introduction frames precaution around “serious or irreversible” threats (LL1-00, p. 13) and latency “pipelines” of harm (p. 3). - Ignorance and its proxies. The editors’ synthesis separates ignorance from risk and uncertainty (Box 16.1, LL1-16, p. 170). It offers “practical proxies” for ignorance: irreversibility, novelty, persistence, dispersal to ubiquity, and scale (pp. 170–171). - Scale and experiment. - A global hazard leaves “only one ‘experimental’ model” (p. 171). - Using “the ‘world as a laboratory’” becomes more problematic as actions become more widespread and less reversible (p. 172). - Neglected complexity. Complex, cumulative, synergistic and indirect effects were neglected (p. 169), and real natural systems proved “far removed from the standard assumptions” of regulatory appraisal (p. 175). - Research and uncertainty. More research can “compound uncertainty” (pp. 172–173). - Monopoly. Technologies that were “one, global, near monopoly” amplified surprise (p. 187). - Conclusions. LL1-17 pairs ignorance with a precaution of property screening (persistence, bioaccumulation), long-term monitoring and technological diversity (Table 17.1, p. 192). It asks for science able to handle feedback, synergy, thresholds and instability (p. 193).

Reports: LL2 (2013). LL2 makes the theme explicit. - Introduction. It gives “fast-changing, multiple, systemic challenges” as one reason for the volume, and warns of technologies that are “relatively new, largely unknown, yet already widespread” (LL2-00, pp. 9–10). - Ecosystems (Ch 17). Equilibrium, “variable-oriented” resource management fails in socio-ecological systems (LL2-17, pp. 408–409). The chapter surveys planetary boundaries, tipping elements, panarchy and resilience (pp. 416–419). - Multicausality (Ch 27). - Multicausality weakens most of the Bradford Hill features as grounds for dismissing an association. - The timing of a dose can matter more than its size. - Twelve “criteria for action” include persistence, bioaccumulation, spatial range, irreversibility and novelty (LL2-27, pp. 649–655; Box 27.4, p. 653). - A “knowledge-to-ignorance ratio” is introduced, and “ambiguity” is added as a knowledge state (p. 655; Table 27.1, p. 656). - Conclusions (Ch 28). - Scale, interconnectedness and feedbacks “have outstripped society’s capacity to understand, recognise and respond” (LL2-28, p. 670). - Three drivers of delay: technology that has “often changed” by the time harm is confirmed (a “moving target”, in the notes’ shorthand), sunk-investment lock-in, and scale that makes monitoring very difficult (p. 672). - An “anatomy of complexity”: co-causality, timing, non-monotonic dose, mixtures, variable susceptibility, colony-level emergent effects and long-range transport (p. 674).

Assessment. - The theme is argued mainly in the synthesis chapters. Those chapters apply a framework that existed before the cases were written (the ESTO risk/uncertainty/ignorance scheme; post-normal science; the resilience literature through Ch 17’s lead author) rather than coding the cases for it (LL1-16 notes; context.md §2). Several of Ch 28’s systems claims have no worked case in that chapter. The notes downgrade “moving target” to suggestive for that reason (LL2-28 notes). - Case chapters vary. Some are genuine systems analyses: fisheries, ozone, acid rain, floods, invasive species, neonicotinoids. But even the ozone chapter omits the threshold chemistry of polar stratospheric clouds, arguably the case’s most transferable complex-systems lesson (LL1-07 notes, limitations). - The reports disagree internally about complexity. Gilbertson (LL1-12, p. 129) argues that “supposed complexity and uncertainty” framings served interests resisting costly remediation, and that simple linear causal inference was the stronger basis for action. LL2’s editors argue nearly the reverse (LL2-27, pp. 650–654; LL2-28, p. 674). Both have case support (§5, item 2).

3. Patterns#

A. Scale and speed#

3.1 Deployment outruns knowledge: exposure becomes widespread before the evidence arrives#

Reports. Hazardous agents reached large volumes and near-universal exposure before the relevant science existed or was applied.

The editors generalise the point: - “largely unknown, yet already widespread” (LL2-00, p. 10); - adoption “faster” while policy is “slower” (LL2-28, p. 670); - innovation “outstripping its ability to predict the consequences” (LL2-A2, p. 701); - “All too often technology outstrips the science needed to assess the risks involved” (LL1-07, p. 83).

Assessment: accelerants the cases document. The cases show what sped diffusion, though no chapter analyses it: - mandates: the US 1990 oxygenate requirement drove MTBE use (LL1-11, pp. 110–111); - bundling into another product: seed-dressings (LL2-16, p. 383); - low price and lack of ownership: DES (LL1-08, p. 88); - compatibility with existing infrastructure: MTBE pipelines (LL1-11, p. 110); - uses dispersive by design: aerosols (LL1-07, p. 82).

Cases. - For: at least 11 sections (listed above). - Complicating: - Vinyl chloride’s concentrated use made control feasible once the harm was recognised (LL2-08, p. 189). - The report’s clearest warning about an emerging technology, mobile phones, has largely not been borne out (LL2-21 hindsight). - Swine flu shows that precautionary interventions also scale (§3.2). - The “moving target” claim, that technology has “often changed” by the time harm is confirmed (LL2-28, p. 672), is asserted without a worked case.

Hindsight. - Strengthened for persistent chemicals: - PFAS became the post-2013 repeat of the pattern (LL2-00 hindsight, including a Commission cost estimate of EUR 440bn); - 92% of European adults exceed EFSA’s 2023 BPA guidance value (LL2-10 hindsight); - Weaker for emerging technologies: - mobile-phone risk evidence is largely reassuring, and exposure fell through unrelated technical change (LL2-21 hindsight); - GM human-health “indications” collapsed with the Séralini retraction (LL2-19 hindsight); - nano produced specific signals, not general harm (LL2-22 hindsight).

Rating. - Strong as a description of the historical persistent-chemical cases: eight or more sections, with production and exposure data, confirmed by hindsight. - Moderate as a general claim about new technologies: the forward-looking cases have mixed outcomes. - Suggestive for the individual accelerants: each rests on one or two cases.

3.2 Scale turns small or rare effects into large aggregate harm, and growth swamps per-unit gains#

Reports. The cases show two different scale effects.

Aggregation (a small effect per person, multiplied across a large population): - Lead: industry dismissed an average loss of about 5 IQ points as “small”, ignoring what a shifted distribution does at both tails (LL2-03, p. 61). The gains from removing lead are valued at USD 100–300bn per US birth cohort (p. 62). - Swine flu: about 40 million people were inoculated; 107 reported cases of a rare neurological side-effect (Guillain-Barré syndrome), including six deaths, halted the programme and were followed by more than 4,100 lawsuits (LL2-02, pp. 26, 28). The authors conclude that “particular care” is needed with large-scale introductions because of “unknown unknowns” (p. 35). - Benzene: diffuse exposure through petrol outlasts workplace controls (LL1-04, p. 45).

Composition (growth in volume offsetting gains per unit): - Ozone: the 1977–80 aerosol cuts were offset by growth in foam blowing (LL1-07, p. 80). - Tall stacks: they improved urban air while European SO2 emissions more than doubled after 1950, to about 57 Mt in the mid-1970s (LL1-10, pp. 101–103). - Minamata: process “improvements” raised the by-product per unit of output. Combined with rising output, they increased total methylmercury waste more than eight-fold by 1959 (LL2-05, pp. 95, 102).

Hindsight. - Lead strengthened: lead-attributable deaths rose from about 0.9m (GBD 2019) to 3.5m (GBD 2023) once adult cardiovascular effects were modelled (LL2-23 hindsight). New causal economics studies link leaded petrol to crime, schooling and earnings (LL2-03 hindsight). - Radiation: CT growth swamped falling per-examination doses, pushing UK collective medical dose up by about half (LL1-03 hindsight). - Benzene: gasoline was classified IARC Group 1 in 2025 (LL1-04 hindsight).

Assessment. - Governance that benchmarks performance per unit (dose per examination, emission per stack, by-product per tonne) can show improvement while totals rise. - Scale cuts both ways. Swine flu is a precautionary intervention whose own rare harm was multiplied by the scale of deployment. The report treats it as a false positive, not as a scale lesson about interventions in general.

Rating. - Strong for aggregation in the lead case (quantified and independently confirmed). - Strong for composition in the SO2 and ozone cases (emissions data). - Moderate as a general claim (four to six sections). - Moderate for the swine-flu lesson that interventions also scale (one case, well documented).

B. Time: persistence and irreversibility#

3.3 Persistence and stocks: stopping production does not stop harm#

Reports.

The editors turn this into two principles: - persistence is a proxy for unknown harm (LL1-16, pp. 170–171); - the “inertness–persistence paradox”: short-term safety “appears to demand” non-reactive chemicals, which will therefore be extremely persistent (LL1-07, p. 83).

Hindsight. Broadly strengthened, with refinements. - CFC-12 is now projected at about 44% of its 2001 level in 2100, because banks were larger than assessed (LL1-07 hindsight). - Exemptions leaked. Emissions from ODS feedstocks run at about 3.6% of production, against the 0.5% assumed when feedstocks were exempted. That could delay mid-latitude recovery by about 7 years (LL1-07 hindsight). - A stock became a resource. The halon bank turned from a stock to be destroyed into a scarce supply that civil aviation depends on (LL1-07 hindsight). - PCBs. The Stockholm Convention’s 2025 elimination goal is off track, and top predators remain above toxicity thresholds (LL1-06 hindsight). - Great Lakes decline was not monotonic. - Invasive mussels remobilised PCBs, and dissolved PCBs in Lake Ontario rebounded from a late-1990s low. - PCB-11, an inadvertent by-product congener, did not decline at all (LL1-12 hindsight). - TBT sediment half-times reach up to 33 years (LL1-13 hindsight). - DBCP exceedances in some Californian wells are projected to about 2080 (LL2-09 hindsight). - Asbestos. UK male mesothelioma deaths peaked at 2,205 in 2016. The timing was right; the magnitude was overstated (LL1-05 hindsight). - Weakened: MTBE. The “everlasting” claim did not hold. Biodegradation and natural attenuation are now established, and US concentrations fell after the phase-out (LL1-11 hindsight). - Institutionalised: EU persistent, mobile and toxic (PMT/vPvM) hazard classes in 2023, a proposed PFAS group restriction, and wider property-based screening (LL1-11, LL1-16, LL1-17 hindsight).

Assessment. Hindsight adds four refinements to the reports’ persistence lesson: 1. After a source is removed, decline is fast and then slow. 2. Residual floors can come from exemptions and by-products, not only from legacy stocks. 3. Stocks interact with ecological change and can be remobilised. 4. A stock can change status from liability to resource, which complicates the case for destroying it.

Rating: strong. At least 12 sections, independent monitoring, and confirmation and extension in hindsight. Specific magnitudes are weaker.

3.4 Irreversibility: central to the reports’ logic, but a matter of timescale and yardstick#

Reports. Irreversibility is a trigger in the definitional core of the principle (LL1-00, p. 13; LL2-27, p. 649; Box 27.4, p. 653). It underpins the argument about asymmetric error: a provisional precautionary measure can be reversed, while missed harm may not be (LL2-28, p. 673). The evidence offered: - Latency pipelines: asbestos, ozone, and DES effects in the next generation (LL1-08, pp. 87–88). - Ecological collapse: - northern cod’s “irreversible demise” (LL2-17, p. 409); - fish stocks that “may never recover” (LL1-16, p. 171). - Converted landscapes: more than 90% of former floodplains in Europe, North America and Japan are “functionally extinct” (LL2-15, p. 349). - Invasions past control: Caulerpa and signal crayfish (LL2-20, pp. 497–498). - Developmental harm: growth effects in exposed Great Lakes infants were called “irreversible” (LL1-12, p. 127).

Farman is more careful. The ozone problem can “neither be avoided, nor rapidly alleviated” once recognised (LL1-07, pp. 82–83): the layer can recover physically, but not on policy timescales.

Hindsight. - Overturned: northern cod. - The moratorium ended in June 2024, and the stock was assessed as in the Healthy Zone in 2026. - Part of that change reflects a limit reference point lowered by about 40%. - Capelin and the wider fish community have not recovered. - The outport fishery lost more than 70% of its ports, vessels and fishers between 1998 and 2023 (LL2-17 hindsight). - Supported: Baltic cod. Western Baltic cod is in “a novel and likely irreversible low productivity state”. For eastern Baltic cod, ICES cannot identify any catch that would allow recovery (LL2-17 hindsight). - Weakened: Arctic summer sea ice. The IPCC finds no tipping point, and losses “potentially reversible” (LL2-17 hindsight). - Weakened elsewhere: MTBE’s “everlasting” risk (§3.3) and the Great Lakes certainty language (LL1-12 hindsight).

Assessment. Four kinds of irreversibility need separating: 1. Committed harm from past exposure (latency). Well supported: asbestos, radiation, DES, ozone. 2. Persistence-driven stocks. Well supported (§3.3). 3. Ecological state changes. Case-specific, and often reversible over decades, but with uncertain completeness and sensitivity to how “recovered” is defined. 4. Social, economic and institutional lock-in. Frequently the most durable: Newfoundland’s fishing communities, beekeepers leaving the trade (LL2-16, p. 383), floodplain settlement (LL2-15, pp. 359, 362).

In the reports, “irreversible” often means “not reversible on policy timescales”. Claims of recovery depend on who sets the yardstick.

Rating. - Strong for kinds 1 and 2 (eight or more sections). - Moderate for ecological irreversibility: roughly as many of the reports’ claims were overturned or softened as were confirmed. - Asserted / normative for the asymmetric-error argument. It is a sound conditional, but its premises must be checked case by case. Critics argue that precautionary measures can also be effectively irreversible (critiques.md §3.3; LL2-02 hindsight on false positives such as saccharin and irradiation lasting decades).

C. Non-linear behaviour in natural and biological systems#

3.5 Ecosystem dynamics: masked decline, equilibrium assumptions and interacting species#

Reports. - Indicators that mask decline. - Rising fishing effort and technology kept landings up while stocks fell. In 1865 Bertram showed nets per boat nearly quadrupling while catches fell. Scottish landings stayed flat from 1898 to 1998 while the number of fishers fell from 36,161 to 7,771 (LL1-02, pp. 17–19). - Templeman warned that catch rates “might continue to increase even as the stock was collapsing” (LL2-17, p. 413). - Structural assumptions. - Northern cod was treated as one unit although it comprised distinct populations. - Recruitment and natural mortality were held constant. - Discards distorted the data. - Inshore data were excluded as “messy and often anecdotal” (LL2-17, pp. 411, 414; LL1-02, p. 22). - Bias that recurred. Assessments that converged only in hindsight overestimated stock size and underestimated fishing mortality. The same bias appeared in three of five North Sea stocks, and had been known since 1977 (LL1-02, pp. 21, 24). - Interacting species. - A Grand Banks multispecies model became less predictable as realism was added: “a gigantic random number generator” (LL1-02, p. 25; LL1-16, pp. 172–173). - The Barents collapse ran through a capelin–cod cannibalism–seal cascade combined with a temperature shift (LL2-17, p. 411). - Regime shifts. Systems exist as “configurations of states” separated by thresholds (Sahel, Baltic algal blooms, Caribbean reefs) (LL2-17, pp. 417–418). Table 17.3 classifies warnings as spatial, temporal, demographic, threshold and cascading (p. 423). - Beyond fisheries. - Flood design codes assume “the past is the key to the future”, on records too short to define rare floods (LL2-15, pp. 351–352, 355). - Forest decline resisted any single-cause explanation (LL1-10, p. 104). - No list of traits predicts which species will become invasive; the best predictor, invasiveness elsewhere, degrades as conditions change (LL2-20, pp. 490, 501).

Cases. - For: five sections (LL1-02, LL2-17, LL2-15, LL2-20, LL1-10). - Complicating: - Ch 17’s claim that “diversity-oriented” analysis should replace “variable-oriented” analysis is asserted, not shown. Most of the cod failures could have been corrected within a better conventional assessment, and the chapter concedes there is no consensus replacement (LL2-17, pp. 409, 415; notes). - Many stocks were rebuilt under model-based rules, including Barents cod after 1990 (LL1-02 hindsight). - The Norway exemplar contains factual errors (LL2-17 hindsight).

Hindsight. - Fisheries. - The bias persists 26 years on (North Sea 2025; northern cod 2025–26), and managers still set catches above advice (LL1-02 hindsight). - DFO’s current model now estimates natural mortality year by year (0.22–0.88) and includes capelin: a better variable-oriented model, not a different paradigm (LL2-17 hindsight). - Recruitment relationships were non-stationary in more than half of 23 North Sea, Norwegian Sea and Barents stocks. - Barents cod fell from a record in 2013 to below its precautionary reference point by 2025, and a 20% cap on year-to-year catch changes delayed the response (LL2-17 hindsight). - Floods. The 2021 Ahr peaks matched unrecorded floods of 1804 and 1910 (LL2-15 hindsight). - Invasions. Rates are “often even accelerating” (LL2-20 hindsight).

Assessment. What survives is not a methodological dichotomy but three mechanisms: 1. Indicators generated by the activity itself can stay reassuring during collapse. 2. Models that hold key parameters constant, or pool heterogeneous units, convert a changing system into a stable average. 3. Institutional rules on admissible evidence exclude the local signals that would reveal both.

Rating. - Strong for the three mechanisms in fisheries: independent official reviews (Alverson, Harris, the Auditor General) and confirmation in hindsight. - Moderate as a claim beyond fisheries: floods and invasions support parts of it. - Asserted for the diversity-versus-variable dichotomy.

3.6 Thresholds and tipping points: receding dose thresholds, unknowable system thresholds#

The reports use “threshold” in two senses that behave differently.

(a) Dose thresholds recede. - Reports. The concluding chapter names “harm expansion”: asbestos, tobacco, PCBs, lead and radiation proved harmful in more ways and at lower doses, and “more often than not, no ‘safe’ threshold” was found (LL2-28, p. 672; also LL2-A3, p. 717). Examples: - the 1925 radiation tolerance dose was about 700 mSv a year, against 20 mSv now (LL1-03, p. 33); - the “safe” blood-lead level fell from 60 to 10 µg/dl (LL2-03, p. 62); - TBT’s 20 ng/l target was set while dogwhelks were sterilised at 3–5 ng/l; it was revised to 2 ng/l (LL1-13, pp. 137–138); - chronic beryllium disease appeared below the exposure limit (LL2-06, pp. 134–135); - the DBCP limit was set below the lowest dose tested, and harm appeared in plants whose measured air levels were 12–125 times lower than that dose (plant-wide sampling, not individual doses; LL2-09, pp. 205–206). - Hindsight. - Strengthened for the named agents: CDC 3.5 µg/dL and an EU “non-threshold” designation for lead (LL2-03); a tenfold cut in the beryllium limit (LL2-06); a 20,000-fold cut in the BPA TDI (LL2-10); EU asbestos limits cut 10–50-fold (LL2-A3). - Counterexamples exist: nickel (2020) and perchlorate (2026) limits did not fall (LL2-26 hindsight). - Hindsight calls the “harm expands” thesis partly a selection effect (LL2-A3 hindsight). - Rating. Strong for the named agents (six or more). Moderate as a general prior.

(b) System thresholds exist but cannot be located in advance. - Reports. - Critical loads. Setting “the highest deposition… that does not change the ecosystem in an unacceptable way” made action tractable despite causal ignorance. It was “not exactly” precaution, and only 60% gap closure was achievable at the worst sites (LL1-10, pp. 106–107). - Planetary boundaries face “many decades of tantalising ignorance” (LL2-17, pp. 416–417). - Generic tipping-point models are “an exercise in futility”; warnings must come from local and regional observation of key processes (p. 417). - 2 °C is “a political decision” (LL2-14, p. 327). - Susceptibility varies, so “tipping point exposures” differ between groups (LL2-28, p. 674). - Omission: the ozone chapter uses 2.5 ppbv equivalent chlorine as a conventional marker (LL1-07, p. 78) but omits the non-linear polar-cloud chemistry (notes). - Hindsight. - Critical-load exceedance for acidity fell to 3.5% of Europe’s mapped ecosystem area by 2024, well beyond the 60% gap closure (LL1-10 hindsight). A threshold-based regime worked. - The planetary-boundaries team reports seven of nine boundaries “transgressed”, but transgression is a precautionary judgement, not an observed crossing. - Generic early-warning statistics proved useful for diagnosing transitions after the fact but not for timing them; physics-based, locally observable signals performed better in at least one major case (LL2-17 hindsight). - Reference points were revised downward: the northern cod limit reference point, and EU biomass triggers that make stock status look over-optimistic (LL2-17 hindsight). - Climate ceilings (450 ppm CO2-equivalent) were crossed and redrawn (LL2-14 hindsight). - Assessment. - System thresholds used in governance are value choices (“unacceptable change”) and can be moved in either direction. - The durable lesson is procedural. Agree diagnostic criteria for action in advance (LL2-17, p. 423), and protect them from convenient revision. That second clause comes from hindsight, not the report. - Rating. Moderate. The concept is strong. The specific claims are mixed, and few cases are worked through.

3.7 Non-linearity and timing: why acute, high-dose and average-based tests mislead#

Reports. - Acute or high-dose endpoints mis-set protection. - TBT acute tests implied effects only at µg/l; real effects were at ng/l, a hazard “grossly underestimated” (LL1-13, pp. 136, 141). - Controlling acute beryllium disease left chronic immune disease untouched (LL2-06, pp. 133–134). - Early radiation limits were calibrated to acute effects (LL1-03, p. 33). - Timing matters more than dose. - DES caused harm after prenatal exposure (LL1-08). - “the time makes the poison” (LL2-10, p. 219). - Developmental windows can make timing more important than dose (LL2-27, Box 27.2, p. 650). - Non-monotonic dose-response curves (LL2-10, pp. 218–219; LL2-28, p. 674). - Amplification by trace agents. A 10,000-to-1 abundance ratio would have led experts to dismiss laboratory hints about chlorine and ozone (LL1-07, p. 82). - Non-linear system response. Flood runoff depends non-linearly on how wet the catchment already is (LL2-15, p. 357). - The wrong dose metric. Surface area, not mass, better predicted the lung response to ultrafine particles (LL2-22, pp. 534, 541). - Small can matter. In interacting systems “‘Small’ can be very significant” (LL2-27, p. 651).

Hindsight. - TBT. The mechanism was corrected (retinoid X receptor activation). A specific biological marker had established causation even though the mechanism was wrong (LL1-13 hindsight). - BPA. EFSA found no convincing non-monotonic dose response. Its 2023 intake limit rests on a monotonic effect (LL2-10 hindsight). - Benzene. The claimed 3–9-fold underestimate from non-linear metabolism remains contested (LL2-A3 hindsight). - Developmental timing is broadly supported: DES, lead, and prenatal PCB effects on cognition (LL1-08, LL2-03, LL1-12 hindsight).

Rating. - Strong that acute or high-dose endpoints mis-set protection (five or more cases, confirmed). - Moderate for developmental timing as a general principle. - Suggestive and contested for non-monotonic responses at environmental doses.

3.8 Mixtures, co-causes and multicausality#

Reports. - Asbestos and smoking. Together they raised lung-cancer risk more than 50-fold, against 5-fold and 10-fold separately. The interaction also complicated attribution and compensation (LL1-05, p. 55). - Oestrogen mixtures. - EE2, E2 and E1, each at or below its own no-effect level, are “particularly potent when present in combination”, so risk assessments “are not adequately protective if based on data for individual substances”. Anti-androgens add to the effect (LL2-13, p. 290). - The authors ask whether it is “scientifically incorrect to blame just EE2” for treatment costs (p. 293). - PCBs. They are a heterogeneous class of congeners. Weathered environmental residues appeared more toxic than the commercial product (LL1-06, pp. 66–67). - PCE. It was involved in 4 of 9 more-than-additive pairs in a 1969 acute-toxicity screen (LL2-04, p. 82). - Growth promoters. Combinations were not assessed (LL1-16, p. 175). - Neonicotinoids. - Colony-level effects of sublethal harm that acute tests cannot detect (LL2-16, p. 375), and pesticide–pathogen synergy (pp. 387, 391). - The regulator asked whether the product was the sole cause of all losses nationally. That question can never be answered yes in a multicausal system, so the advice stayed inconclusive (p. 379). - The editors’ position. - Multicausality weakens consistency, specificity, temporality and dose-response as grounds for dismissing an association (LL2-27, pp. 652–653). - Inconsistency between studies “is to be expected from complexity” (LL2-28, p. 674). - Removing one link can prevent much harm (p. 674). - Acting on whole mixtures (tobacco smoke, fine particles) worked without mechanistic knowledge (LL2-27, p. 651).

Cases. - For: about seven sections. - Complicating: - Gilbertson’s objection that multi-causal framings gave cover for delay (LL1-12, p. 129). - The neonicotinoid authors themselves used national-scale yields when it suited them, the scale they faulted the regulator for (LL2-16 notes).

Hindsight. - Weakened: pesticide–pathogen synergy. A 2021 meta-analysis found the effect additive, and field studies found none (LL2-16 hindsight). - Adopted: group and effect-based approaches. - The EU banned BPA and hazardous bisphenols together after BPS and BPF substitution (LL2-10 hindsight). - The EU mandated effect-based monitoring of total oestrogenic activity (LL2-13 hindsight). - The 2024 wastewater recast requires class-wide “quaternary” treatment, at least 80% producer-funded (LL2-13 hindsight). - Colony losses remain multi-causal (LL2-16 hindsight).

Assessment. Three separable claims: 1. Single-agent assessment understates combined effects. 2. Demanding proof of a sole cause makes an assessment unanswerable. 3. Causal criteria should be relaxed under multicausality.

The third is logically sound, but it can make a hazard claim immune to disconfirmation. LL2-21’s asymmetric handling of null studies shows the risk (§5).

Rating. - (1) Strong: the asbestos–smoking interaction is quantified, and oestrogen mixture effects are shown experimentally. - (2) Strong: documented in the assessors’ own words, and logically robust. - (3) Moderate.

3.9 Self-propagating and adaptive agents: invasions, prions and resistance#

Reports. Some hazards reproduce, spread or adapt after release. The case evidence shows how their dynamics differ from those of a static pollutant. - Invasive species (LL2-20). - A “lag phase” of decades can precede spread, so the origin of an introduction becomes untraceable and liability fails (p. 497). - The decision point comes long before the evidence (p. 497). - Windows for eradication close fast: weed eradication cost rises at least 40-fold with delay (p. 487). Caulerpa was found early in France in 1984 but tackled late, whereas California began eradication 17 days after detection in 2000 and succeeded (p. 498). - The hazard scales with trade and travel. GDP is the top predictor of invasive species richness on islands, which argues for regulating pathways rather than inspecting items (pp. 490, 498–500). - Fixes became hazards: signal crayfish introduced to replace plague-hit native crayfish carried the plague and became invasive (p. 497); cane toad (p. 496). - Invasions show “no indication yet of any saturation effect” despite more than 42 treaties (p. 493). - BSE. Recycling slaughterhouse waste into feed amplified the agent; proposed 1979 rendering standards were withdrawn, with effects the authors call “not yet clear” (LL1-15, p. 158). Partial feed bans then leaked through cross-contamination and exports (pp. 160, 163). - Resistance. - Antibiotic efficacy is a shared, depletable resource that farm use erodes for medicine (LL1-09, pp. 93–97; LL2-A3, pp. 731–732). - Establishing the risk formally would need a 17-step causal chain for each gene and substance (LL1-09, pp. 95–96). - DDT resistance, recognised in 1948–51, was read as a prompt for new chemicals (LL2-11, pp. 241, 243). - Herbicide-tolerant crops set off a resistance “treadmill” (LL2-19, p. 462). - Persistent seed-dressings place constant selection pressure on pests (LL2-16, p. 385).

Hindsight. - Invasions. No saturation, and rates “often even accelerating”. IPBES (2023) puts costs above USD 423bn in 2019, quadrupling every decade. The 2022 Global Biodiversity Framework adopted a target to halve introductions. First records of deliberately traded vertebrates are declining, a possible sign that regulating intentional pathways works. Binding bans also leak without enforcement (LL2-20 hindsight). - BSE. Leakage through feed accounted for about 23% of UK cases. Latent exposure reached blood transfusion, with three clinical transfusion cases. Classical BSE was still being confirmed in 2026 (LL1-15 hindsight). - Antimicrobials. - Growth-promoter bans shrank animal resistance reservoirs. - Co-selection defeats single-agent controls: the ionophore narasin, kept on the market because it is “not used in humans”, was shown in 2026 to co-select for vancomycin-resistant enterococci. - Human-health benefit is inferred mainly from mechanism and reservoir reductions (LL1-09 hindsight). - DDT and pyrethroids. Resistance eroded DDT and now pyrethroids: resistance to pyrethroids is reported in 87% of reporting countries (LL2-11 hindsight). - Herbicides. The treadmill recurred for each stacked replacement trait (LL2-19 hindsight).

Assessment. For agents that reproduce or adapt, the cost of delay is non-linear. Harm can grow after release with no further input. Controls aimed at a single agent are eroded by adaptation and by linked traits.

Rating. - Strong for adaptive resistance wearing down single-mode control: three independent domains, confirmed in hindsight. - Strong for closing windows and the lack of saturation in invasions. - Moderate for quantified human-health benefit of acting on resistance.

D. Systemic rather than product-by-product risk#

3.10 Substitution treadmills and the unit of regulation#

Reports. - Antifoulants. TBT was replaced by booster biocides, “history repeating itself”. The swap exchanged an endocrine hazard for broad photosystem-II inhibition acting on a highly conserved target at the base of food webs (LL2-12, pp. 265, 273–276). The 1989 EC ban on small-boat TBT was itself “triggering the development and use” of the boosters (p. 276). - Ozone substitutes. Farman called the open-ended HCFC/HFC transition “deeply flawed” (LL1-07, p. 81). - Lead to MTBE (LL1-11). - Asbestos substitutes of the same fibre form (LL1-05, Lesson 10). - The editors’ “substitution traps”: substitutes sharing the hazardous property; substitutes judged only against a worse predecessor; “chemical for chemical” swaps (LL1-16, pp. 173–177; LL1-13, p. 142). - DDT to pyrethroids, which carry their own uncertainties (LL2-11, p. 250). - Neonicotinoids: near-equivalents were put forward after the bans (LL2-16, pp. 387, 393). - Substitutes must also be subject to “research, precaution, and exposure control”, the conclusions say (LL2-28, p. 676).

The remedies the reports propose: - assess substitutes on their own terms (LL1-11, p. 117); - alternatives assessment (LL2-02, p. 35); - rethink the function rather than the chemical (LL1-13, pp. 141–142; LL1-07, p. 81); - screen on properties (LL1-17, p. 192); - group similar materials (LL2-22, p. 547); - regulate pathways (LL2-20, pp. 498–499).

Cases. - For: nine or more sections. - Complicating: - HCFCs may have sped the CFC exit (LL1-07 notes). - Lower nitrite plus ascorbate was “smart management” rather than a regrettable swap (LL2-02, p. 25). - DDT’s decline came through bed nets and new insecticides (LL2-11 hindsight).

Hindsight. Strong confirmation across cases:

Assessment. When regulation targets a named product rather than its function or hazardous property, it creates demand for the nearest substitute, which inherits a presumption of safety. The institutional responses were group restrictions, property-based hazard classes, function-based alternatives assessment and pathway rules.

Rating: strong.

3.11 Dispersal and displacement: moving harm across space, media and jurisdictions#

Reports. - Tall stacks are the core case. They improved the visible local indicator while total emissions grew and harm moved to Scandinavian lakes, changing in kind (LL1-10, pp. 101–103, 107). - Minamata: moving the outfall spread the poisoning (LL2-05, p. 100). - Floods: upstream defences raised downstream peaks; dike breaches in one country eased the load on a town in another (LL2-15, pp. 350, 354). - Export of banned products: - DBCP was used on plantations to at least 1985–86 (LL2-09, pp. 207–209); - leaded petrol was exported on the strength of expert assurance (LL2-03, pp. 67–68); - DES use continued abroad (LL1-08, p. 89); - asbestos bans in over 50 countries were offset by growth in China and India (LL2-A3, pp. 724–726); - UK exports of meat-and-bone meal to the EU roughly doubled after the domestic feed ban (LL1-15, p. 163). - Long-range transport: - PCBs reached the Arctic (LL1-06, p. 70); - Greenland bears mercury damage of USD 59m a year with almost no local emissions (LL2-23, p. 569). - Jurisdictional mismatch: - a single market means a species harmless in the UK but invasive in the Mediterranean is traded freely (LL2-20, p. 497); - the most sensitive receptors (corals) lie outside EU jurisdiction (LL2-12, pp. 270–271). - Shift between media: an air-quality mandate harmed water (LL1-11, p. 114).

Hindsight. - Asbestos use was 930,000 t in 2025, mainly in Asian construction (LL1-05 hindsight). - Mercury use migrated to gold mining in the Global South (LL2-05 hindsight). - DBCP was never listed under the Rotterdam Convention (LL2-09 hindsight). - The EU extended its growth-promoter rule to imports in 2026 (LL1-09 hindsight). - Removing sulphate aerosols unmasked warming: one modelling study estimates about 0.45 °C over Europe from 1970–2010 policies, while about 80,000 deaths a year were prevented (LL1-10 hindsight).

Rating: strong. Ten or more sections. The tall-stack and export cases are documented and confirmed.

E. Governing and observing systems at scale#

3.12 Global commons: institutions of matching reach#

Reports. - TBT. Regional bodies could restrict small boats but not shipping. The global body said in 1994 that no further controls were needed, then reversed within three years. The authors conclude “It would seem” only global restrictions can work (LL1-13, pp. 138, 140, 142). - Ozone. - Consensus-paced negotiation with a periodic-review ratchet (LL1-07, pp. 78, 81). - A first-mover penalty: the US claimed its share of world CFC production fell from 46% to 28% after it acted alone (p. 80). - 13 company groups held about 75% of output, and a fund financed developing-country conversion (p. 81). - Acid rain. Jointly produced source–receptor evidence, CLRTAP sustained partly as East–West détente, and national positions that followed costs (LL1-10, pp. 103–107). - Climate. - The UNFCCC was agreed two years after the first IPCC report, before human influence had been detected (LL2-14, pp. 321–322). - Emissions still rose from about 38 to about 50 Gt CO2-eq between 1990 and 2010 (p. 325). - Kyoto compliance was flattered by windfalls (p. 324). - Precaution has mostly been applied where implementers were also beneficiaries (p. 337). - Mercury. A global view emerged only in the 1980s–90s, and action stayed slow despite strong assessments (LL2-05, pp. 125–126). - Invasions. More than 42 treaties, but no saturation (LL2-20, p. 493).

Hindsight. - Ozone became universal (198 parties) and adaptive: HCFC acceleration in 2007, Kigali, and detection and reversal of illegal CFC-11 production (LL1-07 hindsight). - TBT. The convention covered 95.77% of world tonnage by 2025 (LL1-13 hindsight). - Climate. - Emissions reached a record 54.1 Gt CO2-eq in 2025. Most of Kyoto’s over-compliance was “hot air”, and most offset projects were judged unlikely to be additional (LL2-14 hindsight). - The US withdrew from Paris with effect from January 2026 and directed withdrawal from the UNFCCC and IPCC (LL1-07 hindsight). - Mercury. The Minamata Convention was adopted and ratified, but shows no clear effect yet on global emissions (LL2-05 hindsight). - Acid rain. East Germany and Czechoslovakia signed the 1985 30% protocol while their emissions were still rising. The large cuts came with economic collapse (LL1-10 hindsight).

Assessment. The ozone–climate contrast suggests the conditions under which global action worked: - concentrated producers; - available substitutes; - finance for late adopters; - a ratchet built into the agreement; - verification by independent monitoring.

Where costs and harms are diffuse, formal commitments can coexist with rising harm until an outside shock intervenes.

Rating. - Strong that institutional reach must match the source (TBT, ozone, acid rain). - Strong that framework agreements come well before effect (climate, mercury, invasive species, acid rain). - Moderate for the conditions of success: a small-n comparison, and my analysis.

3.13 Seeing complex systems: monitoring, sentinels and the limits of detection#

Reports. - Surprises found by long-running measurement. - Antarctic ozone loss came from systematic long-term measurements begun for exploration. Theory expected the first effects elsewhere, and satellite software flagged very low values as “suspect” (LL1-07, p. 82). - A precipitation network built to study plant nutrients detected acidification (LL1-10, p. 102). - Keeling’s CO2 record (LL2-17, pp. 409–410). - TBT baseline studies are “frequently underrated” (LL1-13, p. 142). - Indicator organisms in the Great Lakes made both success and stalling visible (LL1-12, p. 129). - Surprises found by accident. - Intersex fish were spotted in a routine check (LL2-13, p. 282). - BPA and Irgarol were found incidentally (LL2-10, p. 217; LL2-12, p. 267). - Limits of detection. - A decline of 50% in 15 years would go statistically undetected for 72–90% of whale stocks and 55% of polar bear stocks (LL2-26, p. 634). - Monitoring fails in the extremes it exists to measure: gauges were destroyed and the flood information office itself flooded (LL2-15, pp. 359–360). - Fragility and design. - The global invasive species programme closed “for financial reasons” (LL2-20, p. 489). - Fund long-term databases “even when an immediate need is not perceived” (LL1-03, p. 36). - Monitoring without action thresholds becomes an “academic pursuit” (LL2-12, p. 274).

Hindsight. - Illegal CFC-11 production was detected by atmospheric monitoring (LL1-07 hindsight). - The Lake Ontario rebound and the PCB-11 anomaly were visible only through long series (LL1-12 hindsight). - Russia’s departure from ICES split the Barents assessment (LL2-17 hindsight). - UK radiation surveillance units were closed when no need was perceived (LL1-03 hindsight).

Complicating. - Farman: monitoring at the level needed cannot be funded open-endedly (LL1-07, p. 82). - The editors read the ozone discovery as serendipity, where Farman stresses systematic measurement (LL1-16, pp. 172–173 against LL1-07, p. 82).

Rating: strong. Eight or more sections. Monitoring was decisive for ozone, acid rain, CFC-11 and the Great Lakes, and hindsight supplies further instances.

3.14 Coupled institutions: weakest links, silos, fit and memory#

Reports. - Warning chains break at interfaces between organisations. - At Vaison-la-Romaine an accurate alert did not match the local need. - On the Odra, the authority to warn sat with one body and the knowledge with another (LL2-15, pp. 353, 360). - Attention decays. The “hydro-illogical cycle”: return periods exceed political horizons (LL2-15, pp. 360–361). The conclusions generalise it (LL2-28, p. 680). - Scale mismatch. Poor “institutional fit” between governing bodies and ecosystems (LL2-17, pp. 407, 419). - Turf and deferral. - A health-versus-agriculture dispute delayed ragweed control (LL2-20, pp. 496–497). - Governments cited EU coordination as a reason not to act nationally (Box 20.4, p. 501). - Remit and silo blind spots. - Disciplinary capture split health from environment (LL1-16, p. 174). - Drug regulators had no environmental remit (LL2-13, p. 284; LL1-14, p. 152). - Assessment split by medium missed MTBE’s effect on water (LL1-11, p. 114). - Tightly coupled systems. - A station blackout defeated all redundant safety systems at once. - Probabilistic assessment struggled with common-cause cascades. - An earlier published estimate of tsunami recurrence did not reach the plant’s design basis (LL2-18, pp. 438–439, 447).

Hindsight. - The German floods of 2021 and Valencia 2024 repeated the weakest-link failures (LL2-15 hindsight). - The IAEA’s 2015 report confirmed that probabilistic assessment missed common-cause cascades. The method was broadened, not abandoned (LL2-18 hindsight). - The EU invasive species regulation arrived late and was implemented slowly (LL2-20 hindsight).

Rating: moderate overall. The floods case is strong after hindsight, and Fukushima rests on official inquiries. The other mechanisms are mostly single-case.

4. Interactions between problems#

Assessment, drawing mainly on hindsight. The reports treat each hazard in its own chapter, but several of the most consequential system effects are interactions between chapters.

Lesson. Problems regulated separately interact. Both the benefits and the costs of intervening can arrive through channels outside the frame of the decision. LL2-23’s argument that cost-of-inaction estimates are lower bounds because they count only what is already quantified (p. 568) is one expression of this, and hindsight on lead strengthened it (§3.2).

Rating: moderate. Several documented instances, but mostly identified after the fact.

5. Counter-evidence, complications and critiques#

1. Selection and hindsight. Every historical case was chosen because harm occurred (LL1-00, pp. 12–13; LL2-A2 notes). So the reports can show that system effects were missed. They cannot show how often warnings about system effects were wrong. Hindsight supplies several that were: - northern cod’s irreversibility; - pesticide–pathogen synergy and the honeybee as sentinel; - Arctic summer sea ice as a tipping element; - SO2 as the cause of general forest decline (German crown thinning was worse in 2025 than in 1984, now driven by drought; LL1-10 hindsight); - MTBE as “everlasting”; - the mobile-phone risk; - GM human-health harm; - the claim that geologists had settled the Anthropocene (LL2-17 notes). The complexity-based warnings in the reports are not uniformly reliable.

2. Complexity is a two-edged argument. - LL2-28 concedes that uncertainty “can be a two-edged sword” but treats only its use against hazard evidence as misuse (p. 675; notes). - Gilbertson argues that complexity framings served those resisting clean-up (LL1-12, p. 129). - The French regulator’s “sole cause” framing used multicausality to stall (LL2-16, p. 379). - Conversely, the claim that inconsistency “is to be expected from complexity” (LL2-28, p. 674) can make a hazard claim unfalsifiable. LL2-21 illustrates this: latency is used to discount null studies while short-latency positive findings are accepted (notes). - An analyst needs a symmetric test that asks what evidence would count against the concern.

3. Interventions in complex systems have their own non-linear effects, and the reports underweight them. - Swine flu at scale (LL2-02, p. 28). - Withdrawing DDT in South Africa before substitutes were proven was followed by resurgent malaria (LL2-11, pp. 243, 250). - Precaution about biocontrol prolonged invasions (LL2-20, p. 496). - Fukushima evacuation caused 2,351 disaster-related deaths, against no documented radiation-caused disease (LL2-18 hindsight). - Sulphate removal unmasked warming (LL1-10 hindsight). - French sugar-beet yields fell during a virus outbreak after the neonicotinoid ban (LL2-16 hindsight). - The fire-safety trade-off is omitted from the PCB chapter (LL1-06 hindsight). - Nuclear phase-outs carried diffuse mortality costs (LL2-02 hindsight). - Critics make this point directly: Graham and Wiener on countervailing risk, Sunstein on “system neglect” (critiques.md §§3.1–3.2). LL2 files risk–risk trade-offs outside its count of precautionary errors (critiques.md §3.2).

4. Much large-scale improvement came from structural change, not governance. - SO2 fell with Eastern European economic collapse (LL1-10 hindsight). - Kyoto compliance rode on windfalls (LL2-14, p. 324). - DDT’s exit was driven by resistance, new tools and bed nets, not the road map (LL2-11 hindsight). - Mobile-phone exposure fell through technical change (LL2-21 hindsight). - DES sales fell after 1953 without regulation (LL1-08 notes). Complex systems produce benign surprises as well as malign ones, and credit for improvement is hard to assign.

5. Methodological claims outrun the evidence. - The “diversity- versus variable-oriented” dichotomy (LL2-17) and the critique of “reductionism” (LL2-27, p. 651) are asserted, not demonstrated. - Hindsight shows conventional models improving and stocks rebuilding under model-based rules (§3.5).

6. Scale-based normative arguments do not discriminate. “No mandate for global experiments” (LL1-07, p. 82) and “world as a laboratory” (LL1-16, p. 172) apply to any widely adopted practice, including precautionary ones. The reports offer property proxies (LL1-16, pp. 170–171; Box 27.4) but no method for weighting them. Ch 28 names factors for choosing an evidence threshold without a decision rule (LL2-28 notes). (Assessment.)

7. Thresholds and reference points move both ways. - Beryllium’s provisional limit hardened into a legal anchor (LL2-06, pp. 133, 137). - Cod reference points were lowered after the fact (LL2-17 hindsight). - Harm-expansion counterexamples exist (LL2-26 hindsight).

8. The quantitative systems claims are the weakest material. - The global invasive-species cost of USD 1.4 trillion is misattributed and unsupported by later work (LL2-20 hindsight). - The “tenfold” rise in Eastern European SO2 emissions was about twofold in the chapter’s own source (LL1-10 hindsight). - The 215 GtCO2-eq ozone climate benefit is probably overstated (LL2-23 digest). - Chernobyl mortality extrapolations are contested (LL2-18 hindsight).

9. Standpoint. Most of the systems chapters are written by protagonists: - Ch 17 by the EEA’s own Executive Director, drawing on her DFO career; - the ozone chapter by the lead author of the 1985 discovery paper; - the acid-rain chapter by a scientist in the Norwegian and European monitoring programmes; - the invasive-species chapter by people running the institutions it describes; - the EE2 chapter largely on its lead author’s own research. Only the neonicotinoid chapter carries an opposing industry panel (LL2-16, pp. 401–402).

10. Comparative critiques. Wiener and colleagues find precaution applied selectively on both sides of the Atlantic (critiques.md §3.4). The useful question is less whether complexity was acknowledged than which system risks triggered precaution, at what evidence threshold, and who bore the costs.

6. Technology-neutral diagnostic questions#

Each is tied to the pattern it draws on, and can be asked of any emerging technology and of those developing, financing and regulating it.

1. What persists? If use stopped tomorrow, what stocks would keep releasing effects, for how long, and who would have to manage them? Consider products in service, infrastructure, environmental reservoirs, bodies, and institutional commitments. (§3.3: CFC and halon banks, installed PCB equipment, TBT sediments, pipe linings.)
2. How does the adoption curve compare with the time needed to detect the slowest plausible harm? How much exposure will have accumulated before a study with adequate follow-up could report? (§3.1: MTBE, CFCs, BPA; the Interphone latency problem.)
3. What is accelerating diffusion? Mandates, bundling into other products, low price, compatibility with existing infrastructure, or dispersive use? Could deployment be staged, reversible or confined while evidence accrues? (§3.1; §3.2 on swine flu scale.)
4. Which indicators produced by the activity itself are being read as evidence that all is well? Could they stay reassuring while the underlying system degrades? Are key parameters held constant, or heterogeneous units pooled, in the models that produce them? (§3.5: catch rates, landings, tall-stack urban air.)
5. Is performance judged per unit while the total grows? Who tracks aggregate volume, cumulative exposure and population-level distribution shifts? (§3.2: lead IQ distribution, CT collective dose, methylmercury per tonne.)
6. Were safety limits derived from acute, high-dose, adult or average endpoints? Which groups, life stages or parts of the system are most sensitive, and has anyone measured effects there? (§3.6(a), §3.7: TBT, DBCP, beryllium, DES, lead.)
7. What is the unit of assessment? A single product, or the combined, cumulative exposure, class and function? If this product were restricted, what is the nearest substitute, and has it been assessed on its own terms rather than against the incumbent? (§3.8, §3.10: oestrogen mixtures, bisphenols, antifoulants, HCFCs, MTBE.)
8. Does the activity draw down a shared resource, or select for adaptations that will erode the fix? Can local controls leak through trade, exports, neighbouring uses or linked traits? (§3.9, §3.11: antibiotic resistance and co-selection, herbicide and insecticide treadmills, BSE feed leakage, DBCP exports.)
9. Does a proposed fix reduce harm, or relocate it to other places, media, populations or times? Is a local indicator improving while the total grows or moves? (§3.11: tall stacks, flood defences, air-to-water spillover.)
10. Does the reach of the governing institution match the scale and mobility of the source? Where do costs and benefits fall across jurisdictions? Are the conditions for collective action present: concentrated producers, substitutes, finance for late adopters, a ratchet, and independent verification? (§3.12: TBT and the IMO, ozone against climate, acid rain.)
11. What independent, long-running observation system could detect unexpected change? What statistical power does it have to detect a large decline in time? Is it protected from funding cuts and political rupture? Are the criteria that would trigger action agreed in advance, and protected from later revision? (§3.13, §3.6(b): ozone, CFC-11, whales and polar bears, cod reference points.)
12. Where would a warning break down? Who holds the knowledge, who holds the authority to act, and how does a signal travel between them, including in an extreme event that disables the monitoring itself? (§3.14: floods, Fukushima, invasive-species turf disputes.)
13. What would the precautionary or corrective intervention itself do at scale? Consider rare side-effects, effects on linked systems, and benefits or harms through channels outside the decision’s frame. (§4, §5 item 3: swine flu, sulphate unmasking, evacuation deaths, biocontrol.)
14. When someone claims a harm is irreversible, or an exposure is safely below a threshold, on what timescale and against what yardstick? Who sets that yardstick, and what evidence would count against the claim? (§3.4, §3.6, §5 item 2: northern cod, harm expansion, MTBE, unfalsifiable complexity arguments.)

7. Sources used#

• Digests: all 47 section digests.
• Notes: read in full or in their relevant sections for LL1-02, LL1-07, LL1-10, LL1-13, LL2-14, LL2-15, LL2-16, LL2-17 and LL2-20; the complexity sections of LL1-16, LL2-27 and LL2-28; targeted passages from LL1-12, LL1-15, LL2-02, LL2-03, LL2-05, LL2-13 and LL2-A2. All notes were searched for synergy, mixtures, additivity, non-monotonic dose, scale, diffusion and persistence.
• Hindsight: files read for LL1-07, LL1-10, LL2-11, LL2-12, LL2-13, LL2-16, LL2-17, LL2-19 and LL2-20; targeted passages from LL1-03, LL1-09, LL1-12, LL1-15 and LL2-14; summaries for all other sections relied on.
• External: external/critiques.md (§§3.1–3.4, 5) and external/context.md (§2).
• Verification: quotations and page references spot-checked against working/text/chunks/.