Late Lessons, Jensen Huang and AI
Markdown version

Balance and bias review of 01-late-lessons-analysis.md#

Strand A critical review, 26 September 2026. Scope: fairness, calibration of strength ratings, treatment of false positives, critiques, hindsight reversals and costs of precaution, and whether any framing pre-judges how the lens will be applied. Line numbers refer to 01-late-lessons-analysis.md as of this date. Other paths are relative to working/late-lessons/.

Overall judgement#

On the whole the document is fair. In places it is more candid about the reports’ weaknesses than the reports are about themselves: see the Chapter 17 analysis (§3.3), the three kinds of explanation (§4.3), the structural limits (§5.1), the conditional restatement of the asymmetry argument (§5.2), and the rules and symmetry checks for the lens (§6.1, §6.11). It does not slide wholesale into the reports’ advocacy.

Its balance problems fall into three groups:

Issues are ranked by how much they would distort later use of the lens.

Issues#

1. The lens is built to find missed harm, not misplaced alarm#

Location. The §6 introduction (l. 643); I1 (l. 764–768); I3 (l. 776–780); I6 (l. 794–797); W3 (l. 716–720); §6.11 (l. 968–976).

Problem. - Of the 55 lens entries, only four have the error or cost of acting as their main subject: T3, T4, C7 and S4. The other 51 are framed to detect harm that is being missed, discounted or hidden. - The introduction names the objects of scrutiny as “those who develop, finance, promote and oversee” a technology. It leaves out those who warn about it, campaign against it, litigate over it or restrict it. - The interest entries point one way only: - I1 asks about a gap between the developer’s private and public positions. - I3 asks who funds “the studies behind the safety claims”, but not the studies behind the harm claims, even though its own Limits line concedes that “the reports rarely scrutinise critics’ funding”. - I6 covers liability that rewards not knowing, but not litigation, funding or reputation that rewards finding. - W3 (the reassurance trap) has no mirror entry, although the document’s own evidence shows that alarms and restrictions harden in the same way. - Symmetry appears only as rule 2 and in §6.11, both of which a user can skip. No individual Ask line asks about the stakes of those raising a concern.

Evidence. - The theme files asked for symmetry question by question, and the synthesis compressed that away. This is the same failure §5.1(4) diagnoses in the reports: - themes/T08 §15 preamble (“should be asked of proponents and critics”) and Q15; - themes/T09 §13 preamble (“applies to proponents … and to those warning about it”), Q9 (“Who is writing the warning … does the claim rest mainly on their own studies?”) and Q14 (a route to downgrade a warning). - Alarms and restrictions persist too: the saccharin label lasted 23 years, irradiation approvals stalled for 15–20 years, cyclamate is still banned after 55 years, and MMR did lasting damage (themes/T09 §5, from hindsight LL2-02). - The document already holds the material for judging warning quality: - §5.5 items 3 and 6; - hindsight LL2-21 (large independent null studies); - hindsight LL2-19 (a retracted single study); - hindsight LL2-A3 (contributors’ unpublished claims not confirmed); - LL2-02, p. 31 (the swine-flu warning was over-weighted because it fitted prevailing theory). - Warners can be institutionally committed too: the EEA withdrew from the IARC meeting while its editor co-authored the chapter (LL2-21, p. 520; critiques §6).

Fix. 1. Change the §6 introduction to: “…and to those who develop, finance, promote, oversee, warn about, campaign against or restrict it.” 2. Give each entry a one-line Mirror question. For example: - I1: “Is there a gap between what those raising the concern say publicly and what their own data show?” - I3: “…behind the safety claims and behind the harm claims?” - I6: “Do those raising the concern have litigation, funding, reputational or institutional stakes in its being true?” Evidence: Cranor in Milward (hindsight LL2-24); Ozonoff (hindsight LL2-04); LL2-21, fn 11. 3. Add W7, warning quality: replication by independent groups; dose–response; consistency with population trends; reliance on one group’s or unpublished work; whether the warning concerns direction or magnitude. Rate it honestly, as suggestive to moderate. 4. Add W3b, the alarm trap: an early categorical alarm or restriction makes later de-escalation look like an admission of error. Evidence: hindsight LL2-02 (saccharin, irradiation, cyclamate); LL2-21. For contrast, the BSE measures were wound back through open, costed review (hindsight LL1-15). 5. Move §6.11 to the head of section 6 as rule 0.

2. Strength ratings measure how well a mechanism is documented in a failure sample, and the patterns on the precaution side are under-rated by the document’s own definitions#

Location. Rating definitions in §1.4 (l. 50–55); C7 (l. 876–879); S4 (l. 935–938); L3 (l. 822–826); the weighting guide in §5.8 (l. 626–637).

Problem. - The definitions reward “several cases” that have been “not overturned since”. - In a corpus selected because harm occurred, almost any mechanism of missed harm documented in a few chapters will qualify, and a mechanism is seldom “overturned”. - As a result, 40 of the 55 lens entries lead with Strong. - Of the four entries on the error or cost of acting, three are Moderate (T4, C7, S4). T3 is Strong only “(logic)”, with its frequency “contested”. - Two of these are mis-rated on the document’s own definitions: - C7 has cases across both volumes with independent later support. - S4 is marked down as “mostly identified after the fact”. No harm-side pattern gets that discount, although the whole corpus is retrospective. And L3 (regrettable substitution), a subset of the same phenomenon of interventions causing harm, is rated Strong.

Evidence. - C7 cases: - LL1-14, pp. 153–154 (the hormones ban as a “political risk assessment” with “no good evidence” of benefit, conceded by an author who advised the EU); - LL1-09 hindsight (transition costs; a rise in therapeutic antibiotic use); - LL2-02, pp. 28–29 (swine flu: Guillain-Barré cases and deaths, USD 124m and more than 4,100 lawsuits); - LL2-11, pp. 243, 250 (South Africa’s DDT withdrawal); - LL2-16 hindsight (French sugar-beet yields fell from 851 to 649 q/ha); - Jarvis et al. 2022, €3–8bn a year (themes/T05 l. 263; themes/T09 §5); - themes/T05 §4.2 is headed “The reports document costs of precaution themselves, then underweight them”. - S4 cases: - LL2-02, p. 28; - LL2-11, p. 243; - LL2-20, p. 497 (signal crayfish); - LL1-11, pp. 110–111 (MTBE scaled up under a protective mandate); - hindsight LL1-10 (sulphur dioxide cuts unmasked warming); - themes/T07 summary items 8–9; - themes/T09 l. 153.

Fix. 1. Re-rate C7: “Strong that precautionary responses carry material costs, sometimes irreversible and sometimes regressive; moderate on their size relative to benefits.” 2. Re-rate S4: “Strong (existence); moderate (predictability)”. Delete “mostly identified after the fact”, or apply that discount to every entry. 3. Add to §1.4: “A Strong rating means the mechanism is well documented in the failure histories. It does not mean that its presence distinguishes harmful from benign cases.” 4. Add a diagnosticity note to each entry: does the pattern also appear where the warning proved wrong or the concern faded? For example: - latency reasoning (K1, K4) was used to discount null studies in LL2-21 (p. 514); - peripheral warners (W1) included MMR; - I2 already notes that shifting rationales occur in sincere cases. 5. In §5.8, gloss “High” as “high as a question to ask; not evidence that the mechanism is operating”.

3. The project lead’s chapter, and authors’ standpoints generally, are not treated uniformly#

Location. §1.5 (l. 58–60); Appendix A entries for LL2-22 (l. 1064), LL2-03 (l. 1026), LL2-10 (l. 1040) and LL2-13 (l. 1046); the protagonist bullet in §5.6 (l. 607); the §3.6 analysis (l. 280); the §4.4 row “Who steers direction”, hindsight cell (l. 382).

Problem. - §1.5 says LL2-22 is “treated no differently”. - Appendix A flags authors’ standpoints for LL2-21, LL2-18, LL2-04, LL2-24, LL2-20, LL2-09, LL1-04 and LL1-14. It does not flag LL2-22, although the digest records that the chapter: - draws about a quarter of its references from the authors’ own work; - rests partly on a key early warning (Poland et al. 2008) and a benchmark research agenda (Maynard et al. 2006) that Maynard co-authored; - has no dissenting panel; - lets its hazard language escalate from “preliminary” (p. 536) to “rapidly increasing evidence of risks” (p. 539); - does not consider false positives. - Appendix A quotes only “preliminary”, the most cautious of those phrasings. - §5.6 leaves LL2-22 out of its list of protagonist chapters, although the chapter updates the authors’ own 2008 article. - The nanotechnology research-funding statistic, which bears on one of the chapter’s own recommendations, drops three caveats that the hindsight file records. - More broadly, standpoint omissions in Appendix A fall mainly on chapters that hindsight treated kindly (LL2-03 Needleman, LL2-10 Soto, LL2-13 Jobling). That inflates the pattern criticised in issue 4.

Evidence. - LL2-22: - digests/LL2-22.md l. 89–94; - notes/LL2-22.md l. 27, 43, 865–871; - hindsight/LL2-22.md l. 28 (proximity) and l. 69–70, 328–345: - the EHS share fell from about 10% (2016) to about 4% (2020); - the broader “responsible development” line (PCA 5) was 2.3–2.4% in 2023–25 and 1.1% only in the FY2026 request; - “PCA 5 is broader than EHS, so these shares are not strictly comparable”; - the total NNI budget itself fell from $2,122m to a requested $1,449m; - the 2020 National Academies review said the NNI “has performed exceptionally well” on responsible development, and that EHS research “relieved some unfounded early concerns”. - Other omitted standpoints: - digests/LL2-03.md l. 74; - digests/LL2-10.md l. 64, and hindsight/LL2-10.md l. 10 (“much of the later literature ‘vindicating’ it comes from the same scientific network”); - digests/LL2-13.md l. 57; - digests/LL2-26.md l. 31.

Fix. 1. Give every Appendix A entry a standard Standpoint clause: the author’s role, the self-citation share where known, and whether there is a dissenting panel. For LL2-22: “Protagonist update of the authors’ 2008 article; about a quarter of references are the authors’ own, including a key warning co-authored by Maynard (Poland et al. 2008); no dissenting panel; hazard language escalates from ‘preliminary’ (p. 536) to ‘rapidly increasing evidence of risks’ (p. 539); false positives not considered.” 2. Add LL2-22 to the protagonist bullet in §5.6, noting that its self-cited warning on multi-walled carbon nanotubes was vindicated. 3. Rewrite the funding sentence in §3.6 and §4.4 using the hindsight figures and caveats above. Either drop it from the list of “diluted” reforms or pair it with the National Academies’ contrary assessment. 4. Given the §1.5 disclosure, have the LL2-22 entries re-checked by a reader with no connection to the chapter.

4. “Failures cluster where conviction and self-citation were highest” is a selective generalisation#

Location. §5.5 item 4 (l. 596); the §4.8 table row “The reports’ own mindsets” (l. 473).

Problem. - Protagonist conviction and self-citation are just as present in chapters that held up or were vindicated: - lead (Needleman praises his own “seminal” work; core strengthened); - benzene (Infante’s own cohort; core held); - BPA (Soto co-authored many sources; vindicated in the EU); - Grandjean (his own study in his own journal; PFAS vindicated); - ethinyl oestradiol (the authors’ own studies; held); - multi-walled carbon nanotubes (the authors’ own warning; vindicated); - ozone (Farman). - The generalisation was drawn from the failures alone. - As worded, it primes a user of the lens to discount committed warners, which pre-judges how the lens will be applied.

Evidence. - Digests: digests/LL2-03.md l. 74; digests/LL1-04.md l. 3, 80; digests/LL2-10.md l. 64; digests/LL2-26.md l. 31; digests/LL2-13.md l. 57; digests/LL2-22.md l. 94. - The corresponding rows of §5.4. - §5.5 item 3 in the document already names the better discriminator.

Fix. Replace the sentence with: “Protagonist authorship and self-citation are common both in chapters that held and in chapters that failed. The failures share something narrower: key claims rested on one group’s positive findings, or on unpublished work, and were later contradicted by larger independent studies (mobile phones, the GM health sentence, the Chernobyl mortality figures, the PCB paediatric attributions, Annex 3’s laboratory claims). Where protagonists’ findings were independently replicated (lead, benzene, the BPA endpoint, carbon nanotubes), they held.” Rate it Moderate, and amend the §4.8 row to match.

5. Lessons drawn mostly from failures of prevention are rated for use under genuine uncertainty#

Location. §6.1 rule 4 (l. 650); the strength lines throughout section 6; §5.5 item 6 (l. 598); §5.8.

Problem. - Rule 4 rightly separates prevention (failing to act on strong evidence) from precaution (acting under genuine uncertainty), but the ratings never apply the distinction. - Several Strong entries (I1, I2, W2, W4, C1, G2, G3) rest mainly on prevention failures: asbestos after the 1960s, benzene, vinyl chloride, tobacco, beryllium, Minamata, lead. - The lens is meant for emerging technologies, which are the uncertainty and ignorance cases. There the reports’ only prospective test, their own forward warnings, has a mixed record (§5.5 item 6). - The ratings therefore carry weight from the easiest cases over to the hardest.

Evidence. - themes/T09 §12 item 4 (l. 325): many LL1 cases are “failures to act on known harm”, and “the false-positive debate bears mainly on the smaller set of true uncertainty cases”. - LL1-17, Table 17.1, p. 192; LL1-04, p. 46. - Marchant’s point that genuine ignorance cannot be prevented (critiques §4).

Fix. 1. Tag each entry’s evidence by case type: - [K]: known harm, a prevention failure; - [U]: genuinely uncertain at the time; - [F]: forward warnings in LL2 Parts B and C. 2. Give a separate strength for the [U] and [F] evidence. 3. Add to §5.8: “For emerging technologies, weight entries by their [U]/[F] support.”

6. Interests behind restriction and alarm, including protectionism, are missing#

Location. The §4.3 table (l. 354–365); I7 (l. 799–802); §5.7 (l. 613–624); compare §3.7 item 6 (l. 289).

Problem. - §3.7(6) rightly faults the reports because “every example of misuse given is uncertainty deployed against regulation”. The document then repeats the pattern: - The §4.3 table lists ten interest-driven patterns, all of them on the side of producers or promoting states. - I7 treats a commercial interest in restriction (“profited from the alternative”) as a welcome accelerator of action. It never asks whether competitors, makers of substitutes, domestic producers or trade interests can push restriction beyond what the evidence warrants. - The word “protectionism” appears nowhere, although both the corpus and the critiques file contain examples.

Evidence. - LL1-14, pp. 150, 153–154: the hormones ban was taken against two expert committees and driven “principally” by public concern; it was settled by beef quotas, not science (hindsight LL1-14). - critiques §3.3: Majone on aflatoxin standards and African exporters; hormones as the standard alleged case. - critiques §9.4(2) (l. 280–283): ask of any precautionary response whether it “can be captured to protect incumbents or markets”. - LL2-03, p. 60: GM wanted lead out of petrol to protect catalytic converters. - LL2-27, p. 647: responsible behaviour came mostly from downstream sellers. - LL1-11, pp. 110–111: MTBE was scaled up under a mandate.

Fix. 1. Add a §4.3 row, “Interests behind restriction and alarm”, rated Moderate, citing the cases above. 2. Add lens entry I9. Ask: “Who gains from restriction (competitors, makers of substitutes, domestic producers, advocacy or research programmes), and is that interest shaping the evidence or the threshold? Would the measure look the same if it were applied equally to domestic and foreign, incumbent and new sources?” 3. Add to §5.7: “No analysis of the interests on the side of alarms and restrictions.”

7. The Fukushima figure of 2,351 deaths is attributed wholly to evacuation and used as the flagship cost of precaution#

Location. §5.2, the asymmetry paragraph (l. 518); Appendix A LL2-18 (l. 1056); §4.5 counter-evidence (l. 411); T4 evidence (l. 759); S4 evidence (l. 937); the §4.1 table, surprise row (l. 313).

Problem. - The prefecture’s count covers the combined earthquake, tsunami and nuclear disaster. The hindsight file found no official split between them. - The document also counts emergency protective action after an accident as a cost of “precaution”. That is not the same thing as precaution about deploying a technology before harm is known, which is the reports’ subject. - Dropping the caveat makes the rebuttal of the asymmetry argument look stronger than the evidence allows.

Evidence. - hindsight/LL2-18.md l. 147–148: “The prefecture’s count covers the combined disaster, and I found no official split.” - themes/T09 l. 153 keeps the caveat. themes/T05 §4.2 (l. 262) drops it, so the error entered at the theme stage.

Fix. Use: “Fukushima Prefecture counts 2,351 disaster-related deaths from stress and ill-health among evacuees. The count covers the combined disaster, and there is no official split between the nuclear evacuation and the earthquake and tsunami (hindsight LL2-18). These are costs of emergency protective action after an accident, not of precaution before deployment, but they show that protective measures can cause irreversible harm.” Lead the rebuttal of premise 3 in §5.2 with swine flu (LL2-02, p. 28), which is a clean case of ex ante precaution.

8. Findings that favour the false-alarm review are under-credited#

Location. The §5.2 rating paragraph (l. 516); the §3.5 analysis (l. 268); the LL2-28 row of §5.4 (l. 587).

Problem. - (a) A finding became a norm. themes/T09 rated “Claimed false alarms often turn out real or unresolved” as moderate–strong. §5.2 turns this into a procedural norm (“deserve the same scrutiny”), so the finding disappears. The movement of about 12 checked cases towards harm and about 3 towards reassurance survives only as a bullet, followed by “The check was selective”. - (b) The rarity claim omits the balancing verdict. “False positives are rare relative to false negatives: asserted and contested” leaves out hindsight LL2-28’s verdict that the underlying point “has not been overturned, and many alarms have since been confirmed”. - (c) “Do not reproduce” is inaccurate. It says Chapter 28’s three most-cited statistics “do not reproduce”. In fact: - “4 of 88” has never been re-analysed, so it is unreplicated, not failed; - the 1% figure is unverifiable, but its direction is “very likely right”; - “half of all articles” is overstated about fourfold (verdict: partly held up). - (d) A caution is missing. Section 5 lacks T09’s warning that the counterweight can be overdone. Selection undermines claims about frequency, not the mechanisms documented case by case. The critics’ lists were showcases too. Most Part A false negatives are now mainstream public-health history.

Evidence. - themes/T09 l. 135 and §12.3 (l. 320–323). - hindsight/LL2-28.md l. 33–34, 69, 74, 120, 306, 345. - themes/T09 §1 item 3: the check “vindicates the chapter against its critics”.

Fix. 1. Add the rating: “Alleged false alarms in critics’ showcase lists mostly proved real or unresolved: moderate–strong (a selective check: about 12 of about 18 moved towards harm, about 3 towards reassurance).” 2. Reword the rarity rating: “Unmeasured: neither established nor overturned. Movement since 2013 leans the reports’ way on a selective sample.” 3. Replace “do not reproduce” with: “are fragile. ‘4 of 88’ is unreplicated and was restated without its caveats; the 1% figure is unsourced, though probably right in direction; ‘half of all articles’ is overstated about fourfold.” 4. Add a three-sentence version of T09 §12.3 to §5.1 or §5.5.

9. The table of external critiques is asymmetric in three directions#

Location. §5.3 (l. 522–535).

Problem. - (a) It tilts towards the reports. The row “‘Paralysis’ and incoherence … largely met by the EEA’s two-sided definition” credits the definition’s wording, not the reports’ practice. The document itself shows the gap: - LL2-27 argues only for more precaution (§5.6); - LL2-02 keeps trade-offs out of the error ledger (§5.2); - LL2-28, p. 673 tilts policy by default.

The row also omits the residual force of the critique, which critiques.md records. - (b) It tilts against the reports. The Cox critique is rated Strong, and the risk–risk critique Moderately strong, from abstract-only or metadata-only access, without the authors’ replies: - Hansen et al. 2007b [meta]; - Hansen and Tickner 2008 [meta], who argue that trade-offs are often overstated and avoidable through alternatives assessment.

§1.3 promises that access caveats “carry into this document”. In this table they do not. - (c) Standpoints are treated unevenly. Authors’ standpoints are disclosed throughout (Gee’s NGO background, Cranor, Ozonoff, Hardell’s funding). Critics’ standpoints are not, apart from the European Risk Forum. That fails the document’s own check in §6.11(2).

Evidence. - external/critiques.md l. 70: “Its residual force: once both sides must be weighed, much of the principle’s work is done by the weighing.” - external/critiques.md l. 81 and l. 154–158 (access status). - §6.11 item 2.

Fix. 1. Reword the paralysis row: “Met in the definition’s wording, not in the synthesis chapters’ practice. Residual force: the weighing does the principle’s work, and the definition gives no method for it (critiques §3.1).” 2. Add an access column ([full], [abstract], [meta]) and summarise the replies’ positions. The rating on the false-positive method can stay Strong, but attribute it to the document’s own reading of LL2-02, pp. 18–19 and 33, rather than to Cox, whose paper was seen only in abstract. 3. Either add a standpoint note for each critic at the level used for the report authors, or say plainly that critics’ standpoints were not examined.

10. Smaller over-corrections in wording#

Location and fix. - §4.8, In brief (l. 461). - Text: “Sincere belief did at least as much damage as bad faith.” - Problem: this compares magnitudes that were never measured. themes/T08 l. 30 supports only the claim that sincere error was common, and the documented bad-faith cases include some of the largest harms (lead, tobacco, asbestos). - Fix: “Sincere belief was at least as common a source of delay as bad faith, and did serious damage without any deception.” - §4.1, counter-evidence (l. 318). - Text: “contradicted by the replication crisis and by the reports’ own mobile-phone warning”. - Problem: hindsight LL2-26 (l. 80, 349, 355) finds false positives common “at the level of published findings”, and gives a two-sided restatement that “keeps its precautionary point”. §5.2 already has the calibrated version. - Fix: “too one-directional: low power and publication bias also produce false positives in published findings.” - §5.6 (l. 603). - Text: “its main critics appear only in the bibliography.” - Problem: the audit adds that many supportive works are also uncited (“the trimming was general, not aimed only at critics”), and that Gee does rebut the critics’ reading of the North Sea Declaration (p. 657) (notes/LL2-27.md l. 185, 250). - Fix: add both points. - §5.4, LL2-00 row (l. 560). - Text: “‘irresponsible corporations’ … weakened”. - Problem: the hindsight verdict is “partly held up”. The legal-remedy half was “borne out strongly”; the generalisation is the weaker part (hindsight/LL2-00.md l. 150). - Fix: “partly held up; the generalisation is weakened by harms caused by the state.”

11. Post-2013 policy change is described in one side’s evaluative vocabulary#

Location. - §3.6 analysis (l. 280): “asked little of incumbents”; “Several adopted reforms were later diluted”. - §3.6 table (l. 276): open access “was also turned against research”. - §4.6 counter-evidence (l. 434): “used rigour language to discount epidemiology”. - Appendix A LL1-03 (l. 990): “Threshold and hormesis lobbies failed scientifically and then gained ground politically”.

Problem. These judgements may be right, but each takes a side in a live dispute. - The uptake gradient. It is T10’s own analysis, rated moderate. “Asked little of incumbents” names incumbent resistance as the cause, but other explanations fit too: cost, feasibility, thin evidence for the instruments themselves (bonds and no-fault schemes; see §5.7), or public preference. - The “diluted” list. It contains only protective reforms that were weakened. Elsewhere the document approves of evidence-led relaxation, such as the BSE over-thirty-months rule replaced after a costed review. Changes should be classified by whether they followed evidence, not by their direction. - Executive Order 14303. Hindsight LL2-27 (l. 143–151) records: - the defenders’ case; - critics, including Michaels, an LL2 author; - Science’s editor blaming the scientific community’s own “sluggishness and defensiveness”.

The document gives only the critics’ reading. - LL1-03. - “Failed scientifically” refers to the US regulator’s denial of petitions, which was a regulatory ruling. - The ICRP’s 2025 memorandum concedes that optimisation means “the appropriate level of protection, not simply the lowest dose”, answering “a common criticism of how ALARA is applied” (hindsight/LL1-03.md l. 51, 126, 229–232). - “Lobbies” is Lambert’s word (p. 35) and should be quoted as his.

Fix. 1. Describe these changes neutrally and state the rival explanations. 2. Adopt T09’s sentence (§12.6): policy moving against the reports’ prescription “is not evidence against their diagnosis”, nor is it evidence for it. 3. Include evidence-led relaxations alongside the dilutions.

12. Two calibration fixes that remove a one-directional impression#

What to keep, so the fixes do not over-correct#

The following passages are well balanced and should survive revision unchanged: - §3.3 analysis (Chapter 17 drops Chapter 16’s caveats and ignores the hormones chapter); - §4.2 on warnings from inside industry; - §4.3’s three kinds of explanation and its “whoever controls exposure and information” generalisation; - §4.4’s complications on substitution; - §4.5’s counter-evidence paragraph; - §5.1; - the conditional restatement of the asymmetry argument in §5.2; - the “fairer than their reputation” paragraph in §5.6; - the rules in §6.1.

The fixes above are meant to make the lens two-sided and its ratings honest. They are not meant to make it sceptical of warnings by default. On the reports’ own record, the historical false negatives are real and well documented, and most critics concede that point (critiques §9.3).