Bias audit, scope B: proposals for 03-late-lessons-and-huang.md, sections 5–12 and Appendix A#

Working file, 26 September 2026. It proposes wording-level edits only. No change to the ranking, the structure or the findings. Line numbers refer to 03-late-lessons-and-huang.md as it stands (identical to working/bias-audit/03-before.md). “Official transcript” is working/text/NYT-official-transcript.txt, cited by its printed page number.

1. Quantitative check: qualifiers on criticisms and on concessions#

What was counted. - A softening qualifier on a criticism is an inline clause, parenthesis or appended sentence that is attached to a specific statement adverse to Huang and narrows its scope, lowers its certainty or offsets it. Examples: “though…”, “(context unknown)”, “in qualified form”, “no inference about motive”, “structure, not conduct”, “within search limits”, “No one else has such a method either”. - A qualifier on a concession is the same thing attached to a statement favourable to Huang, including the labelled “Limit:” device. - Labelled blocks are counted separately: “Limit:” in 6.1 and 6.2, and “In his favour:” in 7.1. - Excluded: confidence, strength and transfer columns in tables (these are calibration metadata); document-level method cautions (5.1 bullets, 5.2 “not summed”); Mirror statements about critics; and modal verbs (“can”, “may”). - The count was made by hand. Treat the totals as ±10%.

Results.

Reading. - The raw ratio overstates the asymmetry. Sections 7–10 contain more criticisms than concessions, so the fairer comparison is the like-for-like one between 6.1 and section 7. On that comparison, each criticism carries about one and a half times as many qualifiers as each concession, rising to about three times in the top five. - The asymmetry also shows in placement. Qualifiers on concessions sit in a visible, labelled Limit: slot. Qualifiers on criticisms sit inside the sentence, and are sometimes stacked: DuPont in 7.1 #3 carries five, and I5 carries three in 7.1 #4 and again in 5.3. - Most of the qualifiers are earned. They carry documented context (press-reported, context unknown, the 2030 horizon, case type, an advisory role), and many were added by the earlier balance review after checking the sources (B2, B5, B7, B9, B13, B20, F7). - The proposals below target the minority that are unearned, duplicated or one-sided. They also flag criticisms that go beyond the evidence, since the test runs in both directions.

2. Proposals#

Proposals are in document order. Direction says which way each edit moves the text: towards less sympathy for Huang; towards Huang (a criticism that goes too far, or a Mirror that is too thin); symmetry; or neutral accuracy.

1. §5.4 (with echoes in §5.2 and Appendix A): “No false precision”#

• Location:
• §5.4, l. 494: “- No false precision: he refuses probabilities without a model.”
• §5.2 table, Knowledge row, l. 457: “He refuses false precision (Hinton’s point probability);”
• Appendix A, 4.1 row, l. 959: “Refuses false precision;”
• Rubric: b. Direction: less sympathetic.
• Problem: The concession is stated without its required exception. His own “0% chance” (by 2030) is a point figure with no stated basis. The document says so in §6.1 item 2 and in §7 challenge 2, so §5.4 contradicts both. The source finding (“not false precision”) concerns treating ignorance as calculable risk, and it was made about Hinton’s figure.
• Evidence:
• D01 l. 266, l. 289: the charge “in the sense of false precision” concerns treating ignorance as risk.
• D02 l. 24 (“Neither can ‘0%’”); l. 309 (“‘0%’ is a point estimate of zero”); l. 365 (“Hinton’s 10–20% and ‘0%’ both fail W7”).
• LA1 l. 547: Altman’s approach is “closer to the reports’ approach than either Huang’s ‘0%’ or Hinton’s ‘10 to 20’ per cent”.
• Replacement text:
• §5.4: “- No false precision of the kind the reports name (treating ignorance as calculable risk): he refuses others’ probabilities without a model, though his own “0% chance” of the end of the world by 2030 is also a point figure with no stated basis (section 7, challenge 2).”
• §5.2: “He refuses others’ false precision (Hinton’s point probability), though not in his own “0%”;”
• Appendix A: “Refuses others’ false precision (his own “0%” aside);”
• Confidence: medium-high.
• Note for scope A: §4.1 (l. 203, “The charge of false precision does not hold…”) is the origin and may need the same clause.

2. §5.2 table, Interests row: I1 absence listed as support#

• Location: l. 459: “absence of any documented private–public gap (I1) or knowledge-avoidance by Nvidia (I6)”.
• Rubric: b. Direction: less sympathetic.
• Problem: The absence is listed as support for Huang without the caveat that §5.4 gives two paragraphs later.
• Evidence: §5.4 of 03 itself (“absence proves little at this stage”); hypotheses.md §3, “The lens” (I1: “its absence here proves little either way”); D04 l. 122.
• Replacement text: “absence of any documented private–public gap (I1; weak evidence, since such gaps surfaced mainly through litigation) or knowledge-avoidance by Nvidia (I6)”.
• Confidence: low-medium. Optional.

3. §5.5: the Mirror list is thinner than the supporting files allow#

• Location: l. 500: “shared dependence on lab-generated indicators; costs and consent”.
• Rubric: g. Direction: towards Huang (fuller Mirror on the critics).
• Problem: Two Mirror findings that the supporting files document are missing from the list:
• W4 turned on the warners, which appears only later, in §9.4;
• the I5 pattern on the critics’ side: warners who assess the hazard also sell the remedy.
• Evidence:
• leaders-comparison.md §5, W4: “accurate as description, and it is W4 turned on the warners. But W4 transfers weakly here… Here it is a subjective probability”.
• leaders-comparison.md §5, I5 Mirror: “Warners who campaign on a hazard also assess it and sell the remedy: Anthropic’s research, Suleyman’s ‘humanist’ AI, OpenAI’s defensive AI”.
• LA2 l. 215.
• Replacement text: “shared dependence on lab-generated indicators; warning while building (the labs asking to be slowed down are building the most compute, W4 turned on the warners, though W4 transfers weakly where “knowing” is a subjective probability; section 9.4); assessing the hazard and selling the remedy (warners who campaign on a hazard also assess it and sell the remedy, as with the labs’ safety research and defensive AI: the I5 pattern on the critics’ side); costs and consent”.
• Confidence: medium. The same caveat stays: less depth than for Huang.

4. §6.1 item 1, Limit: “doom narratives drive opposition”#

• Location: l. 526: “and his claim that doom narratives drive opposition to data centres is unverified, with no direct evidence found (FC C213).”
• Rubric: h. Direction: towards Huang.
• Problem: “Drive” overstates what he said. He first lists the industry’s own failures (communication, water, power, taxes, setbacks, being a good neighbour) and only then says the narratives “are not helping”.
• Evidence: official transcript, pp. 51–52 (the [1:40:15] turn: “…we could have done so much better of a job communicating with the communities… And then, of course, all of our narratives about the end of the world are not helping”); HA A6 (02 l. 1004: “He does not say fear is the main obstacle”).
• Replacement text: “and his claim that doom narratives add to opposition to data centres, which he lists after the industry’s own failures to communicate and be a good neighbour, is unverified, with no direct evidence found (FC C213).”
• Confidence: medium.

5. §6.1 item 9, Limit: Box 20.4 “applies directly, against Huang”#

• Location: l. 534: “Limit: the box concerns governments, and applies directly, against Huang, to pre-empting state rules.”
• Rubric: h (with e). Direction: towards Huang.
• Problem: The limit overstates the case against him and conflicts with the document’s own §10.3. That section says pre-emption “fits Box 20.4 by extension”, and that Huang’s December 2025 statement, which pairs one standard with “a federal AI regulation”, “is consistent with G5”. The thematic comparison attached this charge to Huang only at medium-low confidence.
• Evidence:
• D07 l. 375 and l. 668: “Huang has not advocated pre-emption before a federal framework… the Box 20.4 and Shizuoka charge applies to the administration’s position; it is attached to Huang only at medium–low confidence”.
• D04 l. 59.
• 03 §10.3.
• Replacement text: “Limit: the box concerns governments. By extension it bears on pre-empting state rules before any federal framework exists, which is what current federal action amounts to (section 10.3); Huang’s only documented statement on the question pairs one federal standard with “a federal AI regulation” (December 2025), so the point attaches to him only weakly.”
• Confidence: high.

6. §6.1 item 12: “Caution is not costless”#

• Location: l. 537: “12. Caution is not costless (L6).”
• Rubric: b (with e). Direction: less sympathetic.
• Problem: The heading claims more than L6’s evidence shows. The cited meta-analysis finds “statistical insignificance”: it undercuts the reports’ claim that precaution stimulates innovation, but it does not show that caution is costly. The lens record says L6 “fails on both sides”. Costs of precaution belong to C7 (item 1), not to L6.
• Evidence: LA4, L6 row (“Fails on both sides. ‘False choice’, the EEA’s ‘does not stifle’ and Amodei’s… each deny a trade-off”) and l. 402 (“Neither is established”); D05 l. 117, l. 295; LLA l. 428 (Cohen and Tubb 2018).
• Replacement text: “12. The reports’ claim that precaution does not hold back innovation is unestablished (L6).” Keep the body and the Limit as they are.
• Confidence: medium-high.

7. §6.3 Analysis: “well founded… on the moral hazard” and “several points”#

• Location: l. 559, the whole paragraph beginning “Analysis. Late Lessons does not tell Huang that AI must be slowed.”
• Rubric: b (with c). Direction: less sympathetic.
• Problem: There are two issues.
• Late Lessons cannot tell him he is “well founded” on something that, as the same sentence says, lies “beyond anything the reports examined”. Item 16 and §2 call the moral-hazard argument “partly reasoned”, with a limit (the less careful rival).
• The criticism side is compressed to “several points”, although §7 ranks twelve challenges, the top two at high confidence.
• Evidence: 03 §6.1 item 16; LA2 l. 215 (“Huang’s resistance to coordinated pacing is partly reasoned”); HA §7.4 item 2 (02 l. 889: “Limit: the argument does not reach the strongest form of the labs’ case”); 03 §7 table, ranks 1–2.
• Replacement text: “Analysis. Late Lessons does not tell Huang that AI must be slowed. It tells him he is well founded on the costs of alarm and of precaution, on fixing known failures first, on the suspicion that restriction can entrench incumbents and on refusing liability relief; and, on a point the reports never examined, his moral-hazard argument against making safety a collective duty is reasoned, though it does not answer the case of a less careful rival. It tells him he is poorly supported on his claim that the critics’ warnings have failed, and on the institutional core of his own programme (who checks the builder’s containment and tests, what standard of proof governs whom, who holds the trigger), which section 7 sets out.”
• Confidence: medium-high.

8. §7.1 #2: “even for cheap steps such as incident reporting”#

• Location: l. 586: “For new public rules it is high and undifferentiated, even for cheap steps such as incident reporting;”
• Rubric: h. Direction: towards Huang.
• Problem: The sentence reads as though he applied the bar to incident reporting. He was not asked about it. The point is an inference from his general bar (demonstrated harm plus a demonstrated gap), and the source says so.
• Evidence: D03 l. 144 (“A cheap step such as mandatory incident reporting faces the same bar as a licence. He does not address such steps”); D03 l. 354.
• Replacement text: “For new public rules it is high and undifferentiated: demonstrated harm plus a demonstrated gap, a bar that makes no exception for cheap steps such as incident reporting, though they were not put to him;”
• Confidence: medium.
• Note for scope A: §4.3, l. 245 (“In his model mandatory incident reporting faces the same bar as licensing”) is the same inference and is already worded as a feature of his model; no change needed there.

9. §7.1 #3: the DuPont outcome “is not relied on here”#

• Location: l. 588: “(How the pledge played out is outcome, not structure, and is not relied on here: the companion analysis’s hindsight check finds it honoured only after global loss had been formally attributed.)”
• Rubric: a (with e). Direction: less sympathetic.
• Problem: This is the fifth qualifier on one precedent, after “structure, not conduct”, “one case”, “moderate weight” and “no bad faith alleged”. The balance review (B7) asked for it so that outcome would not be used to infer conduct. But how a threshold judged by the pledger compared with a public threshold under the same evidence is the mechanism itself (T1; rule 7’s comparator), not evidence about motive. Leaving it out discards the one piece of evidence that shows the structural point working.
• Evidence:
• LL1-07, p. 80 (text extract working/text/chunks/LL1-07.txt l. 410–423): “It was to deny the existence of reputable evidence until 1986. Nevertheless… the industry… gave substantial financial support to… research into the ozone problem”.
• LL1-07, Table 7.1, p. 83: “1977 United States bans CFCs in aerosols based on ‘reasonable expectation’ of damage”.
• D07 l. 296 and l. 320 (“The DuPont analogue: moderate (one case)”).
• Replacement text: “(How the trigger behaved bears on structure, not conduct: by the chapter’s own account DuPont denied the existence of “reputable evidence” until 1986, while the United States had banned CFCs in aerosols in 1977 on a “reasonable expectation” of damage (LL1-07, pp. 80, 83). A threshold judged by the party that bore its cost sat well above a public one, which is T1’s point. The chapter also records that industry funded ozone research throughout.)”
• Confidence: medium. This partly reverses review fix B7, so it is for the applier’s judgement. Keep “one case, moderate weight, no bad faith alleged”.

10. §7.1 #4: Huang’s link to the promoting state#

• Location: l. 590: “his link to that state is an advisory seat and an alignment of interest (I10), and no inference about motive follows.”
• Rubric: d (with a). Direction: less sympathetic.
• Problem: The rule against inferring motive is applied correctly, but the description of the link leaves out documented, disclosed political action on the rules the state would apply. LL2-25, which this document cites in §11.2, treats political action as open to stricter scrutiny than business action. Stating it is structure, not motive.
• Evidence:
• D04 l. 45, l. 190 (“Nvidia’s political actions are documented and disclosed: Huang’s support for a federal standard in place of state AI laws… lobbying on chip-security and export bills”), l. 289.
• hypotheses.md §5, “The lens” (“LL2-25’s distinction between business and political actions licenses more scrutiny here than anywhere else… these entries bear on who should hold the gate, not on Huang’s sincerity”).
• 03 §11.2 (LL2-25, p. 615).
• Replacement text: “his link to that state is an advisory seat, an alignment of interest (I10), and open, disclosed political action on the rules it would apply (support for one federal standard in place of state laws; lobbying on export-control and chip-security bills), which LL2-25 would hold to a stricter standard than business actions (p. 615). No inference about motive follows.”
• Confidence: medium.
• Note for scope A: In brief item 4 carries the same phrase (“an advisory seat and an alignment of interest”).

11. §7.2 #7: “Three explanations… around a fixed conclusion”#

• Location: l. 597: “Three explanations of the labs’ warnings in a week around a fixed conclusion;”
• Rubric: h. Direction: towards Huang.
• Problem: The criticism is stated plainly here, but §5.1 of the same document records that review found this marker weak. The explanations are of others’ motives, one is charitable, and they coexist rather than replace one another. The qualifier is required but applied in only one place.
• Evidence: 03 §5.1, third bullet; hypotheses.md §3, qualifiers (“one of them (‘too much humility’) is charitable… This is a weak flag”).
• Replacement text: “Three explanations of the labs’ warnings in a week around a fixed conclusion (a weak marker: they explain others’ motives, one is charitable, and they coexist rather than replace one another; section 5.1);”
• Confidence: medium-high.

12. §8.2 pattern 3: the count of “hurt”#

• Location: l. 621: “Nine of his eleven uses of “hurt” are aimed at talk about AI, though four of those nine concern”
• Rubric: h (accuracy). Direction: neutral.
• Problem: The count comes from the machine transcript and includes a stutter (“It hurts. It actually hurts…”). §1.4 removes stutters, and the official transcript has none.
• Evidence: official transcript. On talk about AI (8):
• p. 29: “hurts their reputation”, “hurts their character”, “hurts employee morale”;
• p. 30: “Those predictions are hurtful”, “Is that helpful or hurtful to society?”, “terribly hurtful”, “helpful or hurtful, if it were to happen?”, “It’s hurtful”.

Elsewhere (2): p. 49 “it hurts the whole industry”; p. 53 “they’ve got to hurt you first”. - Replacement text: “Eight of his ten uses of “hurt” in the official transcript are aimed at talk about AI, though three of those eight concern” - Confidence: high on the facts; low importance.

13. §8.4, the frame paragraph: “What is missing fits the frame”#

• Location: l. 650: “What is missing fits the frame: no explicit probabilistic reasoning about rare severe risks, no game theory of coordination, no analysis of distribution.”
• Rubric: h. Direction: towards Huang.
• Problem: The source limits this list to the interview, and the document credits his moral-hazard argument elsewhere (§6.1 item 16). Outside the interview, his view on jobs is “considered and conditional”.
• Evidence: hypotheses.md §2.1 (“the reasoning absent from the interview fits the frame”); §4, H3 (“Outside the interview his view on jobs is considered and conditional: … ‘net generation of jobs doesn’t guarantee that any one human doesn’t get fired’ (Acquired, 2023)”).
• Replacement text: “What is missing from the interview fits the frame: no explicit probabilistic reasoning about rare severe risks, no game theory of coordination beyond his moral-hazard argument, and no analysis of distribution (elsewhere, on jobs, he is more conditional: “net generation of jobs doesn’t guarantee that any one human doesn’t get fired”, Acquired, 2023).”
• Confidence: medium.

14. §8.4, the lag paragraph: “his archive is the richer one”#

• Location: l. 656: “and on that side his archive is the richer one.”
• Rubric: b. Direction: less sympathetic.
• Problem: The paragraph identifies where the reports bear on him most directly, then ends on a comparative that no supporting file makes. The hypothesis file says “Both archives are showcases”.
• Evidence: hypotheses.md §7, H6c (“Rule 0 asks whether examples are ‘a sample or a showcase’. Both archives are showcases”); 03 §6.1 item 1.
• Replacement text: “and on that side his archive, though itself a showcase (rule 0), holds cases the reports’ own ledger left out, radiology among them (section 6.1, item 1).”
• Confidence: medium-low.

15. §8.4 Verdict: “nothing in the record contradicts it”#

• Location: l. 658: “”Sincere” is the reading the reports’ rules require absent documents, and nothing in the record contradicts it;”
• Rubric: f (with d). Direction: less sympathetic.
• Problem: Sincerity is assumed under the rules, not tested. “Nothing in the record contradicts it” goes one step further than the evidence. The hypothesis file records a difference by audience on data-centre opponents: the All-In echo set against “then so be it” [1:40:15], a pattern that H4 predicts and H1 does not. Section 8’s own introduction says public words are imperfect evidence.
• Evidence: hypotheses.md §1.3; §5, H4 evidence; the §8 table note (“No row is a verdict on sincerity”); 03 §8 introduction.
• Replacement text: “”Sincere” is the reading the reports’ rules require absent documents, and no document contradicts it, though under those rules it is assumed rather than tested, and public statements in a live policy fight are imperfect evidence either way (section 8, introduction);”
• Confidence: medium.
• On the verdict itself (rubric f): see section 3 below. No change to “partly” is proposed.

16. §8.5 step 5: the insulation account#

• Location: l. 672: “local opposition to data centres. The costs of AI harm to third parties arrive slowly or not at all.”
• Rubric: d. Direction: towards Huang (and accuracy).
• Problem: There are two issues.
• “Local opposition to data centres” is listed as a cost of alarm reaching Nvidia. That rests on Huang’s own causal claim, which the document elsewhere calls unverified (FC C213).
• The Mirror gives the labs an offsetting clause (“though it has also cost them”), while Huang’s step gets none, although the supporting file records a feedback channel he names himself.
• Evidence: hypotheses.md §2.2 (“‘when they don’t build safe products, it hurts the whole industry’ [1:37:36] recognises a reputational commons, though he does not carry it over to pacing”); FC C213 (02 l. 818); 03 §8.5 Mirror.
• Replacement text: “local opposition to data centres (which he links in part to doom narratives, a link not verified, FC C213). The costs of AI harm to third parties arrive slowly, and mainly through the industry’s reputation, a channel he recognises (“when they don’t build safe products, it hurts the whole industry” [1:37:36]) and one LL2-25 counts as leaky (pp. 608–612).”
• Confidence: medium-low.

17. §9.1, Anti-doomerism row: imputation of motive dropped#

• Location: l. 694: “where Huang calls the labs’ narrative of helplessness “a deflection of blame” (while appearing to endorse the opening of their pacing statement [51:20])”.
• Rubric: a (with d). Direction: less sympathetic, stated symmetrically.
• Problem: The source row said “Motives are imputed by Huang (while disclaiming knowledge of their beliefs), Altman (in part), Musk (to one warner), Mensch and Andreessen”. 03 kept a concession (“appearing to endorse”) and dropped the imputation, which is documented and which §6.1 item 15 mentions (“ulterior reasons”).
• Evidence: leaders-comparison.md §3, Anti-doomerism row; D02 l. 17, l. 286; hypotheses.md §4, H3 “Evidence against”; official transcript, p. 29 (“I can’t talk to you about what they believe”).
• Replacement text: “where Huang calls the labs’ narrative of helplessness “a deflection of blame” (while appearing to endorse the opening of their pacing statement [51:20]), and elsewhere speaks of “ulterior reasons” (CBS, reported) while saying he “can’t talk to you about what they believe” [56:48]. Altman (in part), Musk (of one warner), Mensch and Andreessen also impute motive to warners”.
• Confidence: medium.

18. §9.1, Tail risk row: “Outlier in form and emphasis”#

• Location: l. 699: “Outlier in form and emphasis.“
• Rubric: a. Direction: less sympathetic.
• Problem: The horizon caveat (B5) applies to the “0%” figure, which concerns 2030. The same cell reports a substantive difference that has no horizon: he twice denied that humanity could lose control of AI, where even Zuckerberg names it. The source classes the difference as one of substance. The revision turned “Outlier” into “form and emphasis”, a narrowing that B5 did not ask for.
• Evidence: leaders-comparison.md §3, tail-risk row (“Outlier… Kind of difference: Substance”) and §6 (“Most frontier developers… treat loss of control as a real question… ‘0% chance’ [is a] minority view among model builders”); official transcript, p. 29 (“I don’t think you believe that.” / “No.” / “I think you don’t believe it at all.” / “No.”); revision log B5.
• Replacement text: “Outlier, in substance as well as form.” Keep the rest of the cell, including the horizon clause.
• Confidence: medium-high.

19. §9.1, Chips for China row: “Huang accepts a US-first allocation rule”#

• Location: l. 701: “Huang accepts a US-first allocation rule [1:37:36].”
• Rubric: b. Direction: less sympathetic.
• Problem: Offered as a softener, the concession needs its limit: the rule he welcomes is one Nvidia says it already follows, so it costs nothing.
• Evidence: official transcript, p. 50 (“I’m delighted by that. That’s no problem. We do that naturally, anyway.”); hypotheses.md §3, “What would discriminate” (“so the rule he welcomes would cost Nvidia nothing”).
• Replacement text: “Huang accepts a US-first allocation rule [1:37:36], one he says Nvidia already follows (“We do that naturally, anyway”).”
• Confidence: high.

20. §9.2 pattern 5: “his tone has sharpened”#

• Location: l. 711: “Nvidia backed licensing of high-risk uses in 2023, and Huang now says “We don’t need any new laws” (Dreamforce, reported);” and “Huang’s safety model has been stable since 2023; his tone has sharpened.”
• Rubric: b and h. Direction: both.
• Problem: There are two issues, one in each direction.
• The summary understates the change. The Huang analysis’s verdict on regulation is “Consistent in principle, hardened in practice”, not only a change of tone.
• The Dreamforce quotation appears without the interview counterpart [47:10]. Review fix B3 asked for the two to be paired at every use, and they are paired in §3.4 and §4.7 but not here.
• Evidence: 02 l. 1068 (“Consistent in principle, hardened in practice”), l. 494 (“His position has also hardened at company level”); D07 l. 103; revision log B3; hypotheses.md §3 qualifiers (the 2023 testimony was Bill Dally’s).
• Replacement text:
• First sentence: “Nvidia backed licensing of high-risk uses in 2023 (through its chief scientist’s Senate testimony), and Huang now says “We don’t need any new laws” (Dreamforce, reported), though in the interview “I’m not against laws and regulations… I’m against currently the distraction” [47:10];”
• Second sentence: “Huang’s safety model has been stable since 2023; his regulatory position is consistent in principle but has hardened in practice, and his tone has sharpened.”
• Confidence: medium-high.

21. §9.5: “(though single firms did act)”#

• Location: l. 743: “(though single firms did act);”
• Rubric: a. Direction: less sympathetic.
• Problem: The softener is used without the limit that §6.1 item 10 attaches to the same fact.
• Evidence: 03 §6.1 item 10 (“Limit: July was an easy case, with a legible endpoint and a victim with a voice”); §10.3, W5 bullet.
• Replacement text: “(though single firms did act after July, in what section 6.1 calls an easy case);”
• Confidence: low-medium.

22. §10.3, “Uptake follows a gradient”: “the reports’ record predicts reform will stall”#

• Location: l. 780: “The public layer for AI is mainly informational and voluntary, where the reports’ record predicts reform will stall.”
• Rubric: e (a Late Lessons finding stated more strongly than its evidence). Direction: towards Huang.
• Problem: There are two issues.
• “Predicts” conflicts with rule 1: a pattern is a reason to look harder, not a prediction.
• The gradient finding (moderate) says that informational reforms advanced most, while reforms that move money or power moved least. The sentence runs the two together.
• Evidence: LLA l. 318 (“Uptake followed a gradient… cheap to adopt… went furthest… moved least”, T10 P3; moderate); D12 l. 155; 03 §1.3, rule 1.
• Replacement text: “The public layer for AI is mainly informational and voluntary. On the reports’ record (moderate), reforms of that kind advanced most easily, while those that move money or power, such as binding gates and independently generated evidence, moved least.”
• Confidence: medium.

23. §10.6 Analysis: “which is why the reports press on him hardest”#

• Location: l. 816: “Huang states that model most plainly and defends it most fully, which is why the reports press on him hardest.”
• Rubric: c. Direction: less sympathetic.
• Problem: The sentence gives a single charitable cause, plainness, for the pressure on Huang. The document’s own §7 and §9.4 locate part of it in features specific to him: asymmetric evidential standards (challenge 2) and categorical reassurance (W3 “fits him better than most”).
• Evidence: 03 §7 table, ranks 2 and 6; §9.4; leaders-comparison.md §6 (“The reassurance trap (W3) fits Huang better than most”).
• Replacement text: “Huang states that model most plainly and defends it most fully, which is one reason the reports press on him hardest; the others, asymmetric evidential standards and categorical reassurance, are his own (section 7, challenges 2 and 6; section 9.4).”
• Confidence: medium.

24. §11.3: “Discounting Huang’s framings because Nvidia has a stake in them”#

• Location: l. 887: “- Discounting Huang’s framings because Nvidia has a stake in them. They should be judged on whether their premises are verified.”
• Rubric: d. Direction: symmetry.
• Problem: This is protection for Huang specifically, with no counterpart for the labs, whose warnings Huang discounts as serving incumbents (“deflection”; “ulterior reasons”). The previous item covers bad faith “on either side”, but this one extends a Huang-specific charity.
• Evidence: 03 §1.3 (“the same rule protects his critics from his imputations of motive”); §6.1 item 15; leaders-comparison.md §5, I9 limits (“The allegations here are inferred, not documented”); LA3 l. 47.
• Replacement text: “- Discounting Huang’s framings because Nvidia has a stake in them, or the labs’ warnings because they could gain from restriction. Both should be judged on whether their premises are verified.”
• Confidence: medium-high. The balance review listed the original line under “should not change”; this keeps it and adds its Mirror.

25. §11.3: the labour-market item, “harm is detected fast”#

• Location: l. 894: “where benefits are large and near and harm is detected fast, so T4’s conditions often fail.”
• Rubric: b. Direction: less sympathetic.
• Problem: The concession is stated more strongly than its source. D06 says labour harm is detected “within years, not decades”, and immediately adds “detection is not reversal”. The same list, in the latency bullet, and §3.2 both say that effects on skills and early careers have latency.
• Evidence: D06 l. 190 (“Labour harm is detected within years, not decades… But detection is not reversal… part of the cost of waiting is persistent for the cohort that bears it”); red team D06-B l. 122; 03 §3.2 and §11.3, latency bullet.
• Replacement text: “where benefits are large and near and harm is detected within years rather than decades, so T4’s conditions often fail. Detection is not reversal, though: losses to the cohort that bears them can persist, which keeps T1’s question of who bears the interim cost open.”
• Confidence: medium.

26. §12.1 question 3: “would count strongly”#

• Location: l. 923: “Applying the condition, or naming who “we” is, would count strongly for the reading of his position as a considered philosophy;”
• Rubric: b. Direction: less sympathetic.
• Problem: Applying a costly shutdown condition would be strong evidence. Naming “we” is cheap and is weaker evidence. The source says only that applying the condition “would support H3”.
• Evidence: hypotheses.md §10, item 1; 03 §9.2 pattern 4 (the condition is “weak evidence of sincerity” in expectation).
• Replacement text: “Applying the condition would count strongly for the reading of his position as a considered philosophy, and naming who “we” is would count for it;”
• Confidence: low.

3. The hypothesis verdict (rubric f)#

“Partly holds” (§8.4; echoed in §2 and §8.5) is the calibrated reading of hypotheses.md, not a midpoint, and no change to it is proposed. - The hypothesis has a disjunction. As posed, it says “largely unaware of, or does not engage with”. The file rates the first disjunct low and the second medium-high (search-limited). §8.4 says exactly this: “‘Bounded’ holds in the sense of non-engagement…, not ignorance of history.” - “Partial” for governance does not mean the hypothesis is half-true. It means the engineering frame is one of several documented sources of his governance conclusions. H1a is high for mechanisms and medium for governance, and role, supplier interest, alliance and archive do much of the work (hypotheses.md §2.1, §8.1). §8.4’s “What engineering does not supply” states those sources plainly. - The file’s own conclusion matches. hypotheses.md concludes “H1 is neither dismissed nor confirmed”. Moving to “holds” would overstate H1a for governance; moving to “does not hold” would ignore the medium-high non-engagement finding. - The one miscalibration is in the verdict’s treatment of “sincere”. The verdict treats sincerity as borne out (“nothing in the record contradicts it”), although under the rules it is assumed, not tested (proposal 15). - Two smaller tilts sit in the paragraphs around the verdict: proposal 14 (“richer archive”) and proposal 13 (missing items limited to the interview). They pull in opposite directions.

4. Mirror depth (rubric g)#

• §5.5 already discloses that the Mirror was applied “with less depth”. Its table has 7 rows against section 7’s 12. That imbalance follows from the brief (Huang is the subject) and is disclosed, so it is acceptable.
• The supporting files hold two further Mirror findings that fit without new analysis: proposal 3 (W4 turned on the warners; the I5 pattern on the critics’ side).
• Two other Mirror points already appear elsewhere in 03 and need no addition to §5.5: the critics’ imputations of motive (§7 rank 7 Mirror) and Klein’s misattribution of Apollo’s doubt (Appendix A, 4.1 row).
• Nothing else in the files would add depth without new research.

5. Passages checked and judged calibrated#

6. Echoes outside scope B (for the scope-A applier)#

These are noted only; no proposal is made for them here. - §4.1, l. 203: “The charge of false precision does not hold” (see proposal 1). - §2, In brief item 4: “an advisory seat and an alignment of interest, from which no motive is inferred” (proposal 10). - §2, In brief item 3: the DuPont qualifiers (proposal 9). - §2, “Why he sees it this way”: consistent with the §8.4 verdict. No change needed.

7. Tally#

By rubric letter (primary):

By direction:

Highest-confidence changes: 5 (Box 20.4 “applies directly”), 19 (“We do that naturally, anyway”), 1 (false precision), 6 (“Caution is not costless”), 7 (§6.3 Analysis), 11 (the weak W2 marker), 18 (“form and emphasis”), 20 (hardened in practice, paired with [47:10]), 24 (symmetric discounting).