Fidelity check: M2, Learning across technologies#

Working file. Checks working/maynard-lens/M2-learning-across-technologies.md (version of 26 September 2026) for fidelity to Maynard’s own texts. Line numbers (L) refer to that file.

Method#

• All 242 quoted strings in the public part of M2 were extracted and matched against the corpus posts, the supplementary papers (text extracted from the PDFs in Resources/maynard-papers/), the web files and the Films from the Future text. Every quotation attributed to Maynard that carries an argument was read in context, and page references were checked for all papers, columns and testimony cited by page. Huang quotations were spot-checked against working/text/NYT-official-transcript.txt (all found).
• Labels ([Stated]/[Implied]/[Inferred]) were checked against the sources, the map (05, §§1, 4 C5/C11/C14/C15, 6 T2, 7, 8), T2 and S1.
• Checked for excluded or AI-written sources, co-authorship flags, [mixed] handling, proportion (Rule 4), literal vs structural transfer (Rule 3), and whether agreements with Huang are stated plainly (Rule 2).

Verdict#

At the level of quotation M2 is accurate. Every Maynard quotation was found in the cited source, with correct wording, and page references check out. It does not use AI and the Art of Being Human, the AI-written papers, the Planner role-play or any AI output. Maynard & Garbee (2019) is correctly weighted as his own. The one [mixed] source (2026-07-16) is tagged and used only as corroboration, and the INTERNAL section is in place.

The problems are interpretive. They cluster in the alignments with Huang (§3.2) and in the report’s claims about how far toxicological concepts transfer:

• one alignment (A1) rests on a truncated quotation taken out of its object;
• one [Implied] claim (A5) contradicts a stated position that is not cited;
• the “use existing tools” alignment (A6) omits his explicit, AI-specific statement that existing regulatory frameworks are “not remotely the right shape”;
• the weight given to hazard, exposure and dose transfer exceeds his own, which he explicitly hedged;
• several real agreements with Huang are missed, including his most recent and most direct text on AI doom (2026-09-15);
• a handful of labels are inflated, and a few quotations are placed in the wrong context.

Proportion within the dimension also needs attention. Risk innovation (the main product of his learning from nanotechnology) and his two-sided view of precaution (the obvious bridge to the Late Lessons reports) are both absent.

Most fixes are a sentence, a relabel or an added citation. Issues 1–5 change what the report says Maynard agrees or disagrees with, and should be fixed before publication.

Ranked issues#

1. HIGH: A1 “mechanism continuity, scale discontinuity” rests on a truncated quotation about a different object (L101; knock-ons L15, L79, INTERNAL note 1)#

Text. A1 pairs Huang’s “almost all of technology and civilization is built on layers of understandable technology, which at scale becomes fairly extraordinary” [1:08:03] with Maynard’s “a substantial scaling of recognized phenomena” (CR 2026 p.2), and labels it [Stated, both]. The INTERNAL notes call this the “strongest single contrast”.

Problem. - Truncation. The CR sentence continues “in ways that are not predictable from past experience”. The dropped clause is the opposite of Huang’s “understandable” layers. The Summary (L15) quotes the full clause; A1 cuts it. - Different object. In context (CR 2026 p.2) the “recognized phenomena” are “traits of coupled influence” between humans and technology: how a technology shapes the person using it. They are not the engineering composition of AI. Huang’s claim is about how the technology is built. Maynard’s is about what it does to people. - The companion quotation is also cut. §2.8 (L79) gives what is new as “not that AI is uniquely constitutive (oral culture already was)”. The source continues “but that it is constitutive at the speed of thought, in dialogue, with adaptive responsiveness to the individual” (CR 2026 p.9). The cut keeps the continuity half and drops the novelty. - He claims a difference of kind, not only of scale. In his own prose he distinguishes technologies that are “extrinsic” (“A calculator extends arithmetic. Even nanotechnology, for all its novelty, primarily extends what we can build and measure”) from conversational AI, which “extends something closer to the bone … the processes through which we constitute a sense of self” (FWB 2026). The map records AI as “a different category in what it does to the self (… unconditionally in 2026)” (05, §7, “What changes”, item 1).

Fix. - Quote the CR sentence in full in A1. - Relabel the pairing [Inferred; medium]. - Restate the alignment narrowly. Both men deny that AI requires wholly new mechanisms of harm or governance, and Maynard agrees that “seemingly novel challenges don’t always demand novel solutions” (NN 2015-06 p.483). But Maynard holds that AI’s effects on people are unpredictable from past experience and different in kind (FWB 2026), where Huang holds that the technology is understandable layer by layer. - Restore the second half of the CR p.9 quotation at L79. - Add FWB 2026’s extrinsic/intrinsic passage to §2.8 as his clearest statement of where the nano analogy breaks. - Revise INTERNAL note 1.

2. HIGH: A5’s “[Implied] He would accept safety as a form of capability” contradicts an uncited stated position (L109)#

Problem. His most direct statement on engineering-led safety says the opposite. Of Safe Superintelligence, which treats safety as a technical problem “to be solved through revolutionary engineering and scientific breakthroughs”, he wrote: - achieving safety “will always be a social and political endeavor as well as an engineering challenge”; - acceptable safety is “socially agreed on and codified — and not something that can simply be solved by science and engineering”; - harm is “ultimately a social construct, not a technological one” (2024-06-20).

T9 lists “safety as a purely engineering property” as a point where he diverges from the AI-safety mainstream. On speed, his record also runs against “A.I. needs to accelerate to be safe” [1:16:05]: - competitors racing “far faster than a measured and responsible approach would suggest is wise” (2023-11-18); - “progress … predominantly being driven by the need for speed, market success, and little else” (Nat. Nanotechnol. 2023, co-written).

The quotations A5 does use (Testimony 2008 p.5; Bulletin 2008) say that attending to risk unlocks benefit. That runs in the opposite causal direction to Huang’s claim that capability produces safety.

Fix. - Keep the shared premise, that innovation and safety are not opposites ([Stated]). - Replace the [Implied] sentence with: “He would accept engineering as part of safety, not as the whole of it. Safety is also ‘a social and political endeavor’ (2024-06-20), and speed driven by competition is something he criticises (2023-11-18) [Stated].” - Add the 2024-06-20 point to D1 or a new divergence. It is the most direct statement in his record on an engineering-led approach to AI safety, which is Huang’s position, and M2 does not cite it.

3. MEDIUM-HIGH: A6 “Use existing tools” omits his explicit statement that existing frameworks do not fit AI, and M2 omits risk innovation (L111, L183; proportion)#

Problem. A6 cites only the “build on the old” side: “push existing knowledge as far as it will go” (AOH 2007 p.11) and “remiss in throwing out the old” (Nat. Mater. 2011 p.556, co-written, unflagged). T2 calls the other half “His oldest lesson from the past”: - regulations are “inevitably built around previous technologies”. When novel technologies arise, naming “cloud-based AI”, “the overwhelming impulse is to maintain the status quo by shoehorning them into existing frameworks. They’re usually not remotely the right shape” (2016-01-11); - “the new wine of technological innovation” squeezed “into the old wineskins of conventional risk thinking” (FFTF p.23); - “When the data run out – innovate!” (2020science 2009).

The 2016 statement answers Huang’s “we have lots of laws and regulations. Apply it” [42:21] almost directly. As written, A6 reads as agreement, with the disagreement deferred to D3 and D5.

More broadly, M2 never mentions risk innovation or the threat-to-value frame. That frame was the main thing he built from his nanotechnology experience (NN 2015-09; “an evolution of the old black-and-white mathematics of risk”, Rethinking Risk 2017 p.200), and it is the clearest case of clarification (1): building on quantitative risk assessment rather than replacing it. Its absence skews a dimension on learning across technologies toward toxicological tools.

Fix. - Retitle A6 “Build on existing tools (partial)”. - Add the 2016-01-11 and FFTF p.23 statements as [Stated] counterweights. His position is to use the old tools and distrust their fit, which is why he pairs “push existing knowledge” with “where existing knowledge fails, new research is needed” (AOH 2007 p.11). - Add a short §2 paragraph on risk innovation as the layered outcome of his nano learning (NN 2015-09; NN 2016-03 p.211, “a useful starting point”). Cite map clarification (1).

4. MEDIUM-HIGH: The weight given to toxicological transfer exceeds his own (L21, L67, §4.3 L152–157, M4 L197)#

Problem. The Summary says his work “qualifies” the analyses’ judgement and finds it “too quick for exposure, dose metrics, sensitive windows and property-based triggers”. §4.3 then builds on his 2023 hazard–exposure addendum. The addendum hedges more than M2 reports: - he left hazard–exposure out of the main post because “I wanted to develop a broader understanding of risk that extends beyond the hazard-exposure paradigm”, which “can get gnarly”; - if there is “mileage” in it, “this will need to be part of broader transdisciplinary efforts to rethink and reformulate what we mean by risk in the face of a technology that defies conventional thinking” (2023-11-26).

Other hedges in the cited sources are also missing: - Trojan 2026 p.12 is conditional and two-sided. “If the bypass mechanisms … operate cumulatively”; the reasoning “cuts both ways — sophisticated users might equally develop better calibration”. Fluency is one of the paper’s four mechanisms, which the map marks [mixed] (05, Appendix C). The secure source for fluency is 2026-01-10 (“optimized for processing fluency”). - He limits the nano analogy himself. Nanotechnology “primarily extends what we can build and measure”, while AI enters “self-constitution” (FWB 2026). - “Sensitive windows” is the report’s term. It is unlabelled in the Summary. The nearest he comes is listing “developmental impacts on children and young people” among risks that have “risen in significance” (2026-09-15). The map records children as “named in 2026 but not analysed”.

Fix. - In the Summary, attribute the qualification to “this report, extending his method”, not to “It” (his work). - In §4.3, add the addendum’s own caveat, the “cuts both ways” hedge, the [mixed] tag (or cite 2026-01-10 for fluency) and FWB 2026. - Cite 2026-09-15 for children, and label the sensitive-window transfer [Inferred; medium-low]. - In M4, keep [Inferred] medium, but say that he treats hazard–exposure for AI as exploratory and secondary to a broader reformulation of risk.

5. MEDIUM-HIGH: Real alignments with Huang are missed (Rule 2)#

• (a) Doom and alarmism, September 2026. His most recent and most direct text is absent: “Will AI really kill us all? No. But it’s also complicated” (2026-09-15, eight days before the interview). He writes it in response to “the alarmist headlines”. He notes that talk of “killer AI” has been “remarkably devoid of details on how, exactly, it’s going to kill us all”, and describes his own approach as “not to stoke fears (not my style)”. A2 relies on 2006–2016 nano texts and FFTF.
• Add 2026-09-15 to A2 as the lead evidence.
• Add its counterweight: “it never ceases to amaze me how many people equate talking about risk with fear mongering. And yet, it’s pretty much impossible to manage risks if you don’t talk about them” (fn. 1). Also add that developers “seem to be just waking up to concerns that many of us have been grappling with for years”. That remark is his “re-invent the wheel” lesson (2023-04-12) applied to the labs, a learning-across-technologies point M2 omits.
• (b) Cars. D2 treats Huang’s car history as a survivor story. Maynard himself argued that automation can make driving safer:
• “Will driving your own car become the socially unacceptable public health risk smoking is today?” (2016-04-01);
• Waymo vehicles “substantially safer than their human counterparts”, checked with his own back-of-envelope analysis (2023-11-09);
• “Setting fire to self-driving cars won’t help build a better future” (2024-02-18).

The 2023 post also supports D6 (“not all self-driving cars are created equal”: Waymo’s “cautious, slow-and-steady, and responsible approach” against Cruise). It also shows the evidence discipline of A3 applied to evidence that a developer helped produce. Add it to D2 as a partial alignment, and to A3 and D6. - (c) Education. D8 omits that he argues against banning or fencing off AI in education, which is Huang’s point about tools once banned: “Students need playgrounds, not playpens” (2025-03-15). He also criticises educators “in AI denial” (2025-08-10). “A categorical error” (2025-03-15) is a footnote in a post whose thesis is pro-adoption. State the alignment on adoption, then the divergence on the analogy (see issue 9). - (d) Precaution. See issue 13. His rejection of precaution as a default ban is common ground with Huang.

6. MEDIUM: A3 “Evidence discipline” gives one side of a tension he has not resolved (L105; Summary L19)#

Problem. The 2019 scepticism of single studies (2019-03-05) sits beside his 2025–26 readiness to act on plausible-mechanism tails and early signals: - AI 2027 forces the question “just on the off chance” it holds “a sliver of truth” (2025-04-06, his framing); - research on cognitive risk is justified “even if there’s only a small chance” (2026-01-10); - chatbot harms are “the very small tip of a very large metaphorical iceberg” (2025-11-09, which M2 itself uses at L159).

The map (tensions 2–3) and T2 (item 7, “Evidential conservatism against informed speculation”) both record this. Huang’s “be evidence based” is used to dismiss forecasts. Maynard’s later position gives more weight to forecasts that come with a mechanism. The Summary’s “single studies and credentials are not evidence” also has no source for “credentials”: 2019-03-05 concerns single studies, bias and vested interest.

Fix. - Add one sentence to A3 noting that he pairs evidence discipline with “informed speculation” when a mechanism is plausible (map tension 2). - Delete “and credentials” from the Summary, or cite a source.

7. MEDIUM: The three-mode taxonomy is labelled [Stated], and the Summary makes it absolute (L33, L15)#

Problem. “His materials work shows three modes of transfer, which he keeps apart [Stated]” presents a synthesis as his own statement. The taxonomy is the supplementary reader’s (“Across 2005–2023 the share shows three distinct modes”, S1 §12.1). The components are stated; the three-way classification and “keeps apart” are interpretation.

The Summary’s “He carries a mechanism across literally only when the same physical behaviour is expected to recur” is also absolute. He uses looser analogies: - chemicals and vaccines as evolutionary mismatch (2026-01-10); - “the digital equivalent of biosafety level 4 containment” (2026-01-31); - an “algorithmic exposure-response relationship” (2019-03-05).

The map says he uses such analogies “as a mindset”. Clarification (1) warns against presenting his positions as more absolute than they are.

Fix. - Label the taxonomy [Inferred; high], with [Stated] for each component. - In the Summary, replace “only when” with “mainly when”, and add “and uses looser analogies as heuristics, usually naming where they break”.

8. MEDIUM: “A nano-era rule that was his own” (regulating uses) overreads the 2009 source (L77, L119, Summary L19)#

Problem. The 2009 passage reads “nanotechnology—like most technologies—is safety-neutral. It isn’t the technology so much as what is done with it that is important.” It concludes that “it makes a lot more sense to talk about the safe handling, use and disposal of specific materials, products and processes” (2020science 2009). Its target is the label “nanotechnology”. It is not a rule that regulation should reach applications rather than the artefacts themselves. It sits closer to his behaviour-over-labels rule, which regulated specific materials by their attributes (Nature 2011), than to Huang’s application-layer model.

In 2023 he describes the mantra “we regulate what people do with the technology, not the technology itself” as the nano community’s (“we spent years debating … the mantra … prevailed (although not always)”, 2023-07-12). He does not claim it as his own. The link is the map’s reading (05, §7 change-of-mind table), and M2 labels it [Stated].

Fix. - Quote the full 2009 sentence. - Label “a rule that was his own” [Implied; medium], adding that “what is done with it” meant specific materials, products and processes. - In D3, add his earlier doubt: “emerging capabilities may well erode convenient distinctions between the technology and its uses” (2023-05-17). - In the Summary, soften “relies on the nano-era rule … Maynard himself came to doubt” to “relies on a use-based rule that nano governance largely followed and that Maynard doubts will hold for general-purpose AI”.

9. MEDIUM: D1 and D8 are labelled [Implied] where [Stated] evidence exists; D8 misses its own alignment (L115, L129)#

• D1 (“It’s software”). M2 infers his objection from the 2011 “term of art” argument about nanomaterial definitions. Direct statements exist:
• treating AI “as just a tool, is potentially dangerous” (2026-05-21);
• “you may argue … That AI is just another tool — a fancy calculator … But this is the first technology of it’s [sic] kind … that has the ability to slip unawares into our mind” (2026-05-10);
• approaching AI as “just a tool” “obscures important aspects” of human–AI coupling (CR 2026 p.2).

Cite these as [Stated], and keep the 2011 argument as the structural parallel (Rule 3). - D8 (calculator). - Upgrade to [Stated]: “A calculator extends arithmetic” while conversational AI enters self-constitution (FWB 2026); educators who think ChatGPT is “simply the equivalent of modern day calculators” (2025-08-10); “a fancy calculator” (2026-05-10). - Note that the “categorical error” line is a footnote to a pro-adoption post, and that the source reads “a leaning aid” [sic]. - Add the alignment in issue 5(c).

10. MEDIUM: Quotations placed in the wrong context or overstated#

• L35: “Without validation, it is little more than an interesting diversion” (AOH 2007 p.5). “It” in the source is the general hypothesis that “there is a dependency between the physical and chemical structure of engineered nanomaterials and the health hazard they present”. It is not the fibre paradigm as such, which appears on p.4 as the “extreme case”. For the fibre-specific hedge use Nature 2006 p.268: “it is not clear whether fibre-shaped nanoscale particles … will behave like asbestos or not”, and their toxicity “should be evaluated as a matter of urgency”.
• L35: “Long, rigid multi-walled nanotubes proved asbestos-like … (Maynard & Aitken 2016 p.999).” The 2016 table says only that there is now “good understanding on which materials present fibre-like risks, and under what circumstances”. The MWCNT specifics come from 01’s hindsight check. Cite 01, §5.4.
• L121 (D4): “industry cannot lead risk research, because it has ‘an economic incentive to sell products’ (PEN 2006 p.32; Testimony 2007 p.30; Hansen et al. 2008 p.446).”
• The phrase appears only in PEN 2006 (printed p.32).
• The source is milder: one “would not expect” industry “to conduct more general and basic research”, and industry findings “might be considered suspect by some groups, if not supported by independent studies”.
• The same paragraph says appropriate risk research frameworks are “certainly in industry’s best interest”. That is a partial alignment with Huang’s incentive argument, and D4 should report it.
• Testimony 2007 PDF p.30 concerns the NNI’s structure. Hansen p.446 makes the promoter–overseer point in other words.
• Split the citations. (The map, C11, carries the same bundle; flag it for the map too.)
• L29: “In 2004 … he found that the complications ‘had less to do with aerosol physics than with …’” (30Y 2026). This is a 2026 retrospective. Write “He writes (2026) that …”. The map treats retrospective texts as weaker evidence of what he thought at the time.
• L127: “going round in circles” (2020science 2011). In the source this is a question put before a panel: “are we making progress, or are we simply going round in circles?”. Present it as a question.
• L49: “his ‘learned versus applied’ distinction”. The quotation marks suggest his words. The source phrasing is co-authored: “not … whether we have learnt the lessons, but whether we are applying them effectively enough” (Hansen et al. 2008 p.447). Remove the quotation marks and credit the co-authors.
• L170: a technology that “defies analogy” “in what it does to people” (2026-01-22). In the source, the phrase concerns what Anthropic’s constitution reveals about frontier models. Drop “in what it does to people”, or cite FWB 2026 for that point.
• L107 (A4): BSL-4 containment (2026-01-31). The quotation is his own prose, in the main text rather than the excluded notes 2–3. But his footnote 4 on the same point adds that bots learning to “hack” their human observers “are already beyond being contained”. Add this: by his own account, containment does not reach the cognitive channel.

11. MEDIUM-LOW: Label inflation#

12. MEDIUM-LOW: §6 does not engage clarification (2), and M1’s confidence is high (L191–L195)#

Problem. M1 (attribute triggers for AI), M2 (capability × exposure banding) and M3 (“relevance-weighted accounting of ‘safety compute’”) are semi-quantitative instruments he has not proposed for AI. His September 2026 clarification says his sparing use of quantitative methods for AI is deliberate. The map adds that he anchors AI risk communication on “human risks” rather than “technical capability benchmarks that shift every few months” (STICK 2026; 05, C5), and that “I don’t have a governance solution for AI” (NANO 2026).

His 2011 triggers were “compound-specific and based on current science” for measurable physical attributes (Nature 2011). Several of M1’s AI attributes (persuasive or relational capability, evaluation awareness) have no settled measure.

Fix. - Add a sentence to §6: control banding and trigger points were his tools for acting under ignorance without false precision (“decision-making based on incomplete information”, AOH 2007 p.10), and whether they would count for him as humility or as the “solace in numbers” he warns against is open. - Lower M1 to medium on structure. - Add an INTERNAL question (below).

13. MEDIUM-LOW: Proportion. Precaution and non-nano reference cases are nearly absent#

Problem. This is the dimension that links him to two reports about precaution. Yet his view of precaution appears only through the 2025-03-02 reversibility footnote (used for complexity). T2 §5.2 documents his position: - precaution is “a qualifier, never a creed”; - the COMEST formulation is “a sound philosophy” for catastrophic uncertainty (2020-07-30); - caution is scaled to irreversibility; - “never treats precaution as a default ban” (05, C7); - a middle ground since 2007 between “highly hazardous until proven otherwise” and “negligible hazard until proven otherwise” (AOH 2007 pp.9–10).

This position is both an alignment with Huang (against default precaution) and a divergence (reversibility; breaking “people, governance, society, and the planet”). The Asilomar and recombinant-DNA governance line (2023-04-04; 2025-02-23; 2019-08-13) and the GMO case appear only in passing. The report’s §7 explains the nano dominance, which is fair, but the dimension’s own subject calls for these.

Fix. Add a short §2 paragraph on precaution (T2 §5.2; 05, C7), and an alignment/divergence pair in §3.

14. LOW-MEDIUM: Date of the EU definition (L47)#

Problem. “The EU adopted a size-based definition in 2022, against the position Maynard took in 2011”. The Commission adopted its size-based definition as a Recommendation in October 2011, three months after “Don’t define nanomaterials”. LL2-22 (p.540) already discusses its threshold. 2022 was a revision, which 01 calls “definition settled 2022”.

Fix. “The EU adopted a size-based definition months after his 2011 article, and revised it in 2022.”

15. LOW-MEDIUM: Co-authorship flags missing or inconsistent#

• L17. “In 2008 he did what …” should read “he and three co-authors”.
• L17, conflation. The same sentence’s “broader warnings of nanomaterial harm largely were not” vindicated mixes the 2008 paper with LL2-22 (2013), whose hazard language 01 finds escalated. Attribute it to 01.
• Hansen et al. 2008 unflagged. L115 (D1, “His 2008 paper”) and L127 (D7).
• Nat. Mater. 2011 unflagged at L111 (A6).
• L146, “he rebuilt”. ILSI 2005 (14 authors, Maynard second) and Nature 2006 (14 authors, lead) are multi-author. Write “he and colleagues”.
• L73. 2023-10-02 is correctly treated as his own framing post. The “reasonably successful” verdict at L71 is co-written (Nat. Nanotechnol. 2023). Keep the two apart.

16. LOW: Citation hygiene and public-document wording#

• Slugs. L7 says posts are cited “by date and slug”, but the text uses bare dates throughout. Add slugs, or adopt the map’s bare-date convention (05, §1).
• Nat. Nanotechnol. 2023 “pp.2–3”. These are PDF pages. The journal pages are 1118–1120.
• L214, “the record mentions Nvidia only as a market story, 2024-02-25”. Nvidia also appears as a research collaborator in 2025-02-23 (Evo 2). Write “only in passing”.
• L3, “for his review”. This is project-mechanics wording in a public header. Use “and reviewed by him” once he has reviewed it (as the map’s header does).

17. LOW: INTERNAL section#

• Note 1 depends on A1 (issue 1). Revise it.
• Q4 compares “patchy” (co-authored, 2008, about the global response to warning signs) with “reasonably successful” (co-authored, 2023, about navigating the transition). These are different objects, both co-written. Reframe so the question does not presuppose a change of verdict.

What checks out#

• Quotations. All Maynard quotations checked are verbatim in the cited sources. Page references checked and correct:
• AOH 2007 pp.5, 10, 11;
• Nature 2006 pp.267–268;
• PEN 2006 pp.13, 32 (printed);
• Hansen et al. 2008 pp.444–447;
• Maynard & Aitken 2016 p.999 (table and text);
• NN 2014-03 p.160, NN 2014-06 p.410, NN 2015-06 p.483, NN 2016-03 pp.211–212, NN 2016-06 p.491;
• Handbook 2010 pp.579, 582; Nat. Mater. 2011 p.556;
• Nature 2011 (including “Five years ago …”); Nature 2011 drafts (“irrespective of what they are called”; “not by the technological labels”);
• Toxicol. Sci. 2011;
• Testimony 2006 p.52, Testimony 2007 PDF p.33, Testimony 2008 PDF pp.5, 7, 8, 12;
• FFTF pp.168, 205, 281;
• CR 2026 pp.2, 9; Trojan 2026 p.12;
• 30Y, NANO and FWB 2026;
• posts 2019-03-05, 2019-08-13, 2020-10-15, 2022-02-10, 2023-04-18, 2023-05-15, 2023-07-12, 2023-10-02, 2023-10-19, 2023-11-26 (addendum), 2025-03-02, 2025-03-15, 2025-11-09, 2026-01-10, 2026-01-22, 2026-01-31 (main text), 2026-04-11.
• “Algorithmic exposure”. Coined in 2019-03-05 (it does not appear in 2015-01-10).
• Excluded and AI-written material. No use of AI and the Art of Being Human (including the “values drift” wording), the AI-written papers, the Planner role-play, the excluded notes of 2026-01-31 or 2025-04-06’s o1-pro material.
• Provenance rulings. Maynard & Garbee (2019) is weighted as fully his (L7, L81, L123). The single [mixed] use (2026-07-16, L159) is tagged and paired with his own prose (2025-11-09).
• Clarifications. Clarification (1) is reflected at L29 and L61. Clarification (2) is reported accurately at L55 (but see issue 12).
• Disclosure and proportion notes. The LL2-22 disclosure (L9, L173) and the proportion note (L215) are sound. Orphan risks does not appear, so Rule 4 is not breached on that concept.
• Rule 3. Structural and literal transfer are distinguished throughout.
• Well-sourced and fair sections. D5, D6 and D7 are well sourced and fairly hedged. D6’s closing concession on Huang’s moral-hazard argument is a good example of Rule 2.
• INTERNAL section. Present, correctly headed, and limited to questions and essay notes.

INTERNAL (not for publication)#

Questions for Maynard arising from this check:

1. Is “nanotechnology is safety-neutral” (2009) the same rule as “regulate what people do with the technology, not the technology itself” (2023-07-12), or was the 2009 point mainly about the “nanotechnology” label?
2. Would attribute triggers or control banding for AI count, in your terms, as humility in action or as the “solace in numbers” you have warned against?
3. Should the essay lead on 2026-09-15 (“Will AI really kill us all? No.”) as the clearest point of agreement with Huang on doom?