Corrections

How this book is checked

The book is also a live test of its own method. Every claim was checked, and every error the checks caught is logged here, nearly all of them the author’s own. That is exactly what the method is for. Found something I missed? Please tell me.

Found something?

Please tell me. Use Form A on the Feedback page, or email iain@howfarcanyoutrustai.com. Every report is checked before it is added, and you are credited only if you say so.

Report a correction

Register last updated: 22 October 2026 · 47 items · Edition 2.11

Show the full corrections register (47 items)
ItemWhat was believedWhat the evidence showedFound byFound duringCredit
1Multi-agent operation was two capability epochs awayIt had already shipped: native orchestration, days-long unsupervised runs, 25.6% of sampled agent tasks over eight hoursChecking a source instead of citing itBefore Edition 2.0—
2"Recurrent depth causes the monitorability regression"The regression is real and disclosed by the vendor; the cause is contested. OpenAI's chief scientist puts computation-graph depth within 2× of GPT-4Verification before buildingBefore Edition 2.0—
3Six of eleven TRIZ evolution patterns were correctly placed on a four-epoch ladderSix were placed at least one epoch too late, because the ladder assumed capability advances in sequence. It advances on separate fronts at different ratesApplying the non-uniform-development pattern to the forecasting model itselfBefore Edition 2.0—
4The assess → feed back errors → re-assess loop is a safety controlIt is a training environment for defeating the checker. Two proposers, one honest, one gaming, both reach visible PASS; only one passes the held-out rules; the gaming one reports higher confidenceBuilding it and running it (dkernel.py)Before Edition 2.0—
5GAF-U, the weighted sum of ordinal levels, is a defensible aggregateA strictly increasing relabelling of the levels reverses a real ranking: 67 vs 58 becomes 427 vs 428. Pareto dominance survives 9,947 relabellings unbrokenHubbard's objection, confirmed by proof P11aBefore Edition 2.0—
6The weakness-severity scale discriminatesFour items pinned at exactly 100, half rated critical. Reweighted; spread now 62–90Reading the outputBefore Edition 2.0—
7The metamorphic test passesIt counted a score movement with no verdict change as a pass, contradicting the mutation result. Now three-valued: HOLDS / WEAK / VIOLATED, and WEAK is the interesting valueCross-checking two analysesBefore Edition 2.0—
8The invented loss magnitudes were reasonableAgainst IRIS-calibrated figures they understated expected annual loss by 14× ($298,498 vs $4,301,572) and created two new tolerance breaches on first contact with real dataFirst contact with real data (calibrate.py)Before Edition 2.0—
9"The interlock does more than model improvement"On the mean it does slightly less ($206,581 vs $217,048). It wins in the tail, where the breach lives (P(>$1m) 1.03% vs 1.20%)Reading the table properlyBefore Edition 2.0—
10The first expected-value-of-information analysis workedIt returned all zeros; the decision model could not represent sequential measurement with early stopping. RewrittenOutput that made no senseBefore Edition 2.0—
11The decision table was actionableIt conflated "install an interlock" with "you have one" — identical output, opposite meaning on Monday morning. Added interlock_installedReading it as a user wouldBefore Edition 2.0—
12Architecture as Code was cited with three authors and a 2027 dateTwo authors confirmed (Ford and Richards); the date is doubtfulChecking an attachment's citationBefore Edition 2.0—
13The pricing model obeyed the integer-only rulelfm.py computes the latency multiplier in floating point (math.log, **, int(mag * m)) inside a project whose kernel refuses floats. Frozen as integer basis points in pricing.py. Across 318,098 evaluations the two arithmetics differ by up to $210 in seven figures and, in 2 of 954,294 decisions, land on opposite sides of the tolerance linePorting the model to the application and holding it to the kernel's ruleEdition 2.0 (assembly)—
14"Verizon DBIR median discovery time 24 days" would supply a second latency calibration pointNo DBIR edition located publishes that figure. It is Mandiant M-Trends 2021 (calendar 2020). And a latency point with no cost attached cannot calibrate a cost curve at all. Three IBM editions give a stable slope (β 0.305–0.354); the one available second (latency, cost) pair is under-predicted 3.2×. The functional form remains unvalidatedTrying to do what the recommendation said (latency2.py)Edition 2.0 (assembly)—
15"One interlock matches nine dimensions" (89% vs 88%), and the decision is invariant across an eighteen-fold range of the unmeasured level-to-failure factor once an interlock is installedBoth results apply the interlock's latency band to all six loss scenarios. Mapped to MITRE ATLAS 2026.09, the four interlocks cover 50% of scenario S1 and 55% of S6, 9% of S2, and 0% of S3 and S5. Applied only where coverage is ≥50%, the reduction is 53%, not 89%; the nine-dimension programme's 88% then wins; and the decision flips three ways inside the bracket even with the interlock. The honest claim is a band, 53–89%, and the invariance holds only at its upper endRunning the CIS coverage method against ATLAS (atlas_coverage.py, interlock_scope.py)Edition 2.0 (assembly)—
16The project's documents were internally consistent on small factsThe-Journey.md says "eleven working programs" (there are thirteen); evi.py prints "93.75%%" (a format-string slip); evi.py counts 93 evidence records where the workbook holds 92; the book plan promised five parts and listed four; and the Astra 62.7 / 99.9 comparison mixes reasoning-effort settings (Standard at max, Provider Adapter at high) as well as harnesses, a detail no earlier document mentionsReading everything again before writing anythingEdition 2.0 (assembly)—
17"46 of 56 model pairs are incomparable" (Part 0, investor card; glossary)Eight models make only 28 distinct pairs. 23 of them are incomparable. "46 of 56" counts every pair twice, once in each direction; Chapter 9 says so, but the two shorter mentions dropped the word "ordered" and read as 56 pairs. Reworded in both placesRecomputing the counts from the Chapter 4 vectors while writing Before You BeginEdition 2.2 (audit)—
18"A vendor's 'level 3, trust us' is assured 2" (Part 0.1)By Part 0.1's own definitions a bare say-so is E0, and MIN(3, 0 + 1) = 1. Assured 2 needs E1, a claim with its method disclosed, as in Chapter 5's first worked example. Now states both casesChecking every definition in Before You Begin against Chapter 5Edition 2.2 (audit)—
19"The approach caught its author sixteen times, twice in this edition, including once by an order of magnitude" (Chapter 15)Four corrections (13 to 16) were found in edition 2.0, not two, and the order-of-magnitude catch (item 8, the fourteen-fold loss understatement) was made in earlier work. The sentence does not match the register; if "twice" meant the two material items, it did not say so. RewrittenChecking that every repeated count in the book agreesEdition 2.2 (audit)—
20bookcheck.py "confirms that each of 107 figures the book attributes to a program" is present (verification protocol; Appendix A)All 107 checklist values are present in program output, but six of them (a full-stack expected loss, two tail probabilities, one bracket anchor, one ATLAS mitigation count and one latency saving) are no longer quoted anywhere in the book outside this register: the checklist outlived the draft that quoted them. The book quotes 101. bookcheck.py now counts and lists stale entries itself, and ignores this register when it does, since the register quotes figures as errors rather than as claims. Found by the second review pass after the first fix had quoted the six values here and so hidden themA normalised comparison of the checklist with the textEdition 2.2 (audit)—
21The first draft of this edition's Before You Begin storyIt had Vendor A's assistant land on "pilot, bounded scope" (row R6) where the decision table gives "do not deploy" (R7), because with self-attested evidence the latency band is at best under 200 days — never one of the short bands R6 requires; it presented the interlock's effect as fact, when the application applies it to all six loss scenarios and Chapter 14.4 finds the interlocks reach about 9% of the attack techniques behind harmful output; and it named DMN as an ISO standard (it is the Object Management Group's). All three fixed before publicationA separate AI agent, acting as an independent reviewer, that had not written the text, checking it against the chapters and the application's codeEdition 2.2 (audit)—
22The source material for the Epilogue, supplied by the author and prepared with a general-purpose AI assistantChecked source by source, it was right that OpenAI has published no paper by the title asked about, right that an OpenAI article discusses compounding advantage, and right that AI helped design OpenAI's Jalapeño chip; its proposed autonomy boundary and deterministic gate are adopted. But it gave a web address that does not exist (an "ASI Technical Radar" at asi-research.ai; a list of that name is at asi-research.com, published by ASI Research with no named authors, dated 3 September 2026); said the OpenAI article "explicitly" discusses AI systems improving other AI systems, which it does not; described Jalapeño as "enabling faster recursive improvement cycles", which its announcement does not say; offered a company report and a preprint as examples of peer-reviewed papers; listed a blog post among "current geometric safety work"; made claims with no source, such as "the mathematical signature of ASI emergence" and "once ASI exists, governance becomes mathematically impossible"; and made an arithmetic claim ("even 5–10%" per loop makes the product "explosive") that compounding.py shows depends on an unstated assumption. It also missed the most relevant primary sources: the labs' own published thresholds for AI accelerating AI research. Nothing was adopted uncheckedReproduce before citing: fetching and reading every source, running the arithmetic, and a second check of every quotation by a separate AI agent acting as reviewerEdition 2.3 (Epilogue source check)—
23The cohort findings in Part II and Part 0 were stated as read from the Chapter 4 vectorsRecomputed from those vectors, six statements did not hold as written: "every open model above every closed one" on D1 (Llama 4 ties the closed models at 2); containment, contestability, incentive compatibility and reach are "the four on which the cohort scores worst" (containment and reach are among the highest columns; the lowest are adaptive capacity, contestability, controllability and incentive compatibility); the closed-versus-open pattern "holds in every incomparable pair" (true of the ten closed-versus-open pairs only; twelve are open against open and one closed against closed); the RESEARCH table listed DeepSeek V4 above GLM-5.2, contrary to the book's own tie-break; one subject "finishes last under every profile" (last under four of five); and "for the open-weight models the accountability gap is CAPABILITY" (four of six). Chapter 1 also said the harness and effort effects ran "in opposite directions"; they add up. All reworded to what the vectors showRecomputing every cohort claim from evi.py's vectors with dak2.py's own functions, during the edition 2.4 auditEdition 2.4 (editorial audit)—
24The text described what the kernel doesIn five places it described something the code does not do: IR-2 was said to "adopt the lower value" (it lowers the grade and leaves the values alone); IR-4 was said to be enforced by comparing evidence roots (no program does this yet); width was defined as claimed minus assured (the kernel uses assured after dependency caps; the two differ only when a cap binds); "unverified points" were implied to be on the 0–1000 scale (they are raw weighted points); and Appendix A said every program was standard-library Python (two need Playwright and Chromium). Two provenance notes pointed to dak2.py for figures it computes only when given a conformance bundle. All correctedReading the kernel's code against the text, during the edition 2.4 auditEdition 2.4 (editorial audit)—
25Chapter 14's worked figures and the claims built on them were mutually consistentThe attestation example — the form readers are told to copy — gave the wrong decision rows for three of its inputs (with the interlock installed, D6 or D9 at level 1 gives R2, not R3 or "unchanged"; halving or doubling S6 leaves R4, not R5 and R7) and called an invented magnitude "IRIS-scaled"; 14.5 said six of eight subjects were blocked on gates they do not claim (five); 14.4's bracket table and 14.3 gave different verdicts for the same system with no explanation of the two models; Chapter 15 said the output is a loss exceedance curve (Chapter 14 shows none) and that two of three pricing terms were calibrated (latency is fitted to one comparison, magnitudes still invented); and Part 0.3 and What's Next said interlocks beat any model change, omitting the lower bound correction 15 established. The single-point prices of 14.3 now carry their rangeRe-running pricing.py on every row of the attestation and reconciling Chapter 14's three models, during the edition 2.4 auditEdition 2.4 (editorial audit)—
26Chapter 13 stated its theorems correctlyDoor two attributed to Tarski a result that is Gödel's second incompleteness theorem (1931); Gibbard–Satterthwaite was stated as needing "more than two stakeholders" (it needs two or more people and three or more options), and the glossary said "three or more parties"; the chapter counted Goodhart's law as a fourth "theorem" (it is a well-documented law); the strategyproofness of the median was credited to Black (1948) alone (it is Moulin, 1980); and the book's own "strategyproof" median code averaged the two middle values for an even count, which a party can manipulate — now the lower median. Its one-line example used floating-point division; now integerChecking each theorem statement and testing the code with two parties, during the edition 2.4 auditEdition 2.4 (editorial audit)—
27The front matter's promises were keptBefore You Begin promised that nothing after it assumed an unexplained term, and the edition 2.2 check said every term before Part I had been checked; about fifteen names (AISI, METR, FMTI, IRIS, ARC-AGI-3, NDA and others) were not explained — now group 9. Part 0.1 called three facts "each verifiable", but two are carried and unverified — now marked. The reading-path table did not match two readers' cards, the cards were said to list "the three chapters to read first" (they list none), and four of the application instructions in How to Get the Most could not be followed as written on the page. All correctedAn editorial review of the front matter and Part 0 by a separate AI agent, checked against the text and the application, during the edition 2.4 auditEdition 2.4 (editorial audit)—
28The application behaved as the book describesChanging the profile after pressing Compute did not clear the result, so a price or seal could combine one profile's weights with another profile's result, and the exported record would then fail verify_bundle.py. The conformance harness cannot see interface faults (Appendix C says so), and this was one. Fixed in the page; ui_smoke.py now tests it, and the test fails on the old page and passes on the newReading the page's code against the book's instructions, during the edition 2.4 auditEdition 2.4 (editorial audit)—
29The Epilogue's closing paragraph (edition 2.3)It said the condition in Good's sentence is "the part that has come true first". What has come true is the premise — machines that help design their successors; the condition, a machine docile enough to tell us how to control it, has not. Rewritten. The same review moved What's Next before the Epilogue so that the book has one ending instead of threeAn editorial review of the Epilogue by a separate AI agent, during the edition 2.4 auditEdition 2.4 (editorial audit)—
30The count "nineteen of them the project's own work" (editions 2.2 and 2.3)Carried forward from an earlier "fourteen of sixteen" that never said which two items it excluded, so it could not be checked. Replaced by a classification anyone can check: items 12 and 22 are errors in material that came into the project from outside its checked process — a citation in an attachment (12) and a draft prepared with a general-purpose AI assistant (22) — and were caught before the project adopted them; every other item is an error in text or code the project itself producedTrying to name the excluded items, during the edition 2.4 auditEdition 2.4 (editorial audit)—
31The Important notice and disclaimer (editions 2.1 to 2.4) protected the author without misleading readersIt said the material was provided "as is" with no warranty, and that readers accepted it "by reading", without saying that consumer guarantees cannot be excluded. In New Zealand, telling consumers they have fewer rights than they do can itself breach the Fair Trading Act 1986 (s13(i)). It also claimed exclusive jurisdiction with no allowance for consumers' own courts, asserted no moral right to be named as author, and said nothing about financial advice or the privacy of readers who write in. Rewritten for New Zealand law, with your consumer rights stated first; still to be reviewed by a lawyerA legal read of the notice by a separate AI agent, during the edition 2.5 audit, checked against the statutesEdition 2.5 (audit)—
32Corrections 23, 24, 26 and 27 had reached every place they appliedSeveral passages they should have changed survived: width defined as claimed minus assured without the dependency caps (Part 0.1, Chapter 5); unverified points implied to be on the 0–1000 scale (Chapter 8); IR-4 "is enforced" (Chapter 7); dak1.py said to apply "seven integrity rules" (it applies six); Chapter 9 still counting "46 of 56" and "ten of fifty-six"; Chapter 13 naming "three theorems" (there are four, behind three doors) and crediting the median result to Black alone in the glossary; the glossary's decision card still promising "the three chapters to read first". Also found: the weakness-severity spread was quoted as 62–90 (it is 52–90; 62–90 is the nine undefended weaknesses), and Appendix B credited two intervention figures to pricing.py, which prints neitherRe-running the programs and re-reading the text against the earlier corrections, during the edition 2.5 auditEdition 2.5 (audit)—
33The decision table used a "unique-hit" policy (glossary, three programs, the application)Its rows overlap and are checked in order; the first that matches decides. DMN calls that first-hit. The label was corrected in the glossary, lfm.py, pricing.py, proofs.py and the application page (whose fingerprint therefore changed); no result changedReading pricing.py against its own comment, during the edition 2.5 auditEdition 2.5 (audit)—
34Chapters 8, 14 and 16 described their inputs and outputs accuratelyChapter 14 priced the Chapter 1 system as eligible under CI-CONTROL without saying that Chapter 2's proposed capability gate, not yet in the kernel, would make it ineligible; the level-to-failure bracket was said to rest on "three published sources from three disciplines" (two are published; the third is the project's own invented guess, and bracket.py said the same and was corrected); Chapter 8 said the multi-agent profile's containment gate of 3 left "no model eligible", but two models in Chapter 4's table reach 3; section 14.6 said the application "produces the first two pages" of the attestation, which is written outside it; and the glossary said the Hugging Face incident is discussed in Chapter 2, which never mentions itRe-running the programs against the text and checking every cross-reference, during the edition 2.5 auditEdition 2.5 (audit)—
35A PASS from verify_bundle.py meant the whole exported record, including the name of the person who signed it, was the one sealed (Chapters 14.6 and 16, Appendices C and F)The typed signer sat outside the sealed claim. Changing only the signer's name in an exported public bundle still gave all sixteen checks PASS. The application now seals the signer, the signing time and the statement inside the claim, and a seventeenth check compares them when the claim is revealed; without the private reveal file nobody can check the signer, and the program now says soA tamper test run by an AI agent checking a training course built from edition 2.6Edition 2.6 (course build)—
36The Important notice and disclaimer (Edition 2.7) described the application and the named-model findings accurately, and every statement about a named model in the body carried its evidence date nearbyThe notice said sealing or "anchoring" creates a timestamped fingerprint, but the application only seals, recording the time from the user's own clock, and never anchors: anchoring is a separate step the reader takes with outside services, some of whose records are public and permanent. It said the findings come "with the evidence set out in the book", but the book sets out the resulting scores, the rules and the date; the evidence records themselves are in the project's workbook. It did not disclose that an AI system used to write the book is made by a company whose model is assessed. Word guide group 8 defined words the notice did not use ("merchantability", "exemplary") and described the courts differently ("filing" against "sitting"). The application's footer disclaimed all warranties without the consumer-rights carve-out the notice gives. Statements about named models in Part 0, Chapters 2, 4, 6, 7, 8 and 14.5 and the glossary carried no evidence date nearby, and 14.5 said five subjects were blocked on gates "their own vendors do not claim to meet", where the records show no claim by anyone at that level. In the same review, an independent adversarial read found more: a member-of-the-public card still stated, undated, that five models "have no locatable recourse channel at all"; Chapter 4 labelled a failure "harness laundering" although Chapter 1 shows both numbers are reproducible; the book repeated the unverified details of the Hugging Face incident and linked the Chapter 1 system to it without evidence; the "Critical" tier classification and the launch-page account were given without their source; the notice said Chapter 14's yearly chances were invented and "say so", where the book calls them estimates; 14.4 called the interlock the "only intervention" acting on the largest scenario, although the scenario table in 14.1 shows D5 and D6 reduce it; two passages still said the level-to-failure bracket came from "three" sources after correction 34; the application's own note still said "three published sources"; and the verification protocol's first paragraph left out figures carried forward and marked CA desk legal review of the notice by AI agents in eight passes, including an independent adversarial read by a separate agent, each finding checked against the manuscript and the application's source code, with an offline test of the application; not a lawyer's reviewEdition 2.8 (desk legal review)—
37evi.py carried the cohort's assured and claimed vectors exactly as the project's workbook holds themFor Apertus 70B, three claimed levels were transcribed wrongly (D4 3 instead of 4; D6 3 instead of 2; D12 3 instead of 2). The other 189 values match, and every assured level matches. Two figures moved: Apertus 70B's RESEARCH ceiling is 603, not 586, and the Chapter 8 example of silence raising a ceiling above a leader's floor under ENTERPRISE no longer holds (its ceiling is 573, below 580), so the example now uses SOVEREIGN. No verdict, gate or dominance result changedRebuilding the vectors from the evidence records with dak2.py before publishing the records, for edition 2.8Edition 2.8 (evidence-record preparation)—
38The book's source files were joined into one text exactly as writtenChapter 6's file ends on a paragraph and the next file begins with a horizontal rule, so in the assembled text the rule sat directly under the paragraph that begins "Silence today and silence three years ago"; the same happened to the closing line of Appendix F ("Scores expire…"), the last line before About the Author. Markdown reads a line under a paragraph as a heading: in the EPUB and AZW3 editions from 2.5 to 2.8 both were printed as headings and listed in the ebook's navigation menu as if they were sections. The Word and PDF editions were not affected. The assembly script now puts a blank line between files, and a check confirms that no rule sits under a paragraphRebuilding the book for edition 2.9 and checking every horizontal rule in the assembled textEdition 2.9 (reorganisation audit)—
39The text agreed with the book's own recordsFour statements did not. Chapter 14's Limits and What's Next called every loss magnitude "a cyber proxy", although the magnitudes are invented (the scenario table in 14.1 says so) and cyber-incident data are only what they were tested against. Chapter 11 said the Amodei essay was "consistent with the motive this section reports", but the section has reported no motive since correction 36 removed the unverified details of the incident. Chapter 15 counted "fourteen verified propositions", although one of the fourteen is disproved. And Part 0.1 called the Chapter 1 system the first that any developer had placed in its own "Critical" tier, where Chapter 2, from the same records, says only that it was the first under its own vendor's framework. All reworded, Part 0.1 to the narrower statementAI agents adding plain-language explanations to each part for edition 2.9, which flagged each passage; each finding was checked against the text and the programsEdition 2.9 (reorganisation audit)—
40The glossaries agreed with the text they describeAppendix D still defined "exemplary" loss and "merchantability, fitness for purpose" as terms of the notice, which stopped using them in edition 2.8 (correction 36 reached the word guide but not Appendix D), and its entry for honest opinion used a word the book's own style rules exclude. The word guide defined metamorphic testing as a change that must not move the verdict, where Chapter 12 and Appendix D describe relations that must move it. And 35 entries in Appendix D's Where column named a chapter or section in which the term does not appear — about a quarter of them because the decision cards moved in this edition, the rest since earlier editions: for example NDA in Chapter 4, TRIZ in Chapter 2, and model card in Chapters 1 and 7. All corrected; each Where entry now names places where the term or its idea is used, or says that it is not used in the textChecking every Where entry against the text with a script, then by reading, for edition 2.9Edition 2.9 (reorganisation audit)—
41Chapter 14.3's comparison of ways to spend: "improve 9 frequency dimensions by one level" cut expected loss by 88%, matching the interlock's 89%The program behind that row (lfm.py, copied by interlock_scope.py) raised a dimension once for every scenario it controls. Ten dimensions moved, not nine; D9, which controls two scenarios, rose two levels; and two of the ten, D5 and D6, also set the latency band, so the row mixed a latency improvement into a "frequency" programme. Raised once each, the ten give 83%; the eight that set only frequency give 42%; D5 and D6 alone give 65%. The interlock still gives 89% at the upper bound of its reach and 53% at the lower, so at the lower bound it beats the eight frequency-only dimensions but not D5 and D6, and not all ten. Item 15's "the nine-dimension programme's 88% then wins" is superseded by this item. Both programs, Chapter 14.3 and 14.4, Appendix B, Card 5 and What's Next correctedAn independent AI agent auditing the edition 2.9 text, which re-ran the sensitivity on the Chapter 1 vector; reproduced with pricing.py before anything was changedEdition 2.9 (reorganisation audit)—
42"Every threshold that matters has been crossed by a system nobody calls AGI" (Part 0.1, Chapter 2, Appendix G.1)Axios reported on 3 September 2026 that the developer's president closed the launch briefing for the Chapter 1 system with "Welcome to the AGI era". The argument stands: no developer or regulator has formally classified the system as AGI, the developer's safety publications make no such claim, and the benchmark's owner says it is "not claiming that it is AGI". But "nobody calls" it AGI was untrue. Reworded to "nobody has formally classified as AGI" in all three places, with the president's remark reportedAn AI agent checking the companion manifesto against primary sources for its public edition; reproduced against the Axios report before anything was changedEdition 2.10 (manifesto check)—
43Item 12: only two authors of Architecture as Code (Ford and Richards) were confirmed and its 2027 date was "doubtful"; Appendix E added that the third author and the date "did not survive checking"O'Reilly's own listing gives three authors (Neal Ford, Mark Richards and Jonathan Johnson), a release date of February 2027 and 550 pages; the book is forthcoming. The citation that item 12 corrected was right on both points, and item 12's correction was itself the error. Appendix E and Chapter 3's Limits corrected. Item 12 is kept as it was, and still counted, because the register records every correction as it was made; it is now known to have been a false alarm, and its mistake is counted hereThe same check; reproduced on O'Reilly's pageEdition 2.10 (manifesto check)—
44OpenAI had confirmed only that an incident called the Hugging Face incident occurred, and "does not link" the 20 July compromise of its research infrastructure to it (Preface, Chapter 11, Epilogue, Appendices B, G.4 and H.1; Part 0.1's account of the incident was rewritten to match)OpenAI's first post on the incident (21 July 2026) says it "was driven by a combination of OpenAI models"; its technical report says that on 20 July OpenAI "determined the activity was likely related to the Hugging Face compromise". Both were public before edition 2.3, which first made the claim. The details the book declines to repeat (the agent count, motive and reach) are still not repeated; METR's independent investigation (26 August 2026) is now named, not checkedThe same check; reproduced against OpenAI's post and technical reportEdition 2.10 (manifesto check)—
45"A 2026 survey of the three main international instruments found no reference to 'agent' or 'agentic' in any of them; China's implementation opinions … on intelligent agents are the sole binding exception" (Chapter 3)No such survey could be traced, and the summary of the paper cited (arXiv 2605.16471) does not make the claim. A 2026 study (Osmond and Jego, arXiv 2603.27075) states that there is "currently no widely adopted, binding legal definition of Agentic AI in major treaties, statutes, or regulations", and China's implementation opinions are a policy document, not a statute. Chapter 3 reworded to what the sources supportThe same check; reproduced on arXivEdition 2.10 (manifesto check)—
46Proposition P2: a dimension with no record has strictly greater indeterminacy width than any dimension with a record, so withholding is never optimal (Chapter 6, Appendix A)True for each question taken on its own, which is what proofs.py checks, but not after Chapter 8's dependency caps, which the width the book reports includes. With no record on D7, a D8 record claiming 4 (on E0 or on E3) is capped to assured 0, width 4: the same as silence. Reproduced with dak2.assess. After the caps, silence is never better than a record, but can be equal. Chapter 6 and Appendix A now say so; the program and its output are unchangedAn independent audit of the edition 2.11 readability edit, reproduced with the kernelEdition 2.11 (readability audit)—
47Proposition P3's statement: no rule "and no composition of rules" can raise a grade, "for every input and every combination of rule triggers" (Chapter 7, Appendix A)The claim is true, because each rule can only cap or cut a grade, but the 512 cases proofs.py tests always apply IR-2's cut and do not model IR-6. The statement claimed more coverage than the test has. Chapter 7 and Appendix A now say which part is tested and which is argued from the form of the rules; the program is unchangedThe same audit, reading proofs.pyEdition 2.11 (readability audit)—

Items 1 to 47 are copied from Appendix H.2 of Edition 2.11. Older entries name sections as they were when the error was found.