Revision log: 02-huang-analysis.md#
Strand B. 25 September 2026. Records each issue raised in fairness.md (FA), fidelity.md (FI) and completeness.md (CO), how it was checked, and the outcome. Checks were made against the transcript and the working files in working/huang/. Four points not in the working files were checked live: the CBS article of 20 September (“0% chance”), Fortune’s report of 21 September (full “ulterior reasons” sentence), Zvi Mowshowitz’s post of 25 September (“security mindset”, “outright lie”, “get us all killed”, the company list), and METR’s report (the “peers” sentence, spoofing and Ed25519).
Outcome key: Fixed; Fixed in part (with what was not done and why); Rejected (with reason). Where two reviews raised the same point, the second entry cross-refers to the first.
The document grew from about 40,500 to about 56,300 words. Nothing substantive was removed; passages were corrected or qualified in place.
Fairness review#
| ID | Issue | Verification | Outcome |
|---|---|---|---|
| FA-1 | Straw man: “no regulation, no worries” attributed to Huang by nobody; packaging’s “no new regulation” is accurate | Transcript [01:14], [42:07], [42:21], [47:10]; FC C005 (“accurate”); L1 (Dreamforce via TechCrunch); E3/E1 (Mad Money: “completely unnecessary”); L6 §3 verdict. The CBS phrase “instead of new rules” was not found as a Huang quotation on a live check (it is the reporter’s paraphrase), so it was not used. | Fixed. In brief bullet 1 rewritten; §2.4 rewritten (42:21 objection answers Klein’s “nearly any venue” paraphrase; 47:10 read as opposing new rules now); §7.1 and §7.3(d) rewritten. |
| FA-2 | Klein and Huang graded on different scales; accuracy gap headlined | FC rows C076, C080, C083, C094, C096, C108, C121 checked. C076 and C080 are reports of what the labs say and are correctly graded; C083’s “misleading” treats the disputed question (whether competition compels) as settled; C096 and C121 are graded leniently; C094 is a rhetorical question. | Fixed in part. Headline comparison removed from In brief and replaced with a caveat; new “Why the two records are not directly comparable” and adjusted figures in §6.1 (C094 excluded; C083, and optionally C108, graded contested → 56% / 27–28% / 13–15%; Klein 38/40 if C096 and C121 regraded); notes on C083 and C094 in §6.2 and Appendix A1; §2.4, §6.3 points 4 and 7 updated. Not done: C076 and C080 not regraded (they accurately report what labs say); the fact-check file and Appendix A1 verdicts left as the fact-check’s originals, with notes. |
| FA-3 | Unlabelled evaluative lens; crux reduced to procedure; Klein’s and labs’ gates merged | Transcript; L6 §5 (tempo); E3 §9.1, E4 §§1.2–1.3 (lab positions). | Fixed. Criteria stated in §1.3 and §10.2; new “Where the alternative gates are exposed” in §10.2 (entrenchment, non-signatories, builders’ alarm as interested evidence, false positives, evaluation awareness cuts both ways, speed of public gates); §10.3 rewritten on two levels (substantive, then institutional) with a table separating Huang, Klein, Anthropic, OpenAI and Meta; In brief bullet 6 rewritten. |
| FA-4 | Conditional remarks inflated to “hard commitments”; “ten times” is a prediction; “stricter than they expected” unsupported | Transcript [36:44], [48:58], [1:16:05]; E4 line 264 (“would welcome”) and line 105 (the “stricter” phrase is E4’s own interpretation, not a critic’s). | Fixed. In brief, §7.1 (“conditional statements”; tenfold as prediction; prescription quoted), §4.2 Safety, “welcomed” in §7.1 and §9.3; low expected cost noted in §8.4. |
| FA-5 | Liability-relief charge: evidence missing on both sides (OpenAI blueprint; SB 3444; Bessent insinuation) | E4 line 55 (blueprint quote); S3 line 55 (SB 3444 and retraction); E3 lines 223, 237 (Bessent). | Fixed. One formulation used in In brief, §3.6, §5.3 #4, §6.2 C108, §6.3 #4, §9.3 #3; April–May and 2–3 June rows added to §2.3; Bessent now given as one of two possible bases, beside SB 3444; pre-emption noted as a third, distinct form of relief. |
| FA-6 | Only Huang’s non-answers listed; competition point wrongly marked “not engaged”; “which laws” contradicted §8.3 | Transcript [40:21], [53:36], [38:37]; L5 §4.1 “Limit”. | Fixed. §8.3 row rewritten (“Engaged… less-careful-rival case not addressed”); §5.4 and T6 corrected; “which existing laws” moved from unanswered to partly answered in §3.6; new “Left unanswered by Klein” table in §3.14 (seven items, each checked against the next Klein turn). |
| FA-7 | “None of them was disclosed on air” overstated; interest-alignment table cannot discriminate | Transcript [30:29], [52:16], [1:24:38], [1:25:12], [1:27:32], [1:27:47], [1:34:16]. | Fixed. In brief and §2.2 closing paragraph rewritten; §8.4 table given a “Shared by disinterested experts?” column and a row on liability; §10.4 #8 conclusion moved into the §8.4 Reading. |
| FA-8 | §5.3 #2 misstates labs’ position and Huang’s response; §5.1 row generalises “deflection” to all alarm | Transcript [44:17], [53:36], [55:46]. | Fixed. §5.3 #2 rewritten (Huang accepts two of three elements; “deflection” aimed at a helplessness narrative); §5.1 row now “Claims of helplessness”. |
| FA-9 | Marcus’s “doesn’t realize” quoted unrebutted | Transcript [32:09], [35:36], [36:44]. | Fixed. Rebuttal added in §9.2; §3.5 “Left unanswered” and T2 revised (containment was his first diagnosis; release rule answered the general alignment point). |
| FA-10 | T8 caricatures his handling of the study; CBS “0%” misquoted and mis-compared | Transcript [22:26]; CBS article (live check: “2030 is not going to be the end of the world. There is 0% chance that’s going to be the end of the world”); FC C124. | Fixed. T8 study sentence rewritten; CBS quoted verbatim with date; the Hinton comparison replaced by the narrower point (unsupported point estimate). |
| FA-11 | Interested parties cited as independent support (Delangue; Nvidia blog for GLM forensics; OpenAI’s own figures) | §2.2; L6 line 40 (HF disclosure 16 July); S2 (HF timeline); E4 §1.4. | Fixed. Delangue’s relationship disclosed in §7.3(h) and §9.2; GLM forensics cited to Hugging Face’s disclosure and timeline; OpenAI’s figures flagged as self-reported and paired with METR in §7.3(a) and §4.2. |
| FA-12 | Omitted evidence that anti-alarmism is not idiosyncratic; moral-hazard argument dropped | L5 line 122 (Amodei, “Avoid doomerism”); E4 line 66 (Altman, “trap of doomerism”); L5 §4.1 (moral hazard). | Fixed. Added to §7.3(c), §9.2 (Altman), In brief bullet 4, and §7.4 as argument 2 with its limit. |
| FA-13 | “They are an adversary” wrongly used as evidence of audience tailoring | E1 lines 414–417 (Dwarkesh, paired with dialogue). | Fixed. §9.1 pattern 6 rewritten with lower confidence; §9.1 China row assessment revised. |
| FA-14 | “Ulterior reasons” quotation truncated; source overstated | S6 line 108; Fortune (live check: full sentence plus “It is irresponsible, and I don’t know what their motives are”). | Fixed. §5.6 quotes the full passage, attributed to CBS Sunday Morning as reported by Fortune; Appendix C #3 updated. |
| FA-15 | Escalation “tracks the policy stakes” is a motive reading presented as the pattern | E1 assessment (the reading is E1’s interpretation). | Fixed. §9.1 pattern 2 gives both readings with confidence and notes the evidence cannot separate them. |
| FA-16 | T1 headline states the uncharitable reading; “seventeen minutes” wrong | Transcript [48:58] to [1:16:05] = 27 min. | Fixed. Retitled “Evaluation awareness: mechanism accepted, no method offered”; interval corrected. |
| FA-17 | §7.3(d) lacks GAIN caveat; §9.1 “each concrete proposal opposed” ignores 2023 licensing | E1 (GAIN remark; Dally); E4 line 99 (Zvi). | Fixed. Caveat and Zvi’s charge added to §7.3(d) and T13; §9.1 Regulation row softened (“most specific new AI measures since 2025”; 2023 licensing noted). |
| FA-18 | Audit called “voluntary”, which he never said | Transcript [51:20]. | Fixed. §10.3 now says whether audit is mandatory is not stated; financial-audit analogy quoted. |
| FA-19 | Post-recording disclosures used against claims made before them | §1.4 dating; E4. | Fixed. “(post-recording)” convention added in §1.5 and applied in §2.3, T3, T5, §6.2 C117, §9.2, §9.3 #1, §10.4 #2, Appendix A1 C117, Appendix D. |
| FA-20 | “Nvidia now owns the victim” | 8-K (closing expected H1 2027); §2.2; C057. | Fixed. T5 (“has agreed to buy”); also §3.5 (see FI-7). |
| FA-21 | Referent of “That paragraph’s fantastic” treated as settled | Transcript [51:20]; S3 line 89. | Fixed. Flagged as uncertain in §3.6 and Appendix A2 C109; audit endorsement kept as clear, “buy time” endorsement described as unclear. |
| FA-22 | Trump clip characterised inconsistently; Klein’s “any kind of regulation” gloss unflagged | E3/E1 (CNBC reading); FC C077. | Fixed. §2.3 row marks the referent as disputed; C077 added to §6.3 point 7. |
| FA-23 | Demeanour shown one-sidedly; Witt as “close precedent” | Transcript [19:50], [47:21], [53:36], [1:45:28]. | Fixed. Counterweights added to §5.4 and §2.1; Witt episode recast as background (see also FI-22). §5.3 #11 kept, since the lines it quotes are accurate. |
| FA-24 | Premises “validated” on the data they came from; formation stories presented as causes | E1 (out-of-sample statements). | Fixed. “Predicts” replaced by “fits” (In brief, §4 intro); out-of-sample check added to §4 intro naming where P7 fails; formation-to-premise links labelled “Reading (his own account)” in §2.1 and P1, with a general caveat. |
| FA-25 | A6 overstates (“Fear, not real grievance, is the main obstacle”) | Transcript [1:40:15]: industry’s failures listed first. | Fixed. A6 restated as “Doom narratives materially add to local opposition”; confidence lowered to medium. |
| FA-26 | Energy rebuttal leads with the uncharitable reading | §4.2 Energy already led with electricity; L4 had the charitable reading. | Fixed in part. §4.2 Energy now states the electricity reading explicitly and moves total energy to a parenthesis; §6.2 C207 and T10 reworded to contest the cause, not the flat trend. The C207 verdict (misleading) was left unchanged. |
Fidelity review#
| ID | Issue | Verification | Outcome |
|---|---|---|---|
| FI-1 | Liability verdict stated more strongly than the document’s own evidence | As FA-5. | Fixed (see FA-5). C108 kept as misleading on the narrow meaning (relief requested to enable pacing). |
| FI-2 | Klein’s gloss presented as Selsam’s words (T1) | Transcript [48:21]; S3 note 7. | Fixed. T1 and §3.6 now distinguish Selsam’s sentence from Klein’s gloss. |
| FI-3 | T1 interval wrong | As FA-16. | Fixed. |
| FI-4 | “That paragraph’s fantastic” referent | As FA-21. | Fixed. |
| FI-5 | “What’s stopping them?” probably Huang’s, treated as Klein’s unanswered question | Transcript [1:18:11]; S5 line 102. | Fixed. Added to §1.4; §3.9 and §8.3 reworded. |
| FI-6 | “Huang assents” at [1:20:03] rests on an inferred “Yeah” | Transcript [1:20:03]; S5 E13. | Fixed. §1.4, §3.9, §3.14 #3, §10.3 and Appendix A1 C168 now mark the assent (and the “should not ship” correction) as inferred. |
| FI-7 | Nvidia does not own Hugging Face | 8-K; transcript [38:32] (“while it was your product”). | Fixed. T5 and §3.5. |
| FI-8 | Wilson Research Group misquoted (“routinely”) and overstated | S5 line 97 (correct summary); L5 line 37 (source of error). | Fixed. §4.4 and §7.2 rewritten (“roughly one to one on average… 5-to-1 ‘not unusual’ in processors”; “as much effort… far more in processors”); L5 error recorded in Appendix B. |
| FI-9 | In brief says “none” disclosed; body says “most” | As FA-7. | Fixed. |
| FI-10 | Klein said to favour audit; Klein said to accept “engineering failure” | Transcript [39:02], [54:44]–[54:57], [1:10:51]. | Fixed. §3.14 #3 no longer attributes audit to Klein; §10.3 says Klein reports the labs’ view and restates Huang’s. Appendix C #9 notes that the column might settle Klein’s view on audit. |
| FI-11 | Prediction presented as commitment; critics’ reaction misdescribed | As FA-4. | Fixed. |
| FI-12 | (a) Agent message quoted from Wikipedia as primary; (b) “most aligned model yet” | METR (live check of the “peers” sentence); E3 line 217. | Fixed. (a) replaced with METR’s sentence and OpenAI’s “arguably unauthorized… Yet goal solution” (E4); (b) §2.3 now quotes the system card wording supplied by the reviewer. |
| FI-13 | Anthropic monitor finding quoted selectively | Reviewer’s live check of Anthropic’s 9 September assessment. | Fixed. T1 and §4.2 now say the monitors missed one of four incidents and caught the other three. |
| FI-14 | “Less than a year” ROIC stated as fact; content-per-GW explanation missing | Appendix B; S5 line 131; FC C172. | Fixed. §6.2 C172 row and §6.3 #1 rewritten, flagged as from an unofficial transcript, with the content-per-gigawatt reading added. |
| FI-15 | Mis-cited sources: (a) “ulterior” to E1; (b) “not a nuclear reactor” as Huang’s; (c) “racing ahead” as Huang’s own post; (d) class action to E4 | S6 line 108; E1 lines 107–110, 402–404; E3 line 225. | Fixed. (a) see FA-14; (b) §9.1 labels it Nvidia’s chief scientist’s line; (c) T13 attributes it to an Nvidia statement in Huang’s name; (d) §7.3(e) cites E3. |
| FI-16 | T8 misstates the schooling study exchange | As FA-10. | Fixed. |
| FI-17 | Paraphrases in quotation marks and small misquotations (12 items) | Each checked against the transcript or the named source. | Fixed. “lawless behavior” (§5.1); “I don’t think it’s necessary” (§4.2 Geopolitics); “with agency” (T6; T9’s “has agency” is correctly the Mad Money quote and was kept); “cover tracks” unquoted (§3.5, §6.2); “just going through their transition” (§10.1 #13); C035, C056 (Appendix A1); “every one of them made great contributions” (§3.7, A2 C137); brackets removed from “a, if you will” and “No[,] software”, with the ambiguity noted in §1.4; “virtually prototyped the chip” (§2.1); “30 days” quotation split (§2.1); 19% figure attributed to the Stanford paper (§9.2); Crane and Soto comparison restored (§7.3(j)). |
| FI-18 | “No question in my mind” attached to the wrong proposition | Transcript [11:29]. | Fixed. §4.2 Work quotes “I believe there’s going to be a net creation of jobs”; §4.2 Knowledge and §10.5 give the ambition sentence in full. |
| FI-19 | OS-vocabulary critique says he omits what he said himself | Transcript [1:03:30] (“parent and child… give birth”). | Fixed. §5.2 table row, §7.3(h) and §3.8 rewritten. |
| FI-20 | Zvi summarised selectively; company list incomplete; “security mindset” unverified | Live check of Zvi’s post: all confirmed, including “Engineering mindset is different from security mindset.” | Fixed. Bottom line and “outright lie” added; list given as “among others” (now including Johnson & Johnson and Meta); quotation kept, since verified. |
| FI-21 | “Delayed IPO timelines” unsupported | E3 timeline does record OpenAI’s IPO slipping to 2027 alongside Amodei’s essay, but no working file ties the slip to safety costs; E4 line 256 has OpenAI’s “great cost and delays”. | Fixed (reviewer partly wrong that no file mentions it, but right that it does not support the point). Replaced with OpenAI’s paused RL run (“great cost and delays”) and Anthropic’s redeployment. |
| FI-22 | Witt anecdote drops the discrepancy between reviews and the year’s source | E2 lines 195, 372. | Fixed. §2.1 and Appendix C #10. |
| FI-23 | Smaller points: timestamp; T7 conditional; T12 segment; C110 speaker; guarantee nuance; GLM citation | Transcript [1:29:48], [1:19:12], [27:02], [52:16]; E3 line 95; L6 line 40. | Fixed. All six. The residual-value description of the guarantees rests on the reviewer’s reading of the 8-K (E3 does not use the phrase); flagged for confirmation in Appendix C #20. |
Completeness review#
| ID | Issue | Verification | Outcome |
|---|---|---|---|
| CO-1 | No values layer or self-conception | L1 §1.11; L3 §7; S1 reveals §5; E2 §§3–4, 9, 12.3–12.6. | Fixed. New §4.5 (eight values with anchors; the paternal model with both readings; the moral asymmetry towards speech; what is absent). Linked from §4.1, §10.1 #13 and the In brief. |
| CO-2 | No domain on the public, democratic authority and who decides | Transcript [15:04], [40:21], [51:20], [1:02:59], [1:03:30], [1:40:15]; L3 §7; E1 §§1–2, 4. | Fixed. New §4.2 domain; §10.1 #14; §10.4 #13; accountability row in the §10.3 table. |
| CO-3 | Crux merges Klein and labs; understates Huang’s development-stage pause; omits chip layer | E1 (Dreamforce, Scotland, All-In); E3 §§5, 8.1, 9.1; E4 §§1.2–1.3. | Fixed. §10.3 table (see FA-3); “take a pause” and “hold it back” added to §4.2 Safety, T2, §7.1, §7.4 and §10.5; chip-layer row and §10.4 #12; T2’s charitable reading revised. |
| CO-4 | T9 omits counter-instances ([44:17] “extraordinary care”; [1:37:36] “bigger game… safety”; All-In “extraordinary standards”) | Transcript lines 215, 545; E1 line 252. | Fixed. Added to §3.6, §3.11, §3.14 #2, §5.1 (“mainly in one direction”) and T9 (confidence lowered to low–medium). |
| CO-5 | Distributed-defence (“cybersecurity”) safety model missing | E1 (Rogan [A], Dwarkesh, Lex Fridman, CNBC 3 September); Nvidia blog (company, not Huang); S5 (HF security agent failure). | Fixed. New paragraph and strain in §4.2 Safety; T12 charitable reading in his own words; Nvidia line cited as the company’s. |
| CO-6 | “Verification before commitment” not a core premise | E2 §7, 12.2; transcript. | Fixed. New P8 with anchors and strain; “eight premises” in In brief and §4.1; cross-referenced in A2, T1 and §10.2. |
| CO-7 | No domain on power and concentration | L4 Part 4–5; E3 §§4–5; E2 12.7. | Fixed. New §4.2 domain “Power, concentration and the platform”; §4.4 bullet; §10.1 #15. Trump’s July 2025 “break this guy up” remark was not added (context only; not needed for the argument). |
| CO-8 | Technology model ignores incident scale, persistence and self-organisation | S2 §6; METR (live check); E4 §1.4. | Fixed. Added to §4.2 Technology strain and §9.3 #1, with OpenAI’s 100x harness finding as the counterweight. |
| CO-9 | Reasoning-style gaps: norm versus prediction; acquaintance as evidence; concede execution, contest structure | S5 E13; S4; L3 §7; S1 §9; S6. | Fixed. §4.3 items 11–13; §3.14 #2; §10.2 bullet; §3.9. |
| CO-10 | No domains for human nature or education and cognition | L1 §1.10; E1 §7; E2 12.5; S1 §12. | Fixed. Two new §4.2 domains. Not used: the “commodity… Character, humanity” and “I don’t think my chips will feel those” quotations, which the reviewer could not find in the Lex Fridman transcript; flagged in Appendix C #21. |
| CO-11 | T6 under-explained: management philosophy; pacing conflated with product defect | Transcript [1:32:23], [53:36]; S6 §5; S3 §11. | Fixed. Fuller [1:32:23] quotation in §3.11; T6 charitable reading and §4.4; new §5.3 move 12; race-language tension noted in T6. |
| CO-12 | Work domain “largely silent” ignores his record; Amodei nuance; Klein’s [16:19] inversion | E1 §3; E2 12.5, §13; FC C014; transcript [16:19]. | Fixed. §4.2 Work (conditional optimism; acknowledgements elsewhere); §3.2 (Amodei nuance, Klein’s inversion); §6.3 #4. |
| CO-13 | Not self-sufficient on sources | URLs extracted from the working files. | Fixed. New Appendix D (91 entries, about 110 URLs, grouped, with tiers and blocked/post-recording markers); §1.5 points to it. |
| CO-14 | RSI referents not distinguished; 2017 statement complicates T11 | L6 §7; S5 E8; E1 §7; E2 §8, §13. | Fixed. §3.9, §5.1, T11 and §9.1 Nature-of-AI row. |
| CO-15 | Anthropic incidents and “did no harm” missing from T4/T5 | E1 §1 (All-In [A]); E3 §9.1; L4; E4 §1.4; L6 §6. | Fixed. T4, T5, §2.4, §6.2 and Appendix A1 C117. |
| CO-16 | Huang’s five-layer-cake essay not cited | S1; L1 §1.3; L3 §4.1. | Fixed. §3.1, P6, §4.2 Energy. |
| CO-17 | Formation omits relevant episodes | E2 §§3, 7, 9, 10, 12.1, 12.3, 12.5–12.7. | Fixed. §2.1 threads 6 (retreat and asking for help, low–medium confidence) and 7 (wonder and fear); American dream and Taiwan; reasoning in public; the rest folded into §4.5. |
| CO-18 | Interests gaps: Taiwan, open-models-and-sales quote, rental revenue, Nscale, Anthropic IPO, personal stake, open-weights letter, NYT litigation | E3 §§3.3, 4; S2 O7; S5; S4 §2; L4 Part 5; E4 §1.3; E3 §2. | Fixed in part. All added (§2.2, §3.10, §7.4, §8.4). Unresolved: the exact source of “Whenever there’s more use, you’ll have to sell a lot more NVIDIA computers” (S2 gives only “July”); the current status of the NYT–OpenAI litigation (added as background, marked unchecked). |
| CO-19 | Government domain lacks level-of-government view and 2023–26 hardening | E1 §2; E3 §§8.2, 9.1. | Fixed. §4.2 Government (federal preference, All-In audit specifics, “actual problems”, EO 14409, 2023 licensing); §9.1 Safety row; T3 evidence (Dally, “resides exactly where we put it”). |
| CO-20 | Geopolitics lacks dialogue content, values case, security | E1 §4; E4 §§5.1–5.2, 6; S6 §5; L4 Part 4. | Fixed in part. Added “agree on what not to use the AI for”, the adversary-plus-dialogue line, CAISI finding, Kantrowitz’s values point, divergence from Kratsios and Moolenaar, and the absence of military uses. Not added: “childish”, which the working files do not give in context. |
| CO-21 | Tempo not a named dimension | L6 §5; Dreamforce (reviewer’s live check); transcript. | Fixed. Tempo is a named substantive dimension in §10.3. |
| CO-22 | No consolidated list of stated conditions and confidence | L1 §§2.4, 2.8. | Fixed. New §10.5 table (eleven conditions) and confidence paragraph; cross-referenced from §8.3 and §9.1. |
| CO-23 | Smaller §3 omissions | Transcript lines 74, 77, 89, 386, 416, 419, 467, 485, 509, 557. | Fixed. All items added to §§3.2, 3.3, 3.6, 3.7, 3.8, 3.9, 3.10, 3.12, 3.13. |
| CO-note | Trim the §6.2 table by a third to hold length | Task instruction not to shorten substantially. | Rejected. The table was kept in full; the length increase is recorded above. |
Remaining concerns#
- Length. The additions requested by the completeness review raised the document by about 40%. Sections 4 and 10 are now much fuller; the In brief remains the entry point.
- Reviewer-only verification. Several added details rest on the reviewers’ live checks of primary sources that are not in the working files: the Astra system card wording, the Anthropic monitor finding (“caught the others”), the Hugging Face timeline figures, the full Dreamforce “take a pause” sentence, the residual-value structure of the Ohio guarantees, and OpenAI’s 21 September “would not be licenses” wording (E3, via an archive copy). They are cited in good faith; Appendix C lists the ones worth confirming.
- Klein’s own position in the §10.3 table rests on his episode notes (L6); his column and his unstated proposal [54:44] were not read.
- Grading adjustments in §6.1 are this document’s judgement. The fact-check file itself has not been regraded, and Appendix A keeps its original verdicts with notes.
- Appendix A was corrected only where the reviews flagged rows. Its claim wording is the fact-check’s paraphrase and has not been re-audited line by line.