{"slug": "the-verification-record", "title": "The Verification Record", "summary": "Mark Pesce of the University of Sydney published a set of papers in July 2026 arguing that work should travel with a record of its checking. The papers were produced using an AI system (Claude by Anthropic) as generator and a separate adversarial checker to verify factual claims and citations. The verification record details 75 findings from four adversarial reports, with most repaired, and notes that the process does not yet meet its own standard.", "body_md": "# The Verification Record\n\n[Mark Pesce](mailto:mark.pesce@sydney.edu.au) · University of Sydney · July 2026\n\n*This note accompanies four papers: Defending the Loop, The Check and the Firm, Delayed Neutrons, and The Principles of Loop Governance, together with the previously published Foundations of Post-Watershed Economics. The papers argue that work should travel with the record of its checking. This document is the narrative account of that checking and a finding-level ledger of what it found. It is not formal verification, and it does not yet meet its own standard: the items that remain open are listed in Section 5. This document was itself reviewed twice by the same adversarial process on 11 July 2026; the ledger includes all sixteen findings from those reviews, and this is the twice-repaired version.*\n\n## 1. Method\n\nEach paper was produced by a loop. The generator was an AI system (Claude, Anthropic), drafting and repairing against instructions. The specification was held by the author, who set register, structure, and style, ratified or declined structural changes, and edited directly between passes. The checker was a separate adversarial review, instructed to verify factual claims, citations, and source-use against the sources themselves. The checker was separate from the generator and chosen to reduce correlated error; the degree of correlation between their failure modes was not measured, and separateness is all this record claims. Checking ran in both directions at least once: the generator audited the author's edit pass on the first paper and reported five defects, of which the author accepted three and declined two.\n\nTwo classes of check applied, on different rungs. The adversarial reports are the strong checks: independent and source-anchored. The mechanical audits are the weak checks: style-rule scans, citation-coverage counts, and banned-construction searches, written and run by the generator on its own output; they establish conformance to the author's style constitution and nothing more.\n\nThe process facts in this record, drafting sequences, instruction counts, acceptance and declination decisions, are the account of the participants, generator-compiled and author-ratified. No prompt log, transcript, or separate approval register is attached; the session transcript exists and its retention is listed as an open item.\n\n## 2. Counting rule\n\nA finding is one top-level entry in an adversarial report, as enumerated by the report itself. Several entries contain multiple atomic errors (the first Defending the Loop entry contains five Ariane corrections; two Delayed Neutrons entries contain three and two sub-errors); these count once. Under this rule the four reports contain seventy-five findings: nineteen, twenty-two, nineteen, and fifteen. A different grouping rule would produce a different total.\n\n## 3. The ledger\n\nDispositions: **R** repaired; **RR** repaired by retreat, the claim moved to a weaker, supportable form; **P** repaired, author ratification pending; **O** open, the finding remains unresolved. All repairs were applied by the generator; foundational changes on the first three papers were ratified by the author in session; the Principles repairs await ratification and are marked accordingly.\n\n**Foundations of Post-Watershed Economics** (April 2026) predates this process and did not pass through it. Its published acknowledgements state its method and assign responsibility for errors to the author. It inherits none of the assurances below.\n\n**Defending the Loop**: five drafts under the author's direction; author line-edit pass; adversarial report of 19 findings, all repaired. The opening was rewritten in the process: it no longer asserts that the software was not wrong, resting instead on the supportable fact that the fatal decisions were faithful to premises inherited from the previous rocket.\n\n| ID | Finding | Disposition |\n|---|---|---|\n| DL-01 | Ariane 501 account: timing, failure order, conversion target, \"software not wrong,\" specification reuse (five sub-errors) | RR |\n| DL-02 | Claim the code passed the strongest checking of its day | RR |\n| DL-03 | Absolute claims for proof lacking a stated trusted base | RR |\n| DL-04 | \"Specification is not a proxy\" contradicted the paper's own gap argument | R |\n| DL-05 | seL4 verification scope overstated (C-level 2009; binary and aircraft later, separately citable) | R |\n| DL-06 | \"Standard where failure kills\" overstated (DO-333 permits, does not require) | RR |\n| DL-07 | Knight Capital \"sold within the year\" (merged) | R |\n| DL-08 | Investment-compliance practice mischaracterised as post-hoc sampling | RR |\n| DL-09 | A-Lab episode misstated against the published correction | R |\n| DL-10 | \"Exhaustive exploration\" misdescribed as the semiconductor standard | R |\n| DL-11 | Huntley citation: wrong title and year | R |\n| DL-12 | FDIV note claimed more than its source supports | RR |\n| DL-13 | Three notes missing author lists; topic page mislabelled a survey | R |\n| DL-14 | \"Careers to hours\" comparison unsupported by citations | RR |\n| DL-15 | \"Dependably learns to fool\" overstates the cited safety paper | RR |\n| DL-16 | A-Lab note out of date against the formal correction; C&EN figure wrong | R |\n| DL-17 | Wiener 1948 first-edition publishers wrong | R |\n| DL-18 | Comma splice | R |\n| DL-19 | Comma before example specification | R |\n\n**The Check and the Firm**: author draft; seven structural critique points from the generator, all accepted by the author; full redraft; two author-directed additions; adversarial report of 22 findings, twenty-one repaired and one open pending publication, with the partnership-liability passage rewritten against the current New South Wales Partnership Act.\n\n| ID | Finding | Disposition |\n|---|---|---|\n| CF-01 | Alchian-Demsetz metering materially misrepresented | RR |\n| CF-02 | Porter cited for inimitability that belongs to the resource-based view; date incomplete | R |\n| CF-03 | Williamson reduced to hold-up | R |\n| CF-04 | Lineage conflated with true defect rate | RR |\n| CF-05 | Failure rate alone treated as sufficient for expected liability | RR |\n| CF-06 | Append-only record treated as proof; theft claim overstated | RR |\n| CF-07 | \"Actuarial pricing assumes rough independence\" false in general | R |\n| CF-08 | 2008 models described as pricing correlation as independence | R |\n| CF-09 | Lloyd's Register and UL histories overstated | R |\n| CF-10 | Claim that lineage ends joint-and-several liability legally wrong | RR |\n| CF-11 | Turing Police forecast presented as existing structure | R |\n| CF-12 | Companion-paper note pointing at site homepage | O (closes on publication of the direct URL) |\n| CF-13 | Four-component harness definition attributed to a source that lacks it | R |\n| CF-14 | PEXA statistic uncited by the linked page | R |\n| CF-15 | Porter publication date | R |\n| CF-16 | Two Wikipedia notes where legal and institutional sources required | R |\n| CF-17 | Workflow/workflows agreement | R |\n| CF-18 | Malformed apposition in conveyancing sentence | R |\n| CF-19 | Authorisation/authorisations agreement | R |\n| CF-20 | \"The second this paper adds:\" incomplete syntax | R |\n| CF-21 | Comma splice | R |\n| CF-22 | \"Liability apportions per member\" grammar and law | R |\n\n**Delayed Neutrons**: compressed draft rejected by the author as inaccessible; plain-language rewrite at double length; fifteen self-narrating passages removed at the author's instruction; cybernetic ancestry added; adversarial report of 19 findings, all repaired, including two mathematical corrections to the treatment of accelerating growth.\n\n| ID | Finding | Disposition |\n|---|---|---|\n| DN-01 | \"Audit once and rely on forever\" too absolute | RR |\n| DN-02 | Reward hacking treated as inevitable; \"only kind that compounds\" | RR |\n| DN-03 | Erdős result misdescribed in three details | R |\n| DN-04 | Gödel machine's goals mischaracterised as rewritable | R |\n| DN-05 | Löbian obstacle stated as impassable wall; note lacked draft date | R |\n| DN-06 | Chain-letter example wrong; shrinking doublings conflated with finite-time blow-up | R |\n| DN-07 | Reactor control stated too absolutely; inherent feedback omitted | R |\n| DN-08 | Ashby's law misapplied to numerical singularity | R |\n| DN-09 | \"Nothing keeps any neutron delayed\" contradicts decay physics | R |\n| DN-10 | Insurance repricing overstated; road-safety comparison unsupported | RR |\n| DN-11 | \"Before fire codes existed\" false (1667 Rebuilding Act) | R |\n| DN-12 | Lloyd's List named where Lloyd's Register meant | R |\n| DN-13 | Lineage treated as loss frequency | RR |\n| DN-14 | AIG figures: committed/drawn/repaid conflated | R |\n| DN-15 | Price-Anderson 1957 and modern schemes merged | R |\n| DN-16 | Dependence pricing called unsolved mathematics | R |\n| DN-17 | Sentence fragment | R |\n| DN-18 | Number agreement (\"seconds is a tempo\") | R |\n| DN-19 | Missing article (\"law of nature\") | R |\n\n**The Principles of Loop Governance**: draft; twelve self-references removed at the author's instruction; adversarial report of 15 findings, all repaired, three foundational: the register scheme was rebuilt (theorems to ancestry; five selection claims and two prescriptions), forecasts restored to prospective tense, and the seventh principle renamed The Judgment of Worth.\n\n| ID | Finding | Disposition |\n|---|---|---|\n| PL-01 | Register classification internally inconsistent | P (foundational) |\n| PL-02 | Principle I stated corruption as inevitability | P |\n| PL-03 | Strathern's formulation attributed to Goodhart | P |\n| PL-04 | Own-cause maxim attributed to Montesquieu | P |\n| PL-05 | DGM cited as judge-and-judged reunited; exhibit misassigned | P |\n| PL-06 | Observability treated as mandating records | P |\n| PL-07 | Good-regulator theorem stretched to the audit record | P |\n| PL-08 | Attestation claimed to establish completeness | P |\n| PL-09 | Delayed Neutrons forecasts stated as operating institutions | P (foundational) |\n| PL-10 | Ashby overstated; plurality presented as theorem | P |\n| PL-11 | AIG example misfit to the claim | P |\n| PL-12 | \"Stein's inequality\" does not exist | P |\n| PL-13 | Incomputability claimed without proof; renamed | P (foundational) |\n| PL-14 | Vocabulary semicolon fragment | P |\n| PL-15 | Delayed-fraction list coordination | P |\n\n**The Verification Record** (this document): first version reviewed 11 July 2026, report of ten findings; repaired version reviewed the same day, report of six findings, which confirmed the major corrections and is appended here under the record's own policy.\n\n| ID | Finding | Disposition |\n|---|---|---|\n| VR-01 | Not a walkable lineage: no artifact identifiers, hashes, check dates, or ratification record | O |\n| VR-02 | Broken and misleading artifact paths in the retained reports | O |\n| VR-03 | Finding counts not reproducible; no counting rule | R |\n| VR-04 | Dispositions asserted but not recorded | R |\n| VR-05 | Checker failure modes claimed uncorrelated without measurement | RR |\n| VR-06 | \"Every source exists as cited\" stronger than the reports support | RR |\n| VR-07 | Process and approval facts not independently verifiable | RR |\n| VR-08 | Sentence fragment in the findings summary | R |\n| VR-09 | Number and delimiter error before examples | R |\n| VR-10 | Unmarked parenthetical list | R |\n| VR-11 | VR-01 through VR-10 absent from the finding-level ledger | R |\n| VR-12 | Disposition code O used but undefined | R |\n| VR-13 | CF-12 marked repaired while acknowledged open | O (rides CF-12) |\n| VR-14 | Nonexistent eponym miscalled a law rather than an inequality | R |\n| VR-15 | Wrong preposition and fragment sequence in Section 4 | R |\n| VR-16 | Possessive ellipsis in the closing line | R |\n\n## 4. The findings, classified\n\nThe four paper reports contain seventy-five findings. No fabricated source was found in any report: every underlying work could be identified, though the reports found bibliographic, link, and source-use errors, wrong titles and dates, missing authors, wrong publishers, and pages that did not support the claims resting on them. The substantive findings fall into the classes the papers predict an unchecked generator will produce: register inflation (claims stated a rung above their support, theorems that were design rules, inevitabilities that were tendencies); tense promotion (the programme's forecasts drifting into the present, the corpus citing the corpus until prediction read as fact); precision decay (real events reproduced with plausible inaccuracies: the Ariane sequence, the AIG figures, the Price-Anderson structure); misattribution (right ideas hung on wrong sources, including a named inequality that does not exist); and grammatical residue in the single digits per paper.\n\nGenerator suggestions declined by the author (a restored roadmap clause, an earlier ending, a style purge on the first paper) stand in the record with the same weight as the repairs.\n\n## 5. Open items\n\nThe critic's verdict on the first version of this record was that a narrative account is not a verification record, and the following remain open until the acts they require are performed, all of them outside this document. Immutable copies of each reviewed draft must be preserved as reviewed, with stable identifiers or hashes; the current materials were edited in place, so the reviewed states survive only partially. The broken artifact path in the Check and the Firm report, and the Principles report's links into a since-repaired manuscript, need correction against preserved copies. The session transcript, which contains the author's ratifications and declinations, should be retained as the approval record this document cites but does not attach. The Principles repairs, and this document, await the author's ratification and a further checking pass. Until those items close, the lineage described here is attested but not independently walkable, and this record says so.\n\nFurther checks of this record append to this ledger as VR entries; the record is the accumulator, and its claims are stated to be recomputed, not trusted. The regress this invites terminates where the corpus says trust terminates: in checks that can be re-run, and in a signature.\n\nWhat the record shows is what was done: five papers, six adversarial reports including the two on this document, ninety-one findings under the stated rule, each with a disposition in the ledger above, and the distance still to travel stated plainly.", "url": "https://wpnews.pro/news/the-verification-record", "canonical_source": "https://thewatershed.markpesce.com/the-verification-record/", "published_at": "2026-07-14 06:29:57+00:00", "updated_at": "2026-07-14 06:52:12.729114+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-ethics", "ai-safety", "ai-research"], "entities": ["Mark Pesce", "University of Sydney", "Claude", "Anthropic"], "alternates": {"html": "https://wpnews.pro/news/the-verification-record", "markdown": "https://wpnews.pro/news/the-verification-record.md", "text": "https://wpnews.pro/news/the-verification-record.txt", "jsonld": "https://wpnews.pro/news/the-verification-record.jsonld"}}