About these corrections

The audit covers my published research: the articles, book chapters, pieces of commentary and textbook listed on this site. Unpublished manuscripts are outside it. Each work was audited in its published edition, with its official supplements and corrections where available. A software package counts only as the implementation of the method of a published work.

An error is a confirmed defect: a misprint, or a statement that is false or unsupported under its natural reading. Loose wording is text that the authors meant to write and that is imprecise, too strong or ambiguous as written. If a competent reader takes the intended meaning and that meaning is correct, loose wording is a clarity note, recorded separately from the errors.

Each error has a severity. Severity measures the consequence for a reader, not whether a sentence is literally false.

• Critical: a headline conclusion is false or unsupported in the setting the paper targets, and no routine repair restores it.
• Major: a central result, method or key claim fails for a non-trivial reason, and the repair changes what the paper establishes: a new assumption its applications may not satisfy, a different estimand, a weaker conclusion. If a minor regularity condition repairs it, it is not Major.
• Moderate: a real substantive error with limited reach.
• Minor: real but inconsequential.

Earlier reviews of my work had been conducted by several large language models, including Sol 5.6, Opus 5 and Astra 6, in a nonsystematic manner. I used publicly available materials for those reviews and no internal research documents.

Two AI models carried out the initial audit. The identifiers and runtime versions below describe that run. I call them Opus and Sol. Opus is Claude Opus 5.5 (model identifier claude-opus-5-5), run in Claude Code. Sol is GPT-6.1-Sol (model identifier gpt-6.1-sol), run through the Codex command-line tool, version 0.159.1.

Each work was read twice, once by each model, in separate runs, or sessions, with the earlier lists of errors and the other read withheld. For the paper on combining list experiments and direct questions, both readers also inspected a maintenance report describing earlier discrepancies; they reported doing so after their own calculations. Sol also tried to reproduce the headline tables and figures from public replication materials where available, or by recalculations from reported numbers. Opus merged the two reads, the results of the reproduction and the lists left by earlier rounds of review into one list of possible errors, or candidates. AI models had also carried out those rounds.

Sol checked every candidate against the published source, in a session that had not raised it. It gave a verdict on the facts and no rating. Candidates expected to be minor and kept in batches were checked against the page, and that initial check had no separate re-run. Every other candidate’s initial check required evidence that another person can re-run, and another Sol session tried to re-run that check. In order of preference, the evidence is a proof checked by Lean (a proof assistant) or an exact counterexample, a complete derivation, a calculation reproduced from source data, or a simulation with stated uncertainty. Of the candidates kept in batches, those first raised in this audit and recorded as not confirmed were later each sent to a Sol session of its own for an independent check against the source and the evidence. Only the part of a check that a program carries out does not depend on a model’s judgment.

In the original audit, Opus gave each candidate a class and rated each error by the severity rule, on the reported facts which it could not overrule. It also drafted the original version of this page and the corrections. A Sol session that had not checked the candidate then checked each correction and made any needed changes.

GPT 6 Astra was responsible for a final round of edits to this site.

The Lean formalizations were produced by Sol and Opus. They state selected results in Lean, a proof assistant, as printed or as corrected, with any restriction made explicit. The current package has checked Solution proofs for its Challenge statements. A status belongs to one statement, never to a work. A proof establishes the formal statement under its stated assumptions; correspondence with the printed or corrected passage depends on a separate comparison by another Sol session.

The archived supporting library includes conditional results with explicit published-theorem premises. The current target statements and their definitions specify their own assumptions.

I have not examined the Lean proofs personally. Their statements and proofs are maintained in llm-audit-lean for independent checking and because some may be useful to others. Each formal target on a work’s page links to its current statement and proof, with its assumptions and any restriction.

I set the rules before the work began, among them the works in scope, the edition, the roles of the two models and the severity rule. My review is ongoing. Other findings remain for my decision. I do not apply my comments or suggested severities directly. I rely on independent checks that withhold my identity when testing facts or mathematics, and withhold my comment and the current rating when reassessing severity. I incorporate what those checks support. Where a check disagrees with me, I consider both positions before deciding. When an entry changes, I preserve a link to its classification before my review.

My comments on the corrections are published, exactly as I last saved them, with the entries they concern and on the page My comments, each with the history of the entry. Where the evaluation did not adopt my suggestion, both positions are shown. I decided to defer to the models' judgment in each of these disagreements, so the classification shown is the one that the models settled.

I reviewed 100 randomly selected minor errors and 50 randomly selected clarity notes. I found difficulty in finding a distinction between minor errors and clarity notes, but did not find any that warranted escalation to moderate, major, or critical status.

Each correction and each clarity note has a permanent identifier, which names the work and the entry. The identifier will not be reused for another entry. It is also the anchor used for links directly to the entry on the page of the work, and it identifies the entry in the data files.

A citation of a correction first cites the work itself. It then gives the identifier of the correction, the title of this site, the date printed at the foot of the page, and the address of the entry, which has the form shown below.