How this works

Each paper is read by several AI reviewers, each looking for one kind of slip: numbers that do not add up, tables that disagree with the text, wrong pointers and citations, claims the results do not support, and, for papers with proofs, steps that do not follow. Everything they raise is located on the page and given to a judge model, which confirms or rejects it after looking at the page itself. Arithmetic is recomputed by code. The blog post tells how it was built and what it found.

The reviewers and the judge are Claude models (Anthropic). Each confirmed finding of the three more serious kinds is then checked again by a model from another company, Gemini (Google) or GPT (OpenAI). If that check disagrees, or cannot tell, the finding is not shown here. Typos are shown in a separate, collapsed group and are mostly checked by the judge only.

The kinds: Substance, a reader would take away a wrong number, direction or fact; Reference or label, a wrong citation, pointer, unit or label; Wording, the text overstates or misdescribes the data; Typo.

The PDF checked is the version listed on the paper's page; a finding may already be fixed in a later version or an erratum. Page numbers are those of that PDF.

Please tell us when a finding is wrong, unclear or already corrected: every report is read, and wrong findings are taken down.