1,000 maths questions about a picture (MathVista's testmini split: function plots, charts, tables, geometry figures, puzzles, scientific figures and photos, drawn from 28 existing datasets and three new ones), imported from inspect_evals. 540 are multiple choice and 460 free-form, answered with an integer, a decimal to one or two places, or a list. Each question comes with one picture, saved inside the Case at import, so nothing is fetched at run time. The Candidate is shown the picture, then the eval's own instruction: inspect's single-answer template over the lettered choices for a multiple-choice question, or a fixed 'ANSWER: $ANSWER' template around the dataset's own query, which carries the precision and unit hint, for a free-form one. Grading is the eval's own pattern scorer: the letter, or the text after 'ANSWER:', must equal the answer exactly (case ignored), so no judge tokens are spent; an answer inside a wrapper, such as 'ANSWER: $B$' or 'ANSWER: **42**', is read as the answer (our change, as inspect reads it as none); 'ANSWER: 1.50', '1.5 cm' or a trailing space score wrong against 1.5, so scores here are stricter than, and not directly comparable to, the paper's and the labs' own more lenient grading. Serving order is pinned by a policy seed (inspect_evals shuffles per run). Benchmark score = plain accuracy over the cases run. Answer format: make the last line of the reply 'ANSWER: 12'. On a free-form question the text after ANSWER: must equal the answer exactly, ignoring case and a wrapper around it such as 'ANSWER: **12**', so 'ANSWER: 12 cm', 'ANSWER: 12.0' and 'The answer is 12' score 0. On a multiple-choice question, write ANSWER: then the option letter, bare or wrapped, as in 'ANSWER: (B)'.