1,319 grade-school math word problems (the GSM8K test split), imported from inspect_evals. Each opens with the eval's 10 worked examples, sent as leading input text rather than a system message (a named deviation: a benchmark cannot address a candidate's system role). The model reasons step by step and commits a final numeric answer; grading is inspect's own numeric match against the published solution, so no judge tokens are spent. Benchmark score = plain accuracy over the cases run. Answer format: end the reply with a line like 'ANSWER: 18'. The last number in the whole reply is graded, with dollar signs and commas ignored, so 'ANSWER: $18' also counts, but any number after the answer is read instead: 'ANSWER: 18, from 9 eggs at 2 each' is read as 2 and scores 0.