350 Cases from InfiniteBench's Math.Find task: find one of seven key numbers (the largest three, the smallest three, or the median) in a list of tens of thousands of numbers. Served as a local Task built from the inspect_evals eval, with inspect_evals' prompts. inspect_evals cuts each input to the answering model's context window; this Task cuts every input to one fixed budget, about 446,000 characters (128k tokens by inspect_evals' estimate), whichever model answers. No Math.Find input is that long (the longest is 120,155 characters), so every Case is inspect_evals' whole input. Graded the official InfiniteBench way: the first number in the reply, compared with the answer by value, so 'The largest number in the list is: 91' is right for 91 and no judge tokens are spent. inspect_evals grades with exact match on the whole reply, which scores that answer 0; this is its proposed fix (PR #2346), adopted here. Benchmark score = plain accuracy over the cases run. Inputs are long, about 118,000 characters each: use a model with a long context window. Answer format: the first number in the reply is the answer, compared by value, so '91' and 'The largest number in the list is: 91' both count. Any number before it is read instead: 'Of the 3 largest numbers, the largest is 91' is read as 3 and scores 0, as does a reply with no number.