250 machine-generated murder mysteries (the MuSR murder_mysteries split β inspect_evals' default domain), each a ~1,000-word narrative whose whodunit question takes multi-step soft reasoning over the story, imported from inspect_evals. The model picks a suspect through the eval's own choice prompt; grading is inspect's own choice scorer against the published key, so no judge tokens are spent. Cases are served in a fixed seeded shuffle (the inspect_evals randomizes order per run); benchmark score = plain accuracy over the cases run. Answer format: end the reply with a line like 'ANSWER: B', the letter of one option. A period after it still counts, as in 'ANSWER: B.', and so does a wrapper around the letter, as in 'ANSWER: $B$', 'ANSWER: (B)', 'ANSWER: **B**' or 'ANSWER: \boxed{B}', or a bold label, as in '**ANSWER:** B'. Words after the letter, as in 'ANSWER: B because it is red', or a reply with no 'ANSWER:' line, as in 'The answer is B', are read as no answer and score 0, even when the letter is right.