Cochran's Q

A forensic reproducibility audit of a chest-radiograph vision-language model benchmark found that 297 of 300 planned model-prompt calls yielded nonempty reports, but 60 Claude calls labeled A/B were e…