53 open-answer questions from college exams, quizzes and textbooks across MMMU's 30 subjects: the open-answer part of MMMU's validation split, since the test split's answers are withheld. Imported from inspect_evals. Every question comes with a picture (one with two), so the Candidate must read images; the pictures are saved inside the Case at import and nothing is fetched at run time. The Candidate is shown the pictures, then the eval's own line 'Answer the question using a single word or phrase.' above the question, and answers in free text; most reference answers are a number, and a few list the accepted forms as written in the dataset (['24/7', '3.429']). An LLM judge compares the reply with the reference answer using inspect's model_graded_fact prompt, which quotes the question as the dataset writes it, without that line, with every judge call routed and metered through our gateway. Benchmark score = plain accuracy over the cases run. The eval lets the model under test grade itself; this Benchmark pins our house judge. inspect_evals runs the model at temperature 0; this benchmark imposes no generation settings. Answer format: no fixed format; the question asks for a single word or phrase. A judge reads the whole reply and checks whether it contains the content of the reference answer, ignoring differences in style, grammar and punctuation.