An MMLU Score Is Not Comparable Without Its Evaluation Frame
A model can post 0.781 and 0.79 under the same MMLU label, yet those numbers may be incomparable because the split, prompt, grader, runner, and network access changed. This deep dive shows how content-addressed evaluation frames and hashed claims make those hidden choices explicit, while structural validity still does not prove a score is correct. The practical shift: compare references, not naked benchmark numbers.