The updated MLCR-AA methodology has the following specifications:
➤ Models receive a system prompt framing the task as medical record review, with the question placed after the full ~25k to ~64k token medical record
➤ Unless a question specifies its own format, models answer each part of the question in one paragraph, giving their conclusion and the supporting evidence from the record. With answer lengths now consistent across models, we no longer apply the concision test
➤ Grading uses a three-model judge panel of GPT-6.1 Sol, Claude Opus 5.5 and Grok 4.7, all at medium effort. Each judge lists any claim that contradicts the medical record and any key point of the reference answer the response misses, and a response passes when a majority of judges list neither. All three judges reach the same verdict on 69% of answers, up from 48% under the previous panel
➤ Every question is run three times, and we report the mean across repeats
Scores on the updated leaderboard are not directly comparable with earlier MLCR-AA results. MLCR-AA is not a component of the Artificial Analysis Intelligence Index.