How it works (AI summary)
The authors curated 388 candidate benchmarks via LLM-assisted benchmark cards, screened them through a six-criterion audit and independent reproduction of eight shortlisted suites, and retained four public benchmarks (HealthBench Professional, PhysicianBench, ALE-CLI v1.0: Health and Medicine, and CHI-Bench) alongside a private diagnostic suite (InclusionMed-DxFiles, 312 NEJM MGH cases). Models were evaluated via direct API calls, a FHIR EHR agent, and CLI harnesses (Codex, Gemini CLI, and Claude Code), scored equally (20% each) into the InclusionMed Index, annotated across eight task attributes, and compared against token-based API cost estimates.