Mistral and OpenAI produced identical overall eligibility labels across 200 synthetic patient-study assessments, with criterion-level disagreements concentrated in inclusion criteria where models switch between “Uncertain” and “Not included” when records lack sufficient evidence, so model choice should be driven by deployment architecture, control, and cost rather than assumed interchangeability.
