xinxuxin / keystone-bench

Why you might care

Standard clinical chat benchmarks score replies against a fixed physician rubric and miss whether the assistant reacts when key evidence is removed, contradicted, or replaced. Keystone builds paired twin conversations, each changing exactly one fact, with annotated accepted and forbidden actions, plus negative controls. It suits researchers evaluating medical LLM safety and rubric validity.