Summary
Provides an open-source method for simulating mental health chatbot conversations and scoring responses against a clinician-informed rubric. Its five dimensions assess detection and confirmation of potential risk, guidance toward human care, supportive communication, and appropriate AI boundaries. Version 1.2 revised the rubric with clinician and lived-experience input.
Healthcare Implications
Developers, providers, and purchasers can use the versioned rubric to investigate chatbot responses to sensitive mental health scenarios and compare evaluations run with the same rubric and personas. It supplies a more specific testing method than HAPI's existing general chatbot guidance. A benchmark score is not certification or proof of safety in clinical use.