Build HealthBench and HealthBench Professional, public evaluations that measure model behavior for lay users vs. clinicians (jargon tailoring), hedging under genuine uncertainty, and safety-critical responses when users want reassurance but need to hear otherwise.