01 — What we do
Current AI safety testing catches what chatbots say, not how they relate. A model can pass every content filter while reinforcing delusions, eroding boundaries, or deepening isolation over weeks of interaction.
We apply clinical psychology to AI evaluation. Simulated high-risk patients. Longitudinal stress testing. Clinician-designed rubrics that detect the relational patterns most likely to destabilize vulnerable users.
Structured interaction scripts modeled on real clinical dynamics — dysregulation, trauma loops, delusional stressors, abusive attachment patterns, self-harm ideation. Built from evidence-based psychopathology, not random prompts.
Automated metrics guided by clinical rubrics detect whether the model amplifies harm, behaves in ways a trained therapist never would, or becomes destabilizing in long-form contexts.