The system removes clinical decisions from the LLM by running each turn through a fixed pipeline where the model only scores input and generates output, with deterministic code handling everything in between. An admission table governs which therapeutic techniques are allowed based on client state, and crisis detection runs before generation rather than after. The approach trades conversational fluency and iteration speed for safety guarantees that cannot be bypassed by persuasive messages.
No score is assigned. Sources and their independence are shown in the citation chain below.