The Five Faculties: A Tour of SAFi's Cognitive Architecture

Most attempts at AI governance treat alignment as a prompt-level concern. You write a system message, hope the model follows it, and accept that any sufficiently creative attacker can talk the model into ignoring it. The Self-Alignment Framework Interface (SAFi) takes a different approach. Instead of asking a single LLM to judge its own output, SAFi splits cognition across five specialized faculties, each with a distinct role, a defined interface, and no ability to overstep its bounds. The resul...