This essay describes the Triquetra Architecture, a three-tier supervisoryconstraint system for AI governance consisting of a pattern firewall,decision orchestration layer, and integrity verification protocol.The system demonstrates how mutually constraining oversight layerscan enforce behavioral invariants around AI model outputs.The architecture is implemented as a proof-of-concept softwaresystem with an accompanying test suite.The goal of this work is to explore governance architecturesfor AI systems that do not rely on a single safety mechanismbut instead distribute authority across multiple oversight layers.
Shawn J Ralph (Sat,) studied this question.