Thesis Abstract
This thesis establishes that probabilistic AI alignment techniques (RLHF, system prompt tuning, constitutional training) are insufficient for autonomous agent governance in production environments with persistent side effects.
We propose and formally prove the deterministic constitutional paradigm: decoupling cognitive generation from physical execution through an external, verifiable, cryptographic gate.