Decision models: AI judgment you can actually rely on
The checks inside the gate run on decision models — a different breed of AI, built for one job: answering policy questions with honest probabilities.
You give a decision model a state (“this email”) plus typed questions (“does it contain a forbidden promise?”) and it returns a calibrated probability — p=0.93 really means yes ~93% of the time. The answer is always one of your options. No free text, no hallucinated answers.
The output is a number you set a threshold on: “block if p ≥ 0.8”. Strictness becomes a documented business decision, tunable per rule — not a prompt tweak. Fast, cheap, and they run locally on modest hardware.
The company's policies become a decision tree — Vision → Rules → Policies → Departments → People, every node with rules, questions and thresholds. One call checks the whole stack, and the gate returns which rule, what probability, why — for the audit trail.
Every rule that fails carries its probability into the audit trail — \u201cproof, not promises\u201d: the deliverable regulators ask for, as a product feature.