Every AI call is routed by task and sensitivity tier through a declarative policy. Providers are classified by where they run: A (deterministic/CPU), B (self-hosted neural), C (external API).