The term “AI governance” has become semantically overloaded, applied indiscriminately to logging, guardrails, model alignment, and policy workflows. This paper proposes a formal taxonomy that distinguishes three fundamentally different governance problems—operational visibility, behavioral alignment, and decision authorization—and maps prevalent governance approaches to the problems they actually solve. The paper introduces a precise definition of deterministic AI governance as a pre-execution authorization layer in which identical governed state produces identical governance verdicts and where the system emits verifiable decision artifacts sufficient for independent third-party replay. It provides a normative specification of governed state, a minimum evidence package, an enforcement triad (ALLOW / DENY / ABSTAIN), and practical evaluation frameworks—including anti-laundering tests designed to prevent trust-based or vendor-dependent imitation. The taxonomy is intentionally testable and disqualifying rather than aspirational. It is intended to provide enterprise buyers, regulators, researchers, and auditors with a shared vocabulary and concrete criteria for evaluating AI governance claims in high-stakes, regulated environments.