Governance Concepts
The FERZ Governance Concepts section provides the canonical vocabulary that defines runtime authorization architecture for AI systems in regulated environments. Each concept is a primitive: a term with a single, precise definition, an origin in the FERZ corpus, and a specific architectural function within the framework.
The concepts cluster into a few groups. Some describe temporal properties (ex-ante authorization, execution-time authorization). Some describe structural properties (runtime authorization boundary, non-bypassable AI governance). Some describe failure-mode and verifiability properties (fail-closed design, deterministic authorization). One describes a verdict (ABSTAIN). One describes the artifact every verdict produces (Proof-Carrying Decisions). One describes the doctrinal scaffold under which all of the above sit (the Three Problems Taxonomy). One is not a claim within the framework at all, but a method for evaluating any architecture, including this one (the Governing Question).
These definitions are canonical. They are the terms FERZ uses to position runtime authorization as its own architectural class, distinct from observability, identity and access management, policy engines, and security guardrails. Conflation with those adjacent categories is the dominant failure mode in current AI governance discourse, and the precision in this vocabulary is the corrective.
The vocabulary
The FERZ doctrinal framing that distinguishes Visibility, Alignment, and Authorization as three categorically distinct problems.
Authority for a governed action is established before the action takes effect.
The architectural implementation of ex-ante authorization at runtime, per action.
The structural location where authorization is enforced, independent of the governed system.
Every governed action must pass through a gate that cannot be modified or disabled.
Under uncertainty or error, the system blocks the action rather than permitting it.
The same inputs and policy state produce the same verdict, every time.
The fail-closed verdict that blocks execution pending authorized human override.
Cryptographically verifiable artifacts emitted for every authorization verdict.
The primary question an architecture was built to answer, identified before evaluating how well it answers it.