Observability is not authorization

Observability is a discipline of accounting. It produces records of what an AI system has done, when, and under what conditions: logs, traces, alerts, telemetry, dashboards. These records can be detailed, real-time, and cryptographically signed. They remain downstream of the generative process they describe. They are evidence, not permission.

Authorization is a discipline of permission. It evaluates a proposed action against governing policy before that action takes effect and emits a verdict reconstructable from policy, context, and proposed action specification. The verdict takes a value in the canonical verdict space {ALLOW, DENY, ABSTAIN}, with ABSTAIN blocking execution pending authorized human override. Override here operates inside the authorization boundary, codifying into the active policy and binding future verdicts. This is architecturally distinct from human-in-the-loop review, where human intervention resolves individual actions without policy persistence (see The Override Asymmetry, Meyman, 2026). The verdict exists before execution and is independently verifiable.

The two disciplines produce categorically different artifacts. Observability produces records of behavior. Authorization produces verdicts over a defined decision space. No accumulation of records of behavior constitutes a verdict over the decision space that produced the behavior. A more accurate observability system remains an observability system. A lower-latency monitoring layer remains downstream of execution. A more complete audit record remains evidence of system activity, not a prior authorization verdict.

This is a structural claim, not a question of vendor sophistication. The impossibility result establishes that no architecture deriving its verdicts from observation of an AI system's outputs, attempts, or self-evaluations can produce the artifact that an ex-ante authorization requirement demands. The conclusion is independent of latency, accuracy, automation, or scale.

This doctrine distinguishes FERZ from the categories most often conflated with AI authorization: AI observability platforms, log analytics, monitoring dashboards, compliance instrumentation, and runtime application security. These categories serve necessary functions: visibility, audit support, anomaly detection, and incident response. None of them produces a pre-execution verdict over a defined decision space. They report on what has happened. Authorization decides what is permitted to happen.