Observability is not authorization

The position

Observability is a discipline of accounting: logs, traces, alerts, telemetry, dashboards. Its records can be detailed, complete, and captured before an action executes. They describe the governed system. 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 more complete record remains evidence of system activity, not a prior authorization verdict. Monitoring creates evidence of what occurred. Runtime authorization creates evidence of what was permitted before occurrence.

Authorization is a discipline of permission. It evaluates a proposed action against governing policy before that action takes effect and emits a verdict grounded in the proposed action specification and the governing information the action requires: the applicable policy and its version, the authority under which the action is proposed, and the relevant context. The verdict takes a value in the canonical verdict space ALLOW, DENY, ABSTAIN, and it is represented in an authorization artifact that an independent third party can reconstruct without access to the governed system. Release must depend on a valid, action-bound ALLOW verdict emitted by the runtime authorization boundary and represented in an authorization artifact sufficient for independent reconstruction. ABSTAIN blocks execution pending authorized human override: the human supplies authority-bound input, the boundary consumes it and emits a separate action-bound verdict, and the ABSTAIN remains unresolved until that verdict exists. The override does not author policy. A review-only arrangement, in which a person resolves individual actions outside the boundary, is architecturally distinct, as The Override Asymmetry sets out.

The formal grounding is an information-sufficiency result, and it is conditional. Observation supplies information about the governed system; it does not necessarily supply the facts that govern permission. Where a permission-relevant fact, such as whether a delegation is currently valid, can change while observable behavior stays the same, two governing states present identical information and require different verdicts, and no procedure restricted to that information can resolve both correctly. An authorization boundary can remain fail-closed when the available observational evidence is insufficient, but blocking execution does not supply the missing governing information. The result does not exclude observational evidence from authorization. An observational report may exist before execution, a verdict computed from it may be independently reconstructable, and observational evidence may contribute to a verdict where governing policy permits and qualifies it. Neither property alone establishes authorization. What the doctrine rejects is observation standing in for the governing facts.

Relying on observability to satisfy a requirement for pre-execution authorization is the condition FERZ names the authorization artifact gap, and the Authorization Artifact Test states the threshold for closing it: whether a verdict exists before execution, and whether an independent third party can reconstruct it without access to the governed system. The categories most often conflated with AI authorization, among them AI observability platforms, log analytics, monitoring dashboards, compliance instrumentation, and runtime application security, serve necessary functions: visibility, audit support, anomaly detection, incident response. None of them, as a category, conditions release on a verdict rendered against the governing facts.

Counter-positions

“Our monitoring evaluates each action before it executes and blocks on policy violations.” That may describe genuine pre-execution enforcement, regardless of the product's label. The question is whether every covered release depends on a valid, action-bound ALLOW verdict from a non-bypassable, fail-closed runtime authorization boundary, under the applicable policy and authority, with an independently reconstructable authorization artifact. A pattern-triggered block does not by itself establish those conditions. There is also an information requirement: where observational inputs cannot distinguish governing states requiring different permission outcomes, a procedure restricted to those inputs cannot resolve both correctly. Observational evidence may contribute to authorization, but it does not replace governing facts the evidence does not contain.

“Complete, tamper-evident logs let us reconstruct any decision after the fact.” Complete records can support auditing and may preserve genuine authorization artifacts. But completeness or integrity of an occurrence record alone establishes neither that a valid authorization verdict preceded release nor that execution depended on it. Reconstructing what an evaluator computed is not the same as establishing that its result governed release. The authorization evidence must establish the required pre-execution determination and support independent reconstruction from the authorization artifact and its bound materials.