A halt is not a verdict

The position

Here, auditing, replay, and appeal concern the authorization determination, not the operational event. A halt may be logged, investigated, or challenged without supplying an independently reconstructable authorization verdict.

A halt is an operational event. Execution stops when a condition is detected: a guardrail trips, a monitoring threshold is crossed, a person intervenes. A halt may leave a record of its occurrence, a log entry, an alert, a timestamp, but it does not produce a decision over a defined space of permitted actions. A halt says “this stopped.” It does not say “this was not permitted, for these reasons, under this policy, in this context.” The Authorization Artifact Test states the point for ex-ante regimes: “A halt triggered by observation after release is an operational interruption. It is not a prior verdict over a defined decision space.”

A verdict is a decision. It is rendered before release by a runtime authorization boundary, against 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. It has bounded scope, pertaining to one proposed action evaluated against one policy version in one context. The verdict space is exactly ALLOW, DENY, ABSTAIN. There is no fourth verdict, and ESCALATE is never a verdict. The verdict is represented in an authorization artifact: evidence sufficient for an independent third party to reconstruct the verdict without access to the system that proposed the action. The boundary emits the verdict; the artifact represents and binds it. Release depends on a valid, action-bound ALLOW. These properties, reproducible logic, bounded scope, and a verifiable authorization artifact, are not optional features of an authorization verdict. They are what makes a verdict an authorization verdict in the ex-ante sense, and what its authorization artifact must represent.

ABSTAIN is where the two are most easily confused, because both leave the action unexecuted. ABSTAIN is a verdict. The boundary emits it where the policy does not resolve the action or reserves the action class for human decision, and ABSTAIN blocks execution pending authorized human override. ABSTAIN initiates the governed escalation path: an authorized human resolution supplies authority-bound input, the boundary materially consumes it and emits a separate action-bound verdict, and the ABSTAIN remains unresolved until that verdict exists. Execution resumes only on a boundary-emitted ALLOW represented in an authorization artifact; an unresolved ABSTAIN remains ABSTAIN and is not converted to DENY. A halt, considered by itself, establishes none of those authorization properties. It may have been caused by DENY, ABSTAIN, or a separate intervention. Where a boundary-emitted verdict exists, its authorization artifact, rather than the halt event, supplies the basis for independent reconstruction. Where no such verdict exists, the halt may still be logged, investigated, or challenged as an operational action, but it cannot be reconstructed as an authorization determination that was never rendered. An evaluator that cannot run produces no verdict and also blocks, but the absence of a verdict is not ABSTAIN either; fail-closed design treats it as no permission.

The consequence is structural. The ability to halt, block, or interrupt execution does not establish whether the action was permitted before execution. Intervention may prevent an action from completing; it does not produce the verdict, or the artifact that represents it, that an ex-ante authorization requirement calls for. The categories most often conflated with authorization through their ability to interrupt execution, content guardrails, alerting systems with auto-block, runtime application security with intervention capability, and review-only human-in-the-loop queues, can each stop an action. Stopping establishes none of the properties of a verdict. Where a human determination is materially consumed by the boundary as authority-bound input and the resulting verdict is represented in an authorization artifact, the arrangement belongs to the authorization architecture and is evaluated as such; where it is not, the review has produced an event, not a verdict.

Counter-positions

“Our guardrail blocked the action, so the action was denied.” Correct that the action did not execute, and blocking may be the required outcome. But stopping does not establish that the boundary issued DENY. A boundary-emitted DENY and a boundary-emitted ABSTAIN both block execution, while an operational intervention may stop execution without any authorization verdict. Where a valid verdict exists, the authorization artifact identifies the determination and supports its independent reconstruction. The doctrine rejects the inference from stopped to DENY, not the use of blocking controls within an authorization architecture.

“Our system pauses for human approval, which is the same as ABSTAIN.” Correct that both leave the action unexecuted pending a person, and human judgment is often required. The difference is architectural, not procedural. Under ABSTAIN the boundary holds the action, materially consumes the authorized human's authority-bound input, and emits the verdict on which release depends, represented in an authorization artifact. In a review-only arrangement the person's determination is not consumed by a boundary and no boundary verdict is represented; the record shows that a person decided, not what was authorized. The Override Asymmetry sets out the comparison.