Evidence-backed completion for AI agents
Evidence-backed completion means autonomous work is accepted only when the declared result matches the correct run, an allowed terminal state, the required evidence, and the expected artifact. It replaces an agent’s unsupported “done” message with a verifiable completion record while preserving a blocker and next action when proof is missing or contradictory.
Four checks make “done” durable.
Correct run
Evidence belongs to the intended invocation, project, owner, and stage.
Allowed terminal state
The workflow reached a terminal state permitted by the acceptance policy.
Required evidence
Tests, review decisions, run events, or other declared checks are present.
Expected artifact
The accepted output is linked and can be inspected independently.
Missing proof is a blocker, not a soft success.
When evidence is absent, stale, or belongs to another run, the record should not advance as complete. It should retain the current owner, the unmet condition, and the next action. Midfleet can enforce this rule only for completion transitions routed through its supported control path.
Turn acceptance into a control point.
Specify the state, evidence, artifact, and recovery requirement for one workflow.