Learn index
Completion integrity

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.

When to require proof—and when not to

When to use

  • Completion releases code, money, data, access, or downstream work.
  • A reviewer must reproduce why the result was accepted.

When not to use

  • Exploration has no accepted final state.
  • A human is directly supervising a disposable draft.

Four checks make “done” durable.

01identity

Correct run

Evidence belongs to the intended invocation, project, owner, and stage.

02state

Allowed terminal state

The workflow reached a terminal state permitted by the acceptance policy.

03proof

Required evidence

Tests, review decisions, run events, or other declared checks are present.

04output

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.

Talk to us