Metrics

Research indicators, not universal scores.

Each measure is proposed so that a property of Human–AI work can be inspected. None is presented as a certified benchmark.

Whether successive revisions approach a stable, inspectable outcome.

How much of the path from goal to decision can be reconstructed.

The extent to which the issued decision remains anchored in stated human goals.

How often, and at which stages, humans interrupt or redirect a loop.

How far a loop proceeds without human intervention.

Whether a loop, or a slice of it, can be replayed under documented conditions.

Whether human interventions actually change unjustified paths.

Reliance of later AI steps on earlier unverified AI outputs.

Whether stopping follows a stated condition rather than drift.

Independence and consequence of checks before acceptance.