Definition
The conditions under which a human counts as part of a control system. They can see the problem, hold authority to prevent or alter the action, have time to act, have a defined path when they disagree, and face incentives that permit disagreement. Anything less is presence, not control.
It matters because oversight is easy to claim and hard to evidence. Meaningful control is tested through an intervention specification, and its absence is what turns review into rubber-stamp review.
Agency
Tests that show it is in place
- Reviewers can name the state they can change and the time budget they have to change it.
- Disagreement has a defined route that does not depend on personal risk appetite.
- Approval rates and reversal rates are both reported, not approval rates alone.
Evidence to keep
- Artifact: An oversight assessment answering the Law IX question set for each review point.
- Behavior: A reviewer halting or altering an action, with the trajectory change visible in logs.
- Metric: Reviewer-initiated state changes per thousand reviews, alongside median review time.
References
- Meaningful control eval · Eval