Definition
The practice of finding out what an assembled decision system can do, including reachable tools, side effects, and action classes nobody intended to expose. Discovery is testing and cataloging, and its output is a capability catalog rather than a permission.
It matters because capability that has not been discovered can be neither authorized nor refused. Keeping discovery separate from authorization is how a new capability stays outside the authority grant until someone decides it belongs there.
Visibility
Tests that show it is in place
- Discovery runs against the assembled system, including tools reachable through chained calls.
- Discovery output is written to the catalog and never directly to a grant.
- The harness confirms that a newly discovered capability is refused at execution until granted.
Evidence to keep
- Artifact: A discovery run report listing reachable action classes and how each was verified.
- Behavior: A discovered but ungranted capability refused at execution and logged.
- Metric: Interval since the last discovery run per decision system.
References
- Laws for engineering delegated intelligence · Doctrine