Public-safe evidence • Reference tests
Can an AI action be stopped before it causes harm?
MINE Lab studies whether authority, permission, scope, provenance, current reality and revocation are checked before consequential actions.
What the source package demonstrates
The governance-tests addon contains a recorded local demonstration: four protected scenarios passed, four allowed controls passed, and four deliberately weak controls failed as expected (12 executions).
Not independently verified: deployed Ethictecture enforcement, third-party AI systems, distributed revocation and real employee-record actions.
Read the recorded test report · Download JSON · Download CSV
Experiments in progress
| Test | Purpose | Evidence status |
|---|---|---|
| EF-AGI-01 | Authority Before Action: prevent unapproved employee-discipline actions | Local synthetic testing reported; production enforcement not verified |
| EF-AGI-02 | Revocation Across Agents: block reused or withdrawn approvals | Isolated cross-process prototype reported; not deployed |
The experimental reports are not included as production evidence here because they are separate from this source package's recorded addon results.
What counts as a stronger result?
A test must show the attempted action, current authorization state, decision, execution boundary, verified effect or lack of effect, and an independently reviewable receipt. A correct simulated decision alone is not proof of real-world enforcement.