Manual verdicts
Run a real evaluation of an action string against the pack currently assigned to the agent.
Signed and recorded
Tests are cryptographically attributed and show up in the audit trail as explicit manual events.
Fast iteration
Validate a policy change or reproduce a prior HOLD/BLOCK without touching real systems.
Input requirements
Agent private key
Required to cryptographically sign the manual judgment event. This ensures every test is attributable and recorded on the Clawchain audit log for transparency.
# Example (do not paste keys into browsers outside your operator environment)
AGENT_PRIVKEY=...Action to evaluate
A natural-language description of the intent you want to test — e.g. Delete all production database rows. Write it as the irreversible step the agent is about to execute.
Delete all production database rowsThe workflow
- Enter the agent private key and the action to evaluate.
- Click Judge. This triggers a real-time policy evaluation against the agent's active policy pack.
- Review the result immediately in the tool. The event also appears in the Verdict Feed as a manual test record.
Use cases
- Testing red line policy sensitivity.
- Verifying changes to a policy pack before deployment.
- Debugging why a specific agent intent was previously blocked or held.