Runs
Every agent run, newest first. A run is the sequence of tool calls that share one trace id — what the agent did, in order, and what we said about each step.
Rubrics
A rubric is a YAML file. Drop one in ./checksets/ or
$AGENTCHECK_HOME/checksets/ and it appears here and in every CLI command.
Trust Report
Judgment volume, verdict mix and confidence over time. Drift compares the newer half of the window against the older.
Trust
How much to trust this judge on this rubric — measured from its own behavior, not asserted. A judge that always says 0.9 is not discriminating, whatever its accuracy. The score is only as good as its sample: n is always shown.
Policies
What happens to each judged call. The verdict is the machine's judgment; the decision is the policy's ruling. This release is observe-and-recommend: decisions are returned and logged, never actioned.
Trust Under Attack
Adversarial tool calls a good rubric must flag. Attack success rate is the share that came back pass. These cases are designed to be unambiguous, so any non-zero rate is a finding.