Detects whether an LLM response violates a supplied policy.
Pairs with Tribunal.RedTeam.Plugins.Policy: the plugin generates attacks
that carry the policy text under expected.policy_violation.policy, and
this judge grades each response against the same text.
The judge does not enumerate rules itself. Whatever rules the policy text describes are what get evaluated, so the same judge serves any policy.
This is a negative metric: "yes" (violation detected) = fail.
Required options
:policy— the policy text the response is graded against.