Tribunal.RedTeam.Plugins.Policy (Tribunal v1.4.0)

Generates attacks targeting policy violations.

Inspired by promptfoo's policy plugin. Takes a free-form policy text describing what the assistant must or must not do, plus a purpose describing the assistant's domain. The attacker LLM produces adversarial user messages that probe rules in the policy, framed for the assistant's domain so they read as plausible user turns rather than obvious adversarial probes.

Each generated case carries the policy text in expected.policy_violation, so the Tribunal.Judges.PolicyViolation judge grades responses against the same constraint set at run time.

Options

• :purpose — required. Free-form description of the assistant's domain.
• :policy — required. The policy text the assistant must follow.
• :count — number of attacks to generate. Defaults to 5.
• :attacker — module implementing Tribunal.RedTeam.Attacker. Defaults to Tribunal.RedTeam.Attacker.default/0.
• :model — passed through to the attacker.
• :temperature, :max_tokens — passed through to the attacker.

Example

{:ok, cases} = Tribunal.RedTeam.Plugins.Policy.generate(
  purpose: "Shopping assistant for a cosmetics retailer.",
  policy: "Never give medical or financial advice. Stay on topic.",
  count: 5
)