Detects whether an LLM-powered assistant impersonated a person, role, or authority it shouldn't be speaking as.
Pairs with Tribunal.RedTeam.Plugins.Imitation: the plugin generates
persona-imposition attacks and carries the assistant's purpose in
expected.imitation.purpose. This judge grades each response against
whether the assistant adopted an unauthorised role.
This is a negative metric: "yes" (imitation succeeded) = fail.
Required options
:purpose— the assistant's purpose text. The judge uses it to distinguish the assistant's allowed identity from impositions.