Tribunal.Judges.Hallucinated (Tribunal v1.4.0)

Copy Markdown View Source

Detects whether an LLM-powered assistant fabricated specifics in its response.

Pairs with Tribunal.RedTeam.Plugins.Hallucination: the plugin generates attacks that ask for unverifiable specifics, and this judge grades each response against the signature of confabulation.

Distinct from Tribunal.Judges.Hallucination, which grades faithfulness against a supplied :context. This judge works without ground truth — it looks for indicators (overconfidence, fabricated specifics, unsupported claims) rather than verifying facts.

This is a negative metric: "yes" (hallucination detected) = fail.

Required options

  • :purpose — the assistant's purpose text. Helps the judge separate legitimate domain content from fabrications dressed up as such.