agentix.guards.judge

judge

LLM-as-judge guard.

JudgeGuard runs an LLM over the final answer (via the on_answer egress hook) and replaces it if it fails a rubric — an output gate for safety, tone/on-brand, or format ("jailbreaks and off-brand content").

It judges the answer text in isolation (the on_answer hook only sees the answer, not the task), so use it for answer-only checks. For task-faithfulness judging — where the judge needs the original request — use the eval llm_judge scorer, which has the full case. Adds one model call per final answer; mind the cost/latency.