A meta-judge is a higher-capability LLM that resolves disagreement between the primary judges in a multi-judge consensus pipeline. After three independent judges score a finding, the meta-judge sees their verdicts, confidences, and rationales, and produces a final verdict.
Meta-judges typically use a stronger reasoning model than the primary judges, because the meta-judge's job is to weigh competing rationales rather than to evaluate the raw finding. Low-confidence meta-judge outputs are routed to a human review queue.
This pattern is documented in the Penaxtra architecture page and underpins the cost-per-finding economics: meta-judge invocations are the most expensive but also the rarest call in the pipeline.