J-Zero Frozen judges limit LLM self-evolution; co-evolving the judge enables stronger, sustained improvement across verifiable and unverifiable domains.
