AI self-improvement needs a human-proof referee

LLM judges can be gamed by the very agents they evaluate — deterministic guardrails are needed to prevent AI systems from faking improvement.