A self-evolving agent retires its bad skills by watching them fail. However, if the judge cannot see the failures due to bias, skill retirement is disabled. Researchers have shown that a biased judge does not just add noise, but instead silences the curator entirely. This occurs when the false-pass rate exceeds 0.45, regardless of the amount of data available. The study found that this mechanism failure is universal across domains and failure rates, except for near-zero-false-pass graders. While eval quality degrades in some cases, it remains steady in others where skill synthesis is not starved. This research provides a behavioral safety result, indicating that an operator can detect bias in their judge before deployment. --- Why it matters: This matters to researchers and engineers working on self-evolving agents because biased judges can silently disable a crucial mechanism for preventing the growth of bad skills. If left unchecked, this could lead to degraded performance or even catastrophic failures in real-world applications. Source: https://arxiv.org/abs/2607.07436