An ongoing process combining community reports, task artifacts, agent traces, automated checks, and human review to identify and repair benchmark issues.