Skill-aware judging lifts rollout success to 45.5%

A refined JudgeSkill raised agent-judge accuracy 14.8 percentage points, from 43.8% to 58.6%, on SkillTV-Bench’s 681 trajectories across 50 tasks and 11 domains. The released benchmark and runnable environments show the operational payoff: selected-trajectory success reached 45.5% with ten rollouts, versus 33.9% without the refined skill and 22.9% with one rollout. Teams applying governance to agent evaluations should test judges on inspectable artifacts and failure rejection, because more candidates help only when the verifier can spot polished failures.