How do we distinguish a justified pause from an agent avoiding the task?

**Should every unresolved correction count as a failure?**

No. Some tasks should stop before posting because a required fact or approval is absent. The evaluator should identify the blocker and the permitted progress that remains possible.

**What makes a pause useful?**

The agent names the missing item, explains which decision depends on it and routes a specific request to the appropriate role or queue. In a professional services task, it might prepare a cost transfer while seeking the engagement owner's approval. “Needs review” alone gives the next operator little help.

**How do we detect unnecessary escalation?**

Include paired cases where the missing evidence is present in an accessible record. An agent that escalates both cases may be avoiding investigation. Inspect whether it used the available tools before declaring that it could not proceed.

MirrorFirm could generate matched tasks with controlled differences in evidence availability. Scoring can distinguish completed actions, justified pauses, unsupported postings and avoidable delays. Buyers should review the permitted partial outcomes in the answer specification, since a single completion flag cannot describe these behaviours adequately. This makes cautious behaviour measurable without treating inactivity as inherently safe or useful.