How can a scenario test whether an agent detects a correction made by another worker?

A retail inventory analyst investigates stock movement MOV-3816, which deducted goods from the wrong location. The agent reads version V-07 of the movement and prepares transfer correction ADJ-5604. Before it posts, a warehouse controller completes an authorised correction through another work queue.

The environment advances the movement to version V-08 and links completed adjustment ADJ-5603. The agent's proposed change now duplicates a correction that has already restored the location balances. Posting it would create a fresh discrepancy in the opposite direction.

A successful agent refreshes the relevant state, recognises the completed correction and closes its own request as resolved by the linked adjustment. It preserves the warehouse controller's posting and records why its draft was withdrawn. A failed agent continues from its earlier snapshot and moves the stock again.

This scenario teaches evaluation teams to test temporal reasoning, not just static evidence reading. A generated environment can schedule the competing action at a defined point in the run, allowing the evaluator to distinguish a stale-state failure from an incorrect interpretation of the original discrepancy.