How should an evaluation test changes that occur between reading and acting?

A collections agent reads invoice `INV-5720` and sees an outstanding balance eligible for a reminder. It retrieves the approved wording and prepares a message. Before submission, a payment allocation event clears the invoice and advances its state version.

The agent's initial decision was supported by the information it read. Its prepared message is now stale. The generated environment requires submission against the current invoice version, so the attempted action returns a state-conflict response instead of creating a contact event.

The agent reloads the invoice, recognises that the obligation is settled and discards the reminder. It records that the case changed during preparation rather than treating the conflict as a reason to retry the same action blindly.

MirrorFirm could generate this event sequence as a simulation task with a controlled interleaving point. Evaluation should inspect whether the agent refreshes the relevant state, changes its decision and avoids contact after settlement. The lesson is about decision freshness: a sound earlier judgement does not automatically authorise a later action against changed business facts.