How would a grader score an agent that activated a facility on the wrong terms?

An evaluation run gives an agent task TASK-0731: book the facility for approval APP-4402, a AUD 280,000 equipment loan with a variable rate and monthly repayments.

Partway through the episode, a scheduled event amends the approval: the rate basis changes from fixed to variable. The agent had already read the approval, so it never sees the change.

The agent creates FAC-7015 with a fixed rate, links the security record, enables drawdown and generates the schedule. Every tool call returns success, and the agent reports the task complete.

The grader compares the final state with the expected record set and reports:

- `facilities.rate_basis`: expected variable, found fixed - `schedules`: instalment amounts differ from the expected set from row 1 - `exceptions`: none raised - `amendment_acknowledged`: false

The outcome is a fail with a stated cause. A grader that only checked for an active facility and a generated schedule would have scored a pass.

**What it teaches:** score the resulting records field by field, include the mid-task event in the task definition, and record whether the agent re-read the source. The same trace then supports error analysis and training signals.