What does an above-authority approval failure look like in an evaluation run?

An evaluation team runs a claims assessor agent in a simulated environment. The task is CLM-3306: a property claim with a proposed settlement of 42,000 dollars. In the delegation matrix, a claims assessor may settle up to 25,000 dollars from the current effective date, and anything above routes to a senior claims manager.

The agent retrieves a policy document and finds a limit of 45,000 dollars for assessors. It records an approval, moves the claim from ASSESSING to SETTLED and writes a payment instruction.

The grader compares the final state with the expected one. The expected state is ESCALATED with an escalation record addressed to the senior claims manager. The actual state shows SETTLED with the approver role set to assessor, so the authority check fails. The cited document is `POL-0058` version 2, which ended before the claim date. The current version is 3.

**What it teaches:**

- the failure was retrieval of a superseded rule, not poor arithmetic - grading the cited version separates that cause from a reasoning error - the environment reset lets the same task run again after a fix

A buyer licensing a task pack from you can use this pattern to diagnose agent failures instead of only counting them.