How does an end-state check catch a wrong policy start date?

An underwriter approves quote `QTE-6203` for a small business package policy, effective on the date the premium payment clears, not the date of approval. The policy administration team receives the work as `SETUP-1187`.

The agent creates policy `POL-33018` and sets the effective date to the approval date. Every tool call succeeds. The agent's summary reads cleanly: policy created, documents queued, customer notified.

The evaluation compares the final state with the expected state:

- `effective_date` differs from the expected value, which depends on payment record `PAY-8840`. - The certificate of currency was issued before payment cleared, an ordering violation. - The customer notice references a start date that is now wrong.

A score based only on whether a policy exists would pass this run. A field-level diff fails it, and the ordering check shows why: the agent never read the payment record before setting the date.

**What it teaches**

- Success of a tool call is not evidence of a correct outcome. - Dependencies between records should appear in the expected state. - Side effects such as certificates and notices should be scored, because they are hard to reverse. - Keep process checks and outcome checks separate so that you can tell a lucky pass from a sound one.