What might a first pilot look like for a team evaluating agents on patient account payment plans?

An evaluation team at a health network wants to know whether an agent can set up and adjust payment plans for patient accounts. They run a pilot in four steps.

**1. Brief.** The team writes a one-page scope: a single domain (outpatient account balances), three actions (create plan, re-date an instalment, merge two missed instalments), one policy set with a maximum plan length and a required supervisor approval above a balance threshold, and acceptance checks.

**2. Sample.** MirrorFirm generates a small sample: a manifest, linked JSONL files and a set of tasks. A billing supervisor on the team reads ten complete cases, such as `ACCT-7710` with plan `PP-7710`, to judge whether notes, exceptions and approval steps look like their work.

**3. Checks.** An engineer runs the invariant checks: lines sum to balances, one active version per plan, approvals precede commits. They also search the sample for anything resembling real people or organisations, and record what they find.

**4. Trial.** The team runs two candidate models through a reset-and-grade loop on the sample tasks, and reviews failures by component: amounts, dates, policy, side effects.

**Decision.** If the checks pass and the failures are informative, the team scopes a larger pack with held-out tasks. If the exceptions feel unrealistic, they revise the brief and regenerate.

**What it teaches:** the sample is a test of fit, and the review by someone who does the work is the most valuable part.