What does a scoped pilot look like from the first call to a decision to widen it?

An evaluation lead at a model lab wants to test whether an agent holds work correctly when an approval has open conditions. They ask MirrorFirm for a pilot in one domain: utility network connections.

In the scoping call, the lead and a MirrorFirm contact settle five things. The domain is commercial connection approvals. The records are approvals, conditions, evidence and work orders, delivered as JSONL. The failure modes are outdated evidence, a condition added after approval, and a request to energise before clearance. The environment needs three tools and a reset. The lab's own team will review the sample.

The first sample contains a small set of cases such as CONN-3318, where a network planner approves a connection subject to a revised site plan and a protection settings sign-off. Each case has policy text, a timeline, a label and a grader.

The lab's reviewers read the cases and flag one policy clause that is ambiguous. They run a current model and find that it releases the work order in cases where the sign-off is missing. They decide to widen the pilot to a second domain, and ask for an added group of cases where approvals lapse.

**What it teaches:** a pilot is a bounded loop of scope, sample, review and adjust. The aim is a decision about whether and how to proceed, with licence terms, volume and delivery agreed separately.