How should a model lab scope a first pilot for set-up material?

Keep the first step small and testable. Four decisions shape it.

**Domain.** Choose one, such as lending, utilities connection or professional services engagement. Mixing domains early makes quality harder to judge.

**Use.** State whether you want trajectories for training, tasks for held-out evaluation or an environment for interactive runs. Each needs different fields, and the answer decides what is built.

**Content.** List the record types, the policies and the exception types you want. Specify the share of clean cases against hard ones, and any failure modes you most want to probe.

**Review.** Agree beforehand how you will assess the sample: structural checks, policy walk-throughs by your own staff, a grounding and disclosure review and a trial run against a model you already evaluate.

Also settle practical limits. Ask what format and schema you will receive, how versions are identified, and what the licence permits for training, evaluation and redistribution. Timelines, volumes and terms are set in discussion with MirrorFirm, not assumed.

All output is generated, so treat the pilot as the point to find out whether it is grounded enough for your needs.