What is a useful first pilot for a model lab?

Start by choosing a business distinction that matters to the intended model behaviour. In retail operations, that could be the boundary between an order amendment that fulfilment can still perform and a return request that belongs to after-sales support once dispatch has occurred.

Define the model's job narrowly: choose the next responsible team from the supplied order state and policy. Specify whether the task includes retrieving that state, asking questions or executing an assignment. Those choices materially change the sample and the evaluation burden.

A buyer can ask MirrorFirm for a generated sample covering the boundary, including cases where dispatch status is clear and cases where the available evidence is insufficient. Explicit exclusions might include refund calculation, warehouse optimisation and customer identity checks.

The first review should ask whether the sample captures the intended decision and contains enough evidence to answer it. Agree on corrections to the task definition before requesting broader coverage. This keeps the pilot focused on a learnable operational distinction rather than an undefined promise to represent an entire service organisation.