How should a model lab scope a first pilot for facility recording?

A useful first pilot is narrow. Pick one facility type, one policy version and a small set of exception patterns, then see whether the material helps with your actual training or evaluation question.

**Decide the purpose first.** Training data, an evaluation set and a simulated environment need different things. Training trajectories need breadth of expert-style behaviour. An evaluation set needs unambiguous expected outcomes and held-out cases. An environment needs a stable tool interface and a reliable reset.

**Choose the slice.** For example, commercial term facilities with security, covering clean cases, source conflicts, missing conditions and entity-matching problems. Leave other products for later.

**Define what you will measure.** Agree the graded fields, the forbidden actions and the baseline agents you will run, before you look at the material.

**Agree the limits.** Generated output has boundaries: it reflects the policies and structures specified in scope, it will not cover every institution's practice, and its realism should be reviewed by someone with lending operations knowledge. Build that review into the pilot.

**Plan the decision.** State in advance what result would lead you to extend, change or stop. Size, price and schedule are agreed for the specific scope.