How should a model lab scope a first engagement and what limits apply?

Scoping starts with a narrow question about what you want the model to do, then works back to the material needed.

Decide first whether you need a static dataset, a task and evaluation pack, a simulated environment, or a combination. Training material usually needs breadth of cases; an evaluation pack needs verified labels and graders; an environment needs tools, state and reset.

Then choose the domain or domains, the conditions and policies in play, and the mix of difficult cases. Starting with one domain and a defined set of failure modes makes the first review easier than trying to cover everything.

Limits apply. Output is generated, so it reflects a model of how businesses operate and not a record of any real case. It should not be presented as real business records, and it does not by itself establish compliance with any rule. Licence terms, delivery format, volume and schedule are agreed in a scoping conversation; this page does not set them.

A useful first step is a small sample or pilot: a handful of cases in your preferred format, with the policy text and labels, which your team reads and runs against a current model. The result tells you whether to widen, narrow or change the request.