How should a model lab choose the first pilot scope?

Start by naming the behaviour the lab wants to improve or measure. A useful first scope might be selecting the correct evidence requirement for open applications after a policy amendment. That is more inspectable than asking for a complete organisation with every possible propagation path.

Define the input the agent receives, the decision it must make and the observable result that demonstrates success. Then identify the minimum surrounding context needed to make the decision credible. A pilot can include rich case histories without modelling unrelated functions such as staffing or procurement.

State the exclusions as part of the scope. If the environment cannot represent external customer responses, it should not require the agent to complete a process that depends on receiving one. The task can instead end at a correctly formed evidence request and an appropriate waiting state.

MirrorFirm could generate a sample around that boundary for joint inspection. The buyer can review whether the task tests the intended reasoning, whether the expected outcome is defensible and whether the surrounding records support it. Expand only after that review identifies a concrete need for additional roles, exceptions or actions.