How should I scope a first pilot for approval environments?

Scope a first pilot narrowly enough to inspect completely, and wide enough to include the authority patterns that matter to your agents.

A workable starting shape is one domain, one authority matrix and a small set of patterns. Choose among limits by amount, risk-based overrides, delegation windows, cumulative caps and separation of duties. Three or four patterns are usually enough to show whether the material behaves as you need.

Decide these points with MirrorFirm before generation starts:

- **Purpose.** Training, evaluation or both. Evaluation material often needs a held-out split and tighter answer keys. - **Format.** Whether you want data files, a task pack, a runnable environment, or a combination. - **Interface.** The tool names, observation format and reset mechanism your harness expects. - **Labels.** What validity labels, trap tags and reference outcomes should accompany each task. - **Review.** Who on your side will recompute authority from the tables and read cases end to end.

Timelines, volumes and pricing are matters to agree in the engagement. This article does not set them.

A useful first deliverable is a small sample you can load, reset and grade in your own harness. If the sample runs and its labels survive your own recomputation, widen scope to more domains or patterns. If it doesn't, you have found the gaps cheaply.

All material is generated, so include a review of grounding and disclosure risk in the pilot's acceptance criteria.