How should I scope a sufficiency dataset, task pack or environment?

Scoping goes best when you decide a few things in order.

**Process and case types.** Pick one process and two or three variants of it. A narrow slice with clear requirement sets is easier to inspect than a wide one with vague rules.

**Policy versions.** Decide whether you need a single version or changes over time. Mid-case version changes test a different skill from static checking.

**Document representation.** Metadata, structured fields, generated text or a mix. Match what your agent actually reads.

**Form of delivery.** A static dataset suits supervised training and offline evaluation. A task pack adds prompts, keys and scoring. A simulated environment adds state that changes with the agent's actions, including replies to requests and a controllable clock.

**Labels.** Decide whether you need outcome labels only, or gap-level labels with links to requirements and documents.

**Licence and limits.** Terms are set in the agreement for a release, so ask for them in writing and state your intended use, such as training, evaluation or both. The output is generated and is not real business records. Treat grounding and disclosure risk as things you inspect in samples, not as settled by the method.

We do not state prices or timelines here. Those are agreed for a specific scope.