How should a model lab scope a first pilot for schedule workflows?

A useful first pilot is narrow and designed to answer a few questions. Broad requests tend to hide problems.

**Choose one or two domains.** For example, consumer lending instalments and utility hardship plans, because their policies differ enough to test generalisation.

**State the use.** Training data, preference pairs, an evaluation set or a simulated environment each need different fields and different amounts of variation. Say which one comes first.

**Name the record types and formats.** List the tables, the link keys and whether you need JSONL, Parquet or both. Share your loader's expectations so the sample fits it.

**Define success before the sample arrives.** Write down the checks you will run, such as key integrity, sum reconciliation, policy-version correctness and a baseline model pass profile.

**Agree the limits.** The output is generated and informed by how businesses operate. It is not real business records, and it does not come with assurances about privacy or regulatory standing. Plan your own review of grounding and disclosure risk.

**Leave room to change.** Treat the first sample as a way to find what is missing. Quantities, timing and commercial terms are agreed directly with MirrorFirm and are not set by this description.