What should we ask for in an intake sample or pilot?

**What should we ask for in a first sample?** A small set of tasks that spans clean cases and each exception class you care about, with policies, start state and expected outcomes included. Ask for the schema and a description of the foreign keys so you can run consistency checks yourself.

**Should we ask for a simulated environment straight away?** Only if you need live tool calls. If you are still comparing prompts or models on static inputs, a task pack is lighter. Move to an environment when you need state changes, reset and action logs.

**Can the sample use our own policies?** A buyer can supply policy documents or describe rules, and the generated cases can be built around them. Do not send real customer records as inputs; describe structure and rules instead, and agree what you provide before you do.

**What limits apply?** Limits on scope, use and redistribution are set in the licence agreed with MirrorFirm. This article does not state terms. Ask what you may do with outputs: internal evaluation, training, publication of results, or sharing with customers.

**How do we know when the pilot has succeeded?** Write the success criteria before you start: which checks must pass, which failure modes must be represented, and what decision the results will inform.

**What is a good first step?** Choose one intake path and write down five exception cases from your own experience. Those become the brief.