What should I ask for in a first sample of authority-dependent tasks?

Frame your sample request around what you need to inspect, so the review takes hours instead of weeks of back-and-forth.

- **One domain, stated up front:** for example refunds, claims payments or purchase orders. - **Schema documentation:** field definitions, keys and date semantics for each file. - **Named conditions:** boundary amounts, linked requests, expired or revoked grants and conflicts, each tagged. - **Gold outcomes with alternatives:** permitted decisions, valid approvers and short rationales. - **Policy text matched to limits:** written rules that agree with the structured authority records. - **A small number of deliberate ambiguities:** clearly labelled, so you can see how they are handled. - **Your preferred format:** JSONL, Parquet or both, with names aligned to your harness. - **Generation notes:** a plain description of how the material was produced and what it was informed by, so you can assess grounding and disclosure risk.

Also ask what the sample cannot tell you. A small sample shows fit and structure, not how a model performs across a wide range of conditions.

Write your success criteria before the sample arrives. Decide in advance what pass rate on your own checks would lead to a larger scope, and what would lead you to change the brief. Terms, quantities and timing are agreed in scoping.