What should buyers inspect before using a generated routing corpus?

Apply this checklist to a sample of complete tasks, including their linked context and evaluator notes, before using a generated corpus for training or assessment.

- Confirm every required relationship resolves to a record. - Check that destinations exist in the relevant state. - Verify expected actions against the supplied policy text. - Distinguish missing evidence from deliberately withheld evidence. - Allow alternative actions where policy permits them. - Inspect exception cases for contradictory task expectations. - Confirm explanations cite evidence visible to the agent.

Review complete decision paths rather than isolated rows. A destination can be valid in the queue catalogue yet invalid for the request's current stage. Likewise, a clarification request can be reasonable even when the evaluator prefers another permitted action.

Buyers can ask for acceptance sets instead of a single answer where the process allows discretion. Those sets should explain why each action is acceptable and which constraints remain mandatory.

Record defects separately from agent failures. An unresolved foreign key, contradictory policy or unsupported expected answer should trigger corpus review. Treating these defects as model errors makes comparisons difficult to interpret and can reward agents for guessing around broken task context.