What checks should I run on a pack before using it for evaluation?

Run these checks on a sample before you trust it as an evaluation set; each one catches a different way a pack can mislead you.

- Every record and policy ID referenced in cases and labels resolves to an actual record or policy. - The stated variance equals the difference you recompute from the two records. - A person can reach the labelled cause using only the tools and data the agent is given. - Held-out labels do not appear in any field, note or tool response the agent can read. - Each forbidden action is actually possible in the environment, so avoiding it is a real test. - Matched, fault-free cases are present, so false alarms are measured. - Escalation-correct cases are present, with a defined passing outcome. - Environment reset returns identical state on repeated runs of the same case.

Record which checks fail and why. A failure is more often a data fault to report back than an agent fault to fix, and sorting the two early saves wasted runs.