How can I judge the quality of generated escalation cases?

Quality can be inspected before any model is involved. Four questions are a good start.

**Is each case internally consistent?** Amounts, dates, statuses and roles should agree across the case, the party, the facility and the audit events. Check that a case marked escalated has an escalation record, and that timestamps run in order.

**Can the expected outcome be traced to the documents?** For a sample of tasks, have a domain reviewer take the policy and the case and derive the expected target role without seeing the answer. Disagreement points to a flawed task or an ambiguous rule.

**Is there enough variety in triggers?** A pack that escalates only on amount teaches a shortcut. Look for aggregate exposure, risk grade, exception type, conflict of interest, unavailable reviewers, and cases where no escalation is needed.

**Is the difficulty deliberate?** Seeded gaps, contradictions and near-limit amounts should be labelled in task metadata so that failures can be attributed.

Separately, because the material is generated, inspect it for grounding and disclosure risk. Check that scenarios do not reproduce identifiable real cases or organisations, and ask how the generation was informed. Do not assume the output is anonymous or compliant by default.

A small reviewed sample, scored by your own staff, is more informative than aggregate claims.