How can I inspect the quality of an escalation dataset before licensing it onward?

Start with checks you can run yourself on a sample.

**Consistency.** Every escalated case should have a requested amount or condition that exceeds the reviewing role's limit on the decision date. Cases that did not escalate should sit within it. Mismatches in either direction show weak generation.

**Referential integrity.** Each event should join to a case, each case to a reviewer, each decision to a rule version. Orphans are a warning sign.

**Threshold coverage.** Look at cases just below, at and just above each limit. A release with only obvious cases will not test boundary judgement.

**Text against fields.** Read handover notes beside the structured record. A rationale citing a figure that differs from the record is a defect.

**Variety.** Check that rationales are not near-copies of each other and that outcomes branch across approval, decline and return with conditions.

**Split hygiene.** Confirm all events for a case sit in one split.

**Grounding and disclosure.** The output is generated, so ask how it was produced and look for unusual detail that could point to a real source. Passing these checks does not remove disclosure risk, so keep your own review in the process.