How can I inspect the quality of generated escalation cases?

Start with checks a script can run, then move to checks that need a domain reader.

**Mechanical checks** - Referential integrity: every `related_facility_id`, `evidence_id` and `next_role_id` resolves to a record. - Arithmetic: recompute aggregate exposure and compare it with the authority limit; the expected routing should follow from the numbers. - Timeline order: events are time-ordered, a case is never actioned by a role that does not yet hold it, and status transitions are permitted ones. - Tier paths: every referral follows the matrix, and no case is referred to its own owner.

**Reader checks** - Do rationale texts and referral packages read like what a reviewer would write, or are they uniform and padded? - Are the policy clauses internally coherent, with precedence rules where documents overlap? - Is there a spread of amounts, tiers and outcomes, including cases that should not be escalated?

**Risk checks** - Look for repeated phrases or near-identical cases that would let a model memorise answers. - Inspect free-text fields for anything that resembles identifiable material. The output is generated, but grounding and disclosure risk still need review in a sample.

Record the result of each check against a sample version, so a later release can be compared with it. A dataset that passes mechanical checks but fails the reader checks is consistent yet unconvincing, and that is worth knowing before any scoring begins.