How can we inspect the quality of generated mismatch cases before using them?

Quality can be inspected on a sample before any larger commitment. Five checks cover most of the risk.

**Solvability.** Give a reviewer the case and records, without the answer key, and ask them to find the cause. If an experienced reviewer cannot, the case is underspecified.

**Internal consistency.** Totals, dates and links should agree across files. A payment that settles before it was created, or a link to a missing record, signals a generation fault.

**Leakage.** Search case titles, notes and field names for words that give away the cause. Check whether a model with no tool access can guess the answer from the text alone.

**Process realism.** Ask someone who does this work whether the approvals, handovers and reason codes resemble their own. Differences are not always faults, but they should be understood.

**Disclosure and grounding.** The output is generated, but a buyer should check whether any identifiers, names or patterns look like real entities, and ask how the generation was grounded.

Record the results per case type, not as a single score. A release can be strong on duplicates and weak on cut-off timing, and you want to know that before you build a benchmark on it.