How can a buyer judge the quality of a generated intake dataset?

Inspecting an intake dataset is mostly about consistency, coverage and honesty about origin. Four lenses help.

**Internal consistency.** Do states follow legal transitions? A case should not be `routed` before it is `received`, and a `fail` on a mandatory check should not be followed by a standard-queue routing with no waiver. Simple scripts over the event stream can test this.

**Rule conformance.** Does each check and routing outcome match the versioned policy that the record cites? Where the data intentionally breaks a rule, such as a supervisor override, it should say so in a field.

**Coverage.** Are the awkward cases present: duplicates, partial submissions, reopens, escalations, conflicting details? A dataset made only of clean paths will look good and train poorly.

**Grounding and disclosure.** The output is generated and informed by direct observation of how businesses operate. Check that it does not reproduce identifiable records or distinctive free text, and ask how generation was controlled. Generated does not mean risk-free, so review it as you would any dataset you will license onward.

For training use, add split hygiene: keep all events of a case in one split.

For evaluation use, ask whether the reference outcomes are defensible when a reasonable operator disagrees. Good packs record the reason, not just the label.