How should an evaluation team inspect the quality of generated intake material?

Inspect generated intake material in three layers, and treat each as a separate pass.

**1. Record consistency.** Do links resolve? Does every document belong to a real case, and every case to a party? Are dates in order, with receipt before assessment and amendments after lodgement? Do declared values in the form match the attached documents, except where a contradiction was deliberately planted and labelled?

**2. Policy and key agreement.** Pick a sample of tasks and work out the answer yourself from the policy text alone. Compare your result with the key. Disagreements point to ambiguous rules, wrong keys or missing policy versions. Also check that each key cites a rule identifier that exists and applies on the case's date.

**3. Realism and disclosure.** Ask how the cases and policies were grounded, and have someone who knows the domain read a handful. Look for implausible field combinations, unnatural wording and over-tidy cases. Also check generated names, identifiers and addresses for resemblance to real people or organisations. Generation does not guarantee that nothing resembles a real person, so this check is yours to run on the sample.

A release that passes the first layer but fails the second will produce confident, wrong scores. Weight your effort towards the second layer.