How can an evaluation team inspect the quality of generated booking tasks before using them?

Inspect a sample in layers, starting with the cheapest checks.

**Structural consistency.** Do foreign keys resolve? Does every schedule reconcile to its facility terms, including residual lines? Do status histories follow allowed transitions?

**Policy alignment.** Apply the generated policies to each task by script and confirm the labelled disposition. A file marked `book` should pass every rule; a file marked `hold` should fail at least one, and the stated blocker should be the real one.

**Label correctness.** Have a domain reviewer, such as someone from credit operations, read a small set of tasks cold and write down what they would do. Compare to the labels and discuss disagreements.

**Difficulty spread.** Check that tasks are not all clean bookings or all traps. Tags for difficulty let you report results by slice.

**Leak checks.** Make sure the expected snapshot, labels and reasons are not visible in the agent's inputs.

**Grounding and disclosure review.** The data is generated, but ask how it was grounded and what overlap checks were run. Review a sample for anything that looks like it could identify a real organisation or person.

None of this proves the data is a perfect mirror of any one business. It shows whether it is coherent enough to grade against.