Choose examples without sensitive information and write the expected answer before testing.

Record answer quality and identify cases where the system should defer to a human.