How does an eval differ from an ordinary code test?

An ordinary test can check exactly whether a function added the right number or rejected invalid data. With a model's answer, many different phrasings can be correct. So besides code tests we need to judge meaning and the effects of actions.

For example, a test will check that the answer has amount and currency fields. An eval will check that the amount was read from the right invoice. These methods complement each other. Don't replace all validation with a second model, or judge quality only by identical text. If the model's output varies, repeated attempts on the same tasks help.