Question no one seems to be asking:

Do we need (the equivalent of) unit tests for AI validation?