Comment by naasking

Exactly the way we've been doing it: compare what it produces against known intelligence to find cases where it fails.