Reviewing AI code

Reading generated code adversarially for the failures models reliably make — wrong assumptions, missing edge cases, plausible-but-wrong APIs.