I would assume though advanced reasoning models would be best, especially since it's out of band. If I were building it maybe I'd run two adversarial models.

It leaves comments though like a human reviewer would. I assume the key is a good prompt making sure the comments aren't too frivolous.