Since the whole thing is behind an API- exposing the works adds little value. If the corrections worked at an acceptable rate, one would just want them applied at the source.
What do you mean? The model is for improving their RLHF trainers performance. RLHF does get applied "at the source" so to speak. It's a modification on the model behind the API.
Perhaps if you were to say what you think this thing is for and then share why you think it's not "at the source".
You'd like to get the "correct" answer straight away, not watch a discussion between two bots.