What I would like to do is double model post-check, with a form of "debate", to better catch edge cases.

Unfortunately, I haven't found a way to set that up as I envisioned it, for the time being.