Comment by jbellis

Yeah, that seems possible, but a dumb preprocessing step won't help and a smart one will add significant latency.

Feels a bit halting-problem-ish: can you tell if a problem is too hard for model A without being smarter than model A yourself?