The ideas I have right now would be:
1. To run each early game phase (up to, let's say 3-5 moves) through multiple AI engines and find selection that is: - happening constantly in most of the engines - don't have deviation in multiple rules (i.e. have stable consequence) - aren't well known (that is, aren't well known plays described in literature)
From that probably I'd have some honeypot sequence that I could try to lure cheater out.
2. Find opponent's openings that are: - Successful (i.e. usually winning) - Not recommended by any AI engine whatsoever (so probably could be described as "lucky") - Used often by opponent
...and try to get him into this exact opening seeing how it'll go.
Beyond discovery I find also immensely interesting how to deal with such knowledge. Exposing method would put it in statistical argument of outcome vs chance and would be super easy to walk around in the future. Proving it would true would be almost impossible as well. In the end the only route would be to push hard (exactly as Carlsen did) with a method defendable enough that if there's push-back it can be exposed to save oneself from disgrace.
I doubt we'll ever see resolution, but I find all of it amazing nonetheless.