The main thing I've developed so far is a way of estimating the complexity of each position. The intuition is that matching an engine's top moves in difficult positions is much more informative than doing so when there's an obvious best move.
The detector currently combines complexity with engine agreement and centipawn loss to make a prediction. I'd like to move beyond the heuristics and train an ML model on a decent dataset of known cheating/non-cheating games.
There is an obvious limitation though: I have fewer signals available than than Chess.com/Lichess. I generally don't have a player's game history, and most PGNs don't contain time-per-move data, both of which would presumably be very useful.
Still pretty experimental, so I'd be interested in feedback; especially from anyone who's worked on chess engines, cheating detection, or related statistical problems.