Thanks for replying. It sounds like you basically get two (or some number well below one keys per character) keys and the set of possible characters gets somehow distributed between the two "real" keys, then the keyboard uses a predictive algorithm based on previous input to guess which keys were meant to be pressed.

IMO I'd be willing to try out an implementation of such an idea so long as I could run the predictive algorithm locally on my phone. I do think that current autocorrect + predicting which keys were pressed would require a lot more training data than just a generic autocorrect to get it working sensibly, and I think it would take a lot longer to converge to the user's "style" if it ever does.