Cool! It's nice to know that 50% solvability is possible.

One note: In your example with the 11 configurations, are you sure that just clicking the squares with lowest counts, is the optimal way to go? It might be that while they are cheap, they also give little useful information on which configuration we are actually in? Perhaps some dynamic programming type solution could give the "clicking order" ensuring maximal winning probability.

Yea, that's a problem for sure. I don't know what's the best way to weigh being able to make progress in the future vs surviving the next click. For example, is it better to take a 50/50 chance and be guaranteed to make progress, or take a 80% chance of surviving but likely be stuck next turn? It's an interesting problem to be investigated someday.

Thanks for your post! I remember reading this sometime in 2013 and writing my own minesweeper AI based on it. It didn't get very far, but I sure had a lot of fun.