[al_biglan]

This seems a perfect candidate to test with some well known test games.

Fischer's 17 ... Be6, for example in the Game of the Century

Or 15 ...Nf2 (again Fischer and Byrne) http://www.chessgames.com/perl/chessgame?gid=1008419&kpage=1

It might be interesting to see how many blunders this identifies compared with traditional analysis. (Looking at the "surprising bad move" vs "surprising good move")

In any case, very interesting work. Others have some good ideas for additional checks/considerations, and it would be interesting to see how this evolves!

[zzazzdsa]

A good game to test this on is http://www.chessgames.com/perl/chessgame?gid=1139126 .

Tal played the move 19. Rf6!!, which instantly wins the game. However, (at least when I tested this a few months ago), Stockfish takes a fairly long time to recognize this-- it prefers the more conservative 19. c4 instead.

[Rainymood]

I just went through the game and man 19. Rf6!! Took me by surprise! What an amazing move!

I also would love to see some chess engine that plays ridiculous moves in the beginning to throw people off guard (i.e. increase greatly increase the variance in possible evaluations) and then crushes the opponent by playing all the computer moves while the human players have to walk on their toes in a minefield of possible blunders because the position is so wacky.