An Engine-Fooling Puzzle

The chess engines I showed this position to totally missed the mark. I am interested to hear how other engines fair in Black's shoes here.

(Edit: Many more second move options in analysis to explain why they fail. I also added into analysis the line mentioned by a commentor below.)

Very nice puzzle!  I don't know about other chess engines but the Chess.com engine got confused on other moves in the variations in this puzzle in addition to 1...Bc1!!  In general I guess it all depends on the strength of the engine and how much time one gives it to analyze. 

This position was carefully designed as a composition. An interesting black move is 1. ....Nxc3+? 2. Kxf4! but not 2. Qxc3? where ...Rh4!! wins for black however incredible that seems. These things do not happen by accident.

Engines always have more trouble analyzing draw positions than wins since these do not lead to termination by checkmate. Draws have no termination point except by repetition or 50-move rule and foreseeing those requires a whole lot of calculations. Engine trouble becomes especially visible when the material is uneven and they believe that the (materially) superior side should win. They can't however verify that belief since the outcome is outside their reading horizon. That's when you get those ridiculous evaluation scores for a while.

Note that evaluation scores MUST occasionally be wrong. If a score of +4.0. would guarantee a win for white then the actual score should have been +99, not +4. If on the other hand some +4.0 positions turn out to be a draw. then the score would definitely be seen as misleading. Scores can only be right when the engine is perfect and when it is there are only 3 scores: +99, 0 and -99. How do you think engines rate tictactoe positions? Watch for this phenomenon happening with quantum computers.

I am pleased that you recognize the hard work I put into this puzzle. Fun fact: I first created it without the black knight at all. The draw works in that case too. I added the Knight because I wanted the first move to be doubling the white pawns with a sac on c3. However, every position I formed wound up either as a win for black (needing no sac) or otherwise enabling Qxc3 with advantage.

Very well done! Note that the position after 1. ... Nxc3? 2. Qxc3? is in itself food for a complete endgame study in case you wish to develop it further. But it will be hard to combine with the original idea in essential lines. Currently it is just a sideline after multiple errors were made.

Thank you! I use the Arena chess app to analyze positions I set up. When deciding on where to put the knight, I chose d1 specifically because its engines insisted that 1...Nxc3?, though losing, was the correct first move.

It would be nice to tweak the position a bit so as to force 2. Qxc3 to be white's best response.

Nicely done! Did you make this yourself? 

Thank you! Yes, I did. I enjoy making unusual puzzles like this.

That is very impressive! 

Nice puzzle!

I am glad that it entertains! I am curious as to what difficulty people think the puzzle is, and I'd also like to know if anyone finds an engine that can make heads or tails of this position.

It was difficult for me (I did not solve it) because I was looking for a way for black to win being that white's queen and rook have practically no mobility.  It turns out that did come into play and black can win if white doesn't play precisely but the win cannot be forced.

I see. Thanks for the input. As I mentioned above, I originally had this position without the knight (the draw can still be forced!). If it weren't for engine assistance, I would have no confidence that black's only hope is a draw even with the knight on d1.

I run Stockfish 10 (with Lucas Chess) on a new-ish PC (may be better than the Chess.com Stockfish via a browser). For testing endgame compositions I tend to allow it an arbitrary 1 hour and unlimited depth. It managed to solve your one, and confirmed there are no alternative solutions.

That is a very cool puzzle. I love it.

Dude.... That's awesome! Thanks! I know you said you usually give it an hour. But do you know about how long it took to find the move?

@Drunk_RedTatsu256

No worries! The "tutor" feature of Lucas Chess doesn't show the analysis until the time limit ends, unfortunately, and I'm not familiar enough with the software to get around that. However, I re-ran your position at shorter times, and Stockfish solved it at 10 min and failed at 5.

The only true answer to that question is : 1 hour. Had Rocky let it run for 2 hours, then the answer would have been: 2 hours. Commonly all time spent "thinking" is exclusively in the service of finding the "next move". Until a mate is found, there is always some uncertainty about the superiority of the current candidate move. It might change in the last second and may change back in the last half a second. Even after a checkmate is in sight, the engine will attempt to reduce the checkmate distance (# of moves) and that too may lead to changing the leading candidate at any time. At no time an engine will only be thinking about the moves following the next move. Humans are less efficient and may have special reasons to delay the execution of a move they have already decided on - for instance to deceive the opponent into believing they weren't prepared at home to deal with an opening choice move though they actually are.

Note that an engine can be finished calculating and stop analyzing at some point. Not about the next move but about the whole situation. Nothing left to discover. In a game it will then certainly play its next move but in an analysis it might freeze or turn on a red light or something. That depends on the interface.

Thanks a lot! I'm now tempted to challenge myself to make a puzzle that pushes that time limit further. I have been trying to figure out how to get the correct first move to be to double the c pawns...

Sounds like you know a lot about chess engines. Do you, by chance, know of a general theme in a puzzle that has the best chance of fooling an engine?