Did Leela Zero miss a win in this position from game 106 of the Computer Chess Championship, on move 83 with White to move?

[FEN "8/2p1r3/2P4k/1P1pB2p/K2NbP1R/P5r1/8/8 w - - 0 1"]

Here, Leela Zero played Bxc7? for their 83rd move. I went deep to analyze the position at 34/91 with Houdini 6.02 for one hour, and the game reached a position with White having a rook, bishop, knight and pawn against Black's two rooks. I analyzed it again starting from there, and either White will increase his positional advantage or Black will be forced into simplifying into a tablebase win for White. Is all this true? Rh2 was the winning move according to Houdini 6.02.

  • Houdini 6.02 spent , too , 18 seconds on Bxc7? and more than a quarter of hour on Kb4?! , which is probably insufficient against accurate defense , and then settles for Rh2 , which seem winning. Thanks.
    – Stefano
    Jul 8, 2020 at 3:49
  • 5
    It would speed up this process if you could also provide the entire line Houdini gives after Rh2 so that we can see if there is an improvement amongst all the maneuvering. Jul 8, 2020 at 4:04
  • 2
    This badly needs more details. Who is Leela playing against? What were their PVs? Can you link the game?
    – Allure
    Jul 19, 2020 at 23:10
  • 2
    I will add that Houdini 6.02 is significantly (>100) elo weaker than Leela, and depth 34 is not especially high, so be careful of relying on what Houdini says.
    – Allure
    Jul 19, 2020 at 23:11
  • 2
    Can you show how the game continued after 83.Bxc7 ? I assume from your question than it ended in a draw ?
    – Evargalo
    Aug 5, 2020 at 8:49

3 Answers 3


For what that's worth, Stockfish 11 gives

  • Rh2 as +2.6
  • Kb4 as +2.2
  • Bxc7 as +0.6
  • other moves as 0.0

on my machine at depth 45 (will be updating this answer while it digs deeper). Rh2 and Kb4 occasionally switch positions as #1 and #2, so it's still unclear which one of these is truly best. Bxc7 surely isn't, however, at least if you believe Stockfish.

One important note: The top evaluation score has hovered between +2.0 and +2.6 since depth 15. Apparently, Stockfish has not found a way to substantially improve White's position (let alone win) in the last 30 plies. It's still possible that the game will end in a fortress situation with perfect play, and then Leela would in fact not have missed a win (despite having played a suboptimal move).


At depth 54, it's

  • Rh2 as +3.9
  • Kb4 as +2.2
  • Bxc7 as +0.1
  • other moves as 0.0

Still not fully conclusive, but we're starting to get somewhere.

In any case, Bxc7 was certainly not the most optimal move that Leela could have played.


Houdini is severely weaker than LC0. I have seen their evaulations of positions, such as in the Salvio Gambit, which Houdini evaluates as only slightly worse for White(clearly not true), where LC0 states that Black has a sizeable advantage.

Thus, it is more likely that Leela had played the best move. We cannot know for sure until we prove a forced win for White in N moves.


Dunno but BxP on c7 looks like a winner to me.

Whether black captures with the rook or not the b pawn marches quickly to become a Q

  • 1
    True, but remember that after the bishop sac, White is a rook behind - forcing Black to trade one extra rook for the new queen won't win by itself, e.g. 1. Bxc7 Rxc7 2. b6 Rc8 3. b7 Rb8 4. c7 R3g8 5. cxb8=Q Rxb8=.
    – Annatar
    Aug 5, 2020 at 10:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.