Where was Stockfish's fatal mistake in this game against Leela?

Question

This is a game from the TCEC Season 18 superfinal, which Leela won with a beautiful positional bind.

[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQkq - 0 1"]
[Event "TCEC Season 18 - Superfinal"]
[Site "https://tcec-chess.com"]
[Date "2020.06.27"]
[Round "65.1"]
[White "LCZero v0.25.1-svjio-t60-3972-mlh"]
[Black "Stockfish 202006170741"]
[Result "1-0"]
[Annotator "archive"]

1. d4 Nf6 2. c4 e6 3. Nf3 Bb4+ 4. Nbd2 O-O 5. a3 Be7 6. e4 d5 7. e5 Nfd7 8. Bd3 c5 9. h4 g6 10. O-O Nc6 11. Nb3 Bxh4 12. Bh6 Re8 13. Re1 cxd4 14. Qc2 dxc4 15. Bxc4 Nb6 16. Rad1 Bd7 17. Nc5 Rc8 18. b4 Nxc4 19. Qxc4 Be7 20. Ne4 Nxb4 21. Qxb4 Bxb4 22. axb4 f5 23. Nf6+ Kh8 24. Rxd4 Rc7 25. Red1 Re7 26. b5 b6 27. Kh2 Rb7 28. Ng5 Qc8 29. R1d2 Rc7 30. Rd6 Rb7 31. R2d4 Rc7 32. Rd1 Rb7 33. R6d4 Rc7 34. f4 Rb7 35. Nxe6 Rxe6 36. Nxd7 Kg8 37. Nf6+ Kf7 38. Rd8 Qc5 39. Nxh7 Re8 40. e6+ Rxe6 41. Ng5+ Kf6 42. Rf8+ Qxf8 43. Bxf8 Rc7 44. Rd4 Rb7 45. Kg3 Rc7 46. Rd3 Rb7 47. Kh4 Rc7 48. Kg3 Rc4 49. Rd7 Re3+ 50. Kf2 Rxf4+ 51. Kxe3 Ra4 52. Be7+ Ke5 53. Kf3

Link to the game from the TCEC archives if one wants to see the engine evals and PVs.

Where did Stockfish go wrong? Stockfish knew it was lost after 22. axb4, so the critical error must have happened before this. Leela didn't like 20...Nxb4, but the analysis by GM Matthew Sadler is that White has a pleasant advantage anyway after that line. Meanwhile newer versions of Stockfish with NNUE enabled don't seem to think 20...Nxb4 is a mistake either. In the position after 20. Ne4, the version of Stockfish on Lichess (Stockfish 13+ NNUE) at depth 33 thinks 20...Nxb4 is the best move with +0.1 eval; the next-best move is 20...a5 with +1.3 eval.

Playing through the game with Stockfish 13+ NNUE, it's clear that somewhere around move 20, it is missing some critical game-winning line, but it's not clear how to avoid that line and the losing move is. Can anyone tell?

Answer 1

The hardware allocated to you by Lichess for running SF is really under-powered when compared to a modern high-end desktop - let alone the TCEC hardware.

SF 14.1 NNUE on my hardware (32 logical cores) thinks 20...a5 (+1.47) is best for Black. Its next best try is 20...a6 (+2.31). This is at depth 52/73.

After analysing 20...Nxb4? 21 Qxb4! to a depth of 46/70, my SF evaluates the position at +3.91! So it seems that 20...Nxb4 is indeed a mistake.

As to "Where did [TCEC] Stockfish go wrong", SF on my hardware sees a gradual decline between 3...Bb4+ (+0.58 at depth 45) and 9...g6 (+1.03 at depth 47). Then 18...Nxc4 seems to have been a mistake (+1.56 at depth 52) - as noted by Sadler, 18...Be7 would have been better.

The real way to find the mistakes would be to play a set of engine matches from various positions in this line. This would give you the main positional and tactical themes.

Answer 2

UPDATE: After viewing graffe's comment, I analyzed the alternative option that Stockfish could play, 18...Be7 instead of 18...Nxc4. It turns out that 18...Be7 is also losing for black, meaning that Stockfish was lost after it played 17...Rc8. Black's best chance to draw would be to play 17...Be7. Here is the best resistance black can offer, but it's still losing after 17...Rc8:

[FEN "2rqr1k1/pp1b1p1p/1nn1p1pB/2N1P3/2Bp3b/P4N2/1PQ2PP1/3RR1K1 w - - 1 1"]
[Event "TCEC Season 18 - Superfinal"]
[Site "https://tcec-chess.com"]
[Date "2020.06.27"]
[Round "65.1"]
[White "Leela"]
[Black "Stockfish"]
[Result "1-0"]
[Annotator "archive"]

 1. b4 Be7 2. Ba2 Nd5 3. Qd2 Bxc5 4. bxc5 Nc3 5. Ng5 Ne7 6. Qf4 Nf5 7. Nxh7 Kxh7 8. Bg5 Qxg5 9. Qxg5 Nxd1 10. g4 Nh6 11. Rxd1 Rxc5 12. Rxd4 Bc6 13. Rd3 f6 14. Qd2 Rxe5 15. f4 Rc5 16. Rh3 Kg7 17. g5 Ng8 18. Bb1

Stockfish 15 is out and stronger than ever. It takes little to no time to analyze this position (in fact, only 500 million nodes) and identifies the critical error with 18...Nxc4, booming to +2 within little time.

Black's seemingly only hope to defend is with 20...a5, after which white responds with 21.b5 Nb4 22.Qxd4 Nc2. According to Stockfish 15, there is only one winning move here, and it isn't 23.Qxd7 (Stockfish only evaluates this as +0.9). The winning resource here is another brilliant idea, 23.Qd2! The line is: 23...Nxe1 24.Qxe1 Qc7 25.b6 Qc6 26.Qe3 Qa4 27.Rd4!

The idea here is stunning - it's a very deep sacrifice involving a king hunt that Stockfish finds in mere seconds. Here's the full line:

[FEN "2rqr1k1/pp1bbp1p/2n1p1pB/2N1P3/1PQp4/P4N2/5PP1/3RR1K1 w - - 0 1"]
[Event "TCEC Season 18 - Superfinal"]
[Site "https://tcec-chess.com"]
[Date "2020.06.27"]
[Round "65.1"]
[White "Leela"]
[Black "Stockfish"]
[Result "1-0"]
[Annotator "archive"]

1. Ne4 a5 2. b5 Nb4 3. Qxd4 Nc2 4. Qd2 Nxe1 5. Qxe1 Qc7 6. b6 Qc6 7. Qe3 Qa4 8. Rd4 Rc4 9. Rxc4 Qxc4 10. Qf4 Rc8 11. Nh2 Bf8 12. Ng4 Bxh6 13. Nxh6+ Kg7 14. Kh2 Qc1 15. Qxf7+ Kxh6 16. g4

Answer 3

There is no one particular mistake. It is white building the attack around black king 7.e5 ... 12 Bh6 and finally deadly 23 Nf6

They are playing somewhat imprecise in some segments around 20. so it is more of a "who made a worse choice" than about one particular error. Not that they are playing wrong, it is in the sense of position being complicated enough to exactly name all details in it.

However, the basic error that allows the attack build-up is 9. ... g6? (h6 is better). This move is just calling for an attack.

Before that, you can take as an imprecision: black castled too early and then played 6 ... d5 forcing the only figure that is defending the king side, a knight, out of the game.

The rest is just an execution. Notice as well how useless black rooks are after 24. Covering nothing, defending a poor locked bishop. It is a disaster in making.

Sometimes people think that chess is about one mistake you make. No. Many times it is a small connected chain of poor decisions that just created a momentum, a combination, and when that starts, like here, if you know what you are doing, there is really no defense.

Answer 4

It would be impossible for an average player like the ones on here to answer this. There are types of positions where engines are weak but even so I would rely on an engine to analyze it and who's to say one engine is better than another?

Engines are so strong it would take weeks of analysis to say one was wrong without some long term strategic error. I doubt the people on here giving SF lines are doing much more than repeating what SF says.