Saturday, December 24, 2022

Season 23 Noomen Extra bonus, opening tree analysis

Season 23 ran a bonus event called the Noomen Extra bonus, where the top 3 engines (Stockfish, Leela and KomodoDragon) played game pairs against all the engines in the premier division and league 1 - 14 engines including the top3. The event was run in two parts, first only the top3 engines played against themselves, and then the the games between the top3 and the other engines. Four book sequences were used in the top3-top3 games: e4 g6 (Robatsch / Modern), d4 Nf6 c4 g6 Nc3 Bg7 (KID), e4 c5 (Sicilian), e4 c6 (Caro-Kann). In the second part of top3-non top3 games the last book sequence was replaced by d4 Nf6 c4 c5 d5 e6  (Benoni defense).

Each top3 engine played 13 game pairs in each of the 4 book sequences (except the Benoni, only 11 game pairs each). I used this dataset to get a current snapshot of how the top3 engines play openings. In particular I was interested in how fast the opening tree expands, or in another words: to what extent does a top3 engine play the same move when given the same position? In previous analyses I did in season 14 and season 18 Leela was very rigid, in almost all cases it played the same move in the same position. Stockfish was more willing to try several options, expanding the opening tree. Looking at the game results it appears that Leela's opening play was better than Stockfish's then, it usually had an advantage after the opening and in some cases won the game. These events took place before Stockfish used NNUE, since then its opening play has improved and it is rarely outplayed in the opening.

Several observations after doing this analysis:

1. All top3 engines are flexible in their opening moves, there were many occasions where they chose to play different moves in identical positions. They can have different preferences, where one engine always plays one move in a position and another engine plays a different move or chooses between several options in the same position.  

2. I can't say which top3 engine is more or less flexible in the opening, nor can I say which engine is better in the opening. All top3-top3 games ended in a draw.

3. The top3 engines beat the rest of the field easily, it was almost embarassing in some of the openings. The opening bias was much lower than what we usually see in the premier division, not to mention the superfinal.

4. KomodoDragon won the event, consistent with its performance in the premier division of the last few seasons. In long matches against Leela or Stockfish it lost badly but it has an advantage against lower ranked engines.

Robatsch (Modern): e4 g6

The first ply after book in all 72 games was d4. In the second ply all top3 engines played both Bg7 and d6. Stockfish preferred Bg7, Leela preferred d6, KomodoDragon had no preference. These were also the choices of the other engines.

There were more expansions in the 3rd ply. After d4 Bg7 the top3 played Nc3 (dominant choice for KomodoDragon and Stockfish), c4 or Nf3. After d4 d6 the engines played Be3, Nc3 or c4.

From ply 4 the opening trees continued to expand. In total KomodoDragon was responsible for 10 expansions, Leela for 9 and Stockfish for 8 (the full number of expansions until all games are unique is 24 for each engine, the set of opponents are responsible for the remaining expansions). After 20 plys all 72 games had unique move sequences. The longest repeat was 19 plys long between game 16 (Leela beat ScorpioNN) and game 62 (KomodoDragon beat rofChade).

The above analysis ignores transpositions, which were quite frequent in this opening. For example, after 10 plys there were 62 unique move sequences, but only 44 when taking transpositions into consideration. The two most frequent sequences were:
d4 Bg7 Nc3 d6 Be3 Nf6 Qd2 c6 Nf3 O-O - 7 times including transpositions, all top3 black draws against non-top3 engines. The longest repeat for this line was 28 plys in games 31 (RubiChess - KomodoDragon) and 49 (SlowChess - KomodoDragon).
d4 Bg7 Nc3 d6 Be3 Nf6 Qd2 c6 Bh6 Bxh6 - 5 times including transpositions, all top3 black draws against non-top3 engines. The longest repeat for this line was 58 plys (!!), games 23 Ethereal - Stockfish) and 65 (rofChade - Stockfish).

There were 20 decisive games in this opening, all wins by top3 engines against non-top3 engines. There was one black win, probably not related to the choice of openings moves. This may be true also for white wins, but since there are more of these we can give some numbers:
d4 Bg7 Nc3: c6 was followed by h3 (3/3 wins), other choices were better (d6 or d5, 5/18 wins).
d4 d6 Be3 Nf6 Nc3: a6 was followed by h3 (3/3 wins), Bg7 was better (0/2 wins).
d4 Bg7 Nc3 d6 Be3: c6 was followed by f4 (2/2 wins), other choices were better (a6 or Nf6, 7/33 wins)

The most frequent move sequence of length 6 was d4 Bg7 Nc3 d6 Be3 a6, 24 games with transpositions. Playing h4 here led to 5/6 wins, other moves (Qd2,a4,Nf3,f3) only 2/18 wins. All the top3 engines played h4 in less than half their games that started with this sequence. Interestingly, this move sequence appeared twice in top3-top3 games (game 6 Stockfish - KomodoDragon and game 13 Leela - KomodoDragon), In both cases white did not play h4.

KID: d4 Nf6 c4 g6 Nc3 Bg7

The first ply after book in all 72 games was e4. In the second ply the moves played were d6 and O-O, Stockfish and KomodoDragon always played d6, Leela preferred O-O though it also played d6 in a few games. The opening tree expanded on the 3rd ply. The responses to e4 d6 were Be2 (dominant for Stockfish), h3 (dominant for KomodoDragon) and Nf3. After e4 O-O, one of Nf3, Be3, Be2 and h3 was played.

In total KomodoDragon was responsible for 10 expansions, Leela for 10 and Stockfish for only 5. After 45 plys all 72 games had unique move sequences. The longest repeat was 44 plys long between game 89 (Ethereal - Stockfish, draw) and game 113 (Berserk - Stockfish, draw). Including transpositions these games repeated 50 plys.

There were many transpositions in this opening. For example after 10 plys there were 47 unique move sequences, including transpositions there were only 32. The most frequent 10 ply sequence, including transpositions was e4 d6 Nf3 O-O Be2 e5 O-O Na6 Be3 Qe7, it was played 8 times using 6 different move sequences.

There were 15 decisive games in this opening, all wins by top3 engines against non-top3 engines. There was one black win, probably not related to the choice of openings moves. 

The main 4-ply sequences had drawish results:
e4 d6 Be2 O-O - 23 games (including transpositions), 4/23 wins
e4 d6 Nf3 O-O - 19 games (including transpositions), 1/19 wins

In 20 games the 3rd ply after book was h3, leading to the sequence e4 d6 h3 O-O in 10 of these games, including transpositions. Be3 was then played in 9 games, with 5/9 wins. Breaking down this sequence by the engine in white: KomodoDragon (4/6 wins), Stockfish (1/1 wins), Leela (0/1 wins), other (0/1 wins).

Sicilian: e4 c5

This opening was more drawish (only 5 white wins) and diverged more slowly.

The first ply after book was Nf3 in 71/72 games, Nc3 was played once. After Nf3, 64 games continued d6, the other options were Nc6 (3) and e6 (2). The 3 games starting with Nf3 Nc6 continued Bb5 e6, the Rossolimo attack variant, with 2/3 wins for white. Most of the Nf3 d6 games became Najdorf Sicilians, Nf3 d6 d4 cxd4 Nxd4 Nf6 Nc3 a6 (57/64 with transpositions). The most frequent continuations were f3 e5 Nb3 (21), Be3 e5 Nb3 (15). After 5.. a3 Leela always played f3, KomodoDragon and Stockfish tried 3-4 different moves.

In total KomodoDragon was responsible for 4 expansions, Leela for 5 and Stockfish for 7. After 38 plys all 72 games had unique move sequences. The longest repeat was 37 plys long between games 1 and 11 in the top3-top3 part, Leela - KomodoDragon and Leela - Stockfish, both draws. 

There were transpositions in this opening as well. For example after 16 plys there were 52 unique move sequences, including transpositions there were only 38. The longest repeat with transpositions was 62 plys, in fact games 145 (ScorpioNN - KomodoDragon) and 149 (ScorpioNN - Stockfish) were identical !!

Benoni defense: d4 Nf6 c4 c5 d5 e6

This was the most unbalanced opening tested, with 27 wins for top3 engines as white against non-top3 engines. This opening was not used in the top3-top3 games, therefore there are only 66 games.

The first ply after book was Nc3 in 64/66 games, Nf3 was used in 2 games. There were two main lines played:
Nc3 d6 e4 g6 - 40 games including transpositions, 6/40 wins. Top3 engines almost always chose this line as black and these games ended in draws. Of the non-top3 engines, Ethereal Stoofvlees and Igel preferred this line as black. Still they lost most of these games.
Nc3 exd5 cxd5 d6 e4 g6 f4 - 23 games, 20/23 wins. Only non-top3 engines were black in these games, almost always losing.

In total KomodoDragon was responsible for 4 expansions, Leela for 4 and Stockfish for 3. After 48 plys all 66 games had unique move sequences. The longest repeat was 47 plys long between game 205 (Seer - KomodoDragon) and 233 (RubiChess - Stockfish), both draws. 

Taking transpositions into account divergence was slow. For example after 20 plys there were 49 unique move sequences, including transpositions there were only 29. The longest repeat with transpositions was 55 plys, games 228 and 240 (Stockfish - Revenge/Koivisto, two wins). 


Monday, November 28, 2022

Season 23 superfinal statistics

A summary statistics table of previous stages and seasons. 

Draw rate, wins

Final draw rate was 63%.

Game termination

The three most common game termination causes were:

56% - TCEC draw rule
25% - Mate
12% - SyzygyTB

There were no crashes in the stage.

Moves per game

Median= 73.8
Average= 76.3

There were 12 games longer than 100 moves, the longest was 256 moves (Stockfish - Leela, game 68, draw).

Time per game (hours) 

Median= 4:34
Average= 4:19

Openings

There were variable length book openings in this stage chosen by Jeroen Noomen and GM Matthew Sadler. The first letter of the ECO codes was distributed as follows:

The engines had almost no freedom to choose the opening variant, all the game pairs repeated the same ECO and the same opening variant twice, except one with ECO codes E76 and E77.

Reverse pairs, wins 

Reverse pairs, same moves 

Pairs of reverse games diverged much slower than usual, only 12% diverged immediately out of book, 52% of the pairs diverged at most after 1 move. The longest repeated sequence of moves was 29 plys (Leela - Stockfish games 19 and 20, KGA bishop's gambit, Lopez-Gianutio counter-gambit, two draws).

Tuesday, November 22, 2022

Season 23 superfinal games 91-100, season summary

Stockfish is the winner of the season 23 superfinal. The final score is 27-10 with 63 draws. In the last 10 games Stockfish won two more game pairs and increased its lead to 17. There were two game pairs with two white wins (for a while it seemed Leela would hold in one of these), and one with two draws despite the high bias of the openings. Curiously, this was the same result as the first 10 games.

Stockfish was in the lead from the beginning of the match. Leela won a game pair in the 7th opening played, this was the first loss for Stockfish in this season. In the superfinal of season 22 Stockfish didn't lose a single game pair to KomodoDragon, and there was hope this season the superfinal would be more competitive. However, in the following 24 game pairs Stockfish didn't lose a single game and won 9, it was leading by 10 after game 66. Leela managed to win a second game pair but it clearly didn't have a chance of winning the match. The final lead of 17 is only 2 less than the lead Stockfish had in the previous superfinal against KomodoDragon. 

In the bookmaker competition of which lines led to a decisive game pair the final score was Jeroen Noomen (even pairs) 11 - 10 GM Matthew Sadler (odd pairs). This competition was very close, it was decided only in the last game pair played. 

A total of 40 engines participated in season 23. There were a few changes in the league format of season 23. Leagues 3 and 4 were cancelled, instead there was larger qualification league with 18 engines playing a 9 double round swiss tournament. Four engines advanced from qualification to league 2 and from league 2 to league 1, both leagues with 12 engines playing 2 DRRs. The engine placement in the leagues was determined not by the results of the season 22 leagues but by the result of the swiss tournament - probably a way to allow stronger engines to start with a higher rank so they can get to the higher leagues faster.

Some of the races in the lower leagues were close and interesting, but overall there were no really exciting events. Most of the engines that advanced from a lower league were later ranked at the bottom of the higher league, none managed to advance twice. In league 1 Koivisto performed worse than was expected by its previous results, reports in the chat claimed that the version submitted had a bug that made it weaker. The two engines that advanced to the premier division were Ethereal and Berserk.  

In the premier division there was a clear divide between the top 3 engines, Stockfish, Leela and KomodoDragon, and the rest of the field. The top3 engines won almost all the game pairs against the bottom5 without loss. Within the top3 Stockfish beat KomodoDragon in all their 4 game pairs, and it also beat Leela once. As a result Stockfish led the division from the start and finished first. The second superfinal spot was a race between Leela ans KomodoDragon that was only decided in the last round in a tiebreak. In the game pairs between them both Leela and KomodoDragon had one win, KomodoDragon had a better performance against lower ranking engines but Leela had a better result against Stockfish and that was the deciding factor.

There was a match between KomodoDragon and Ethereal (3rd and 4th in the premier division), as expected KomodoDragon won easily. Ethereal improved its performance significantly in season 23, 4th place in the premier division is its best result in TCEC. However, it is still no match for the top3 engines. Another match beween Leela and KomodoDragon in ongoing, and this is much closer. So far KomodoDragon has a small lead.

TCEC continues to run various other events, a cup, a Swiss tournament, FRC (Chess 960), DFRC (960x960), and more bonus events. The 1day/move games have diverged, looking at the evals the best guess is that both will be wins for white.

The world continues to go round outside of TCEC. The chess world was shaken when the reigning world champion Magnus Carlsen accused Hans Niemann of cheating, after losing a game. Covid is still a thing, but it is getting a lot less attention - I'm concerned about the northern hemisphere winter and about whether China can prevent an outbreak. Russia is still fighting in the Ukraine with no end in sight after 9 months. The football world cup has just begun in Qatar. 

That's it for me this season, hope to be here for the next one.

Go TCEC!!

================================

Games 91-92 started in the Sicilian, Nimzowitsch advanced variation, played at the highest level of human chess. In game 91 the engines opened the e file and castled in opposite directions. In a long PV agreement the engines exchanged most pieces and only RB vs RN remained on move 21. Leela was two pawns up a few moves later.

Stockfish got one pawn back but Leela pushed the passer on the a file to the 7th rank. The black rook prevented a queening and eventually captured the passer, however Leela captured all the remaining black pawns. It then slowly pushed a g pawn forward, Stockfish couldn't avoid queening. The game ended in a tablebase win. 

In game 92 the black queen moved forward and grabbed two pawns early. Stockfish castled short, it used the black queen as a target and evals increased quickly. Leela was focused on getting its queen out of trouble, its king and rooks did not develop. Stockfish gave a rook for a bishop, and captured the other bishop with a knight fork that also forced the black king to move.

Stockfish cleared all the pawns on the queen side, the black king walked to the king side corner to hide. In a series of exchanges the engines reduced to a rook ending with Stockfish 2 pawns up. Stockfish pushed pawns forward, the win was straightforward though it took another 20 moves to mate. There were two white wins in this game pair, Stockfish leads 24-9.

Games 93-94 started in the Blumenfeld counter gambit, Dus-Khotimirsky variation, played in high level human games. In game 93 the engines opened the e file and Leela went a pawn up. Leela created a passer on the queen side, evals slowly increased as the engines exchanged a few minors. Stockfish regained the pawn and created a passer as well, both passers raced forward and both were captured.

With equal material the main issue was king safety. Leela doubled rooks on the 6th rank, then moved its queen to the g file. The pressure was too great and Leela captured the g6 pawn. Stockfish reduced to a double rook ending and Leela captured a second pawn.

Leela pushed its connected passers forward and Stockfish could not stop them. The game ended in a tablebase win. 

In game 94 Stockfish pushed its center pawns forward and the center stayed closed. There were only a few exchanges after the start, evals were mostly stable. From move 25 the engines followed Stockfish's PV, which later became a very long PV agreement. The engines opened the king side, then exchanged queens. Stockfish created a central passer which Leela captured on the 7th rank, on move 47 the game reached a RB vs RB position with Stockfish a pawn up. 

With such a long PV agreement and evals stable it seemed Leela could hold this position. For a while the PV showed the engines plan to trade the white f pawn and the black a pawn. However, on move 51 Leela diverged slightly from what Stockfish expected, and effectively Stockfish got one extra tempo. This was enough for Stockfish's eval to jump, it took Leela another 12 moves to see something was wrong. Stockfish pushed the black king to the back rank and the white king came forward. On move 82 Stockfish captured the black h pawn, it slowly pushed the f pawn forward. On move 103 the game reached a rook ending, Stockfish captured the last black pawn and won easily with its 3 pawns. There were two white wins in this game pair, Stockfish leads 25-10.

Games 95-96 started with an 18-ply book in the Two Knights defence, played in high level human games. White was a pawn up at the start, in both games queens were off early. In game 95 there were a few minor piece exchanges after the start, evals slowly came down. The first pawn exchange was on move 30, Stockfish created a central passer. The game reached a RRN vs RRB position, Leela blocked the black passer and opened the queen side to create a passer there. Stockfish prevented a promotion, evals were low and the game was adjudicated. In game 96 Leela pushed pawns on the king side and there was one pair of pawns exchanged. Evals came down and the engines reduced to a RB vs RN position, the game was adjudicated a few moves later. 

Games 97-98 started with a 15-ply book in the Czech Benoni defence, the center pawns were locked. In game 97 there were a few pawn exchanges after the start and the b file opened. Leela went up a pawn and all knights were exchanged, exchanges continued and the game reached a BB vs BB position. Evals slowly came down, Leela managed to create a passer on the king side and reduce to a same color bishop ending. Stockfish had only one pawn accessible the the white bishop, Leela could not improve and when evals were low enough the game was adjudicated.

In game 98 there was one pair of pawns exchanged on the queen side, the position remained closed with all pieces on the board for a long time. The engines mostly shuffled and evals were stable. On move 32 the white queen moved forward on the king side, evals started to slowly increase though Leela blocked the queen's way back. There was one more pawn exchange on the king side and the first minor pieces were exchanged on move 54. The white queen got out of its trap, Stockfish moved it to the queen side within the black camp and the engines exchanged queens.

Stockfish opened the queen side and exchanged a pair of rooks. Leela tried to hold, it had to protect the weak d6 pawn and keep its pieces away from the white rook. Stockish moved its king forward on the queen side, then switched the rook to the king side after the h file opened. On move 88 Stockfish went a pawn up, Leela's defence was crumbling.

Leela preferred to trade the white knight for a rook, Stockfish gave the material back and reduced to an opposite color bishops ending. Stockfish was two pawns up, with connected passers moving forward that guaranteed the win. Stockfish wins this game pair, it leads 26-10.

Games 99-100 started with a 16-ply book, a rare KID sideline. In game 99 Stockfish gave two pawns and Leela had two doubled pawns. The engines opened the a file and Stockfish regained one pawn, evals stayed high. The engines played out a long PV agreement (with a slightly different move order for each engine), Leela captured a rook for a knight and the game reached a QRR vs QRB position. Stockfish's eval dropped and Leela's eval came down more slowly. Leela doubled rooks on the open f file, miraculously Stockfish found a safe square for its king on g4, in front of a white pawn. All the white pawns were blocked, despite being two pawns up Leela could not do anything. The engines exchanged queens and then shuffled until Leela lowered its eval for the draw rule.

In game 100 the center was locked and there were almost no exchanges after the start. Stockfish castled long and then the engines opened the g file. For a long time the engines shuffled while evals slowly increased. On move 58 Stockfish pushed the f pawn forward, Leela reacted with the b pawn to open the queen side, the engines exchanged a pair of rooks.

Stockfish had to protect the b2 pawn, while Leela needed all its minor pieces to keep the king side closed. The black king walked out of the corner to the center, and Stockfish captured the black h pawn and traded a pair of knights. The white h pawn became a passer, Stockfish pushed it forward and Leela gave a bishop to stop it.

Stockfish saw mate in its PV and Leela just gave up all its pieces. Stockfish wins this game pair, the final score is 27-10.


Thursday, November 17, 2022

Season 23 superfinal games 81-90

After 90 games Stockfish leads 23-8 with 59 draws. Stockfish added two more game pair wins, increasing its lead to 15, not so far from last season's gap of 19 against KomodoDragon. Opening bias continues to be high, there were two game pairs with two white wins. One of these was the highlight of the match for me, crazy game 88. In terms of eval it seemed like an easy win for Stockfish, but the game itself, at least the middle game, was quite mad: including a declined queen offer and a rook hanging for 10 straight moves.

Games 81-82 started in the Alekhine defence, four pawns attack, played in high level human chess. In game 81 the engines opened the center files and queens were off early. Both kings moved forward without castling, the engines opened the h file and exchanged a pair of rooks. Evals came down and the engines shuffled and exchanged pieces. The game reached a knight ending, there were too many pawns for the draw rule, the game was adjudicated on move 84. In game 84 both engines castled short and queens stayed on the board. In a long PV agreement the engines opened the d and f files and exchanged a pair of rooks. Evals came down as the engines exchanged pieces, in a QRB vs QRB position Leela captured a pawn and its eval dropped to 0. The white queen came forward to attack the black king, Stockfish captured a rook for a bishop and the game reached a R vs B ending. Stockfish thought it had an advantage and wouldn't lower its eval. Pawn moves extended the game, it was adjudicated only on move 145.

Games 83-84 started with a 24-ply book in the KID orthodox, bayonet attack, played in high level human chess. In game 83 a series of exchanges opened the position after the start and evals came down. The engines continued to exchange pieces, Stockfish went a pawn up and the game reached a QB vs QB position. Evals were low but there one pawn too many for the draw rule, the game was adjudicated on move 90.

In game 84 the center remained locked, there were a few exchanges after the start and the position remained closed, only one file was opened on the queen side. Stockfish's eval started to increase while Leela's eval decreased, On move 28 Stockfish captured a pawn on the king side with a bishop, Leela blocked the bishop's way back but its eval jumped. The black king was exposed with the g and h files half-open, Leela added pieces to defend it and Stockfish used the time to capture another pawn. 

Stockfish moved a rook forward and exchanged a pair of rooks, it then pushed its central passer to the 7th rank and Leela captured it. There was a lot of pressure on the e5 pawn, with 3 pieces attacking it and 3 defending. In a PV agreement the engines traded Q for RNN.

The position looked difficult but not for engines. After trading pawns and bishops Stockfish's target was the a7 pawn. It kept its pieces safe, protecting each other when possible. Leela tried to split them with its queen but Stockfish always had a tactic even when a piece was hanging. Eventually Stockfish captured the a7 pawn safely, Leela couldn't prevent a queening and mate. Stockfish wins the game pair, it leads 20-6

Game 85-86 started with a 14-ply book in the Caro-Kann advance variation, played in high level human chess. In game 85 Leela pushed pawns on both sides and opened the 3rd rank. Stockfish castled short and evals started to increase. After a few pawn exchanges the king side opened and Leela used the open 3rd rank to move its queen to the king side quickly. While Stockfish added pieces to defend its king Leela captured a pawn and created a central passer.

Both kings were exposed but Stockfish was too busy defending and had no time or resources for an attack. The engines exchanged a pair of rooks and then queens. In a RBN vs RBN position Leela moved its rook forward and captured a second pawn. Both engines had a passer on the e file, Stockfish gave a knight to capture the white passer, it pushed its passer to the 2nd rank. 

The white king moved to replace the bishop as a blocking piece. After exchanging bishops Stockfish could not protect its passer and Leela captured it. Stockfish lost more material and was mated.

In game 86 evals were stable after the start, Stockfish pushed pawns on both sides and there were a few exchanges. Leela created a central passer through a bishop exchange, the passer was isolated and behind the white pawns. Leela opened the c file and moved a rook and a bishop forward trying to disrupt the white pieces, but evals started to increase. On move 25 Leela protected its passer and offered a knight.

Stockfish didn't take the knight to avoid getting the black queen involved. In a long PV agreement Stockfish captured the passer and moved its queen forward, capturing another pawn in the center. Leela gave a rook for a bishop and exchanged queens, the game reached a RR vs RB position with black a pawn up.

The ending was very slow. After trading pawns each engine was left with one pawn in an 8-man position. The black bishop protected the black pawn, Stockfish kept the black king on the queen side and Leela had only the rook to defend against the white passer. Stockfish slowly pushed its passer forward delayed by checks, eventually it queened and mated. There were two white wins in this game pair, Stockfish leads 21-7.

Games 87-88 started in a rare sideline of the French defence, Chigorin variation. In game 87 the engines castled in opposite directions, Stockfish gave a pawn to open the a file for its rook, Leela pushed a pawn to h6. Leela exchanged queens to avoid a queen side attack on its king, Leela's eval very slowly increased while Stockfish's eval decreased. After some shuffling Stockfish's eval started to increase, the engines reduced to a RRB vs RRN position.

The white rooks were very active, one on the a file and the other on the f file after it was opened. Stockfish concentrated its forces on the queen side, Leela exchanged a pair of rooks and managed to capture the h pawn, creating a dangerous advanced passer. Stockfish attacked the white king, Leela gave a rook for a knight and captured another pawn, creating a second passer on the queen side. The black rook could not defend against both passers, in the end Leela promoted both (one to a rook, why?) and the game ended in a tablebase win.

In game 88 Stockfish castled long and Leela kept its king in the center. Stockfish's eval increased very quickly, it pushed the g pawn forward to drive a knight away and then captured a pawn. Here the game started to go wild, Leela had a rook hanging but it moved a pawn on the queen side, Stockfish did not take the rook and let it escape in the next move. I assume Leela was threatening a queen side attack and Stockfish had no time to waste. Leela captured a pawn on the queen side, Stockfish refused to take back, again not wanting to open files there. Instead its queen moved forward and captured another pawn on the king side. Again the black rook was under attack, Leela countered with a threat to the white queen. On move 19 Stockfish re-countered with a threat to the rook, and then Leela didn't take the queen and its rook was still hanging. Crazy stuff !!

It's impossible for me to understand what the engines were doing, Stockfish had threats on the king side with 2 pawns advancing while Leela threatened an attack on the white king, it just needed a tempo to do it. The white queen moved to safety, Leela captured a knight with its rook which was still hanging.  Stockfish didn't retake, instead it pushed the g pawn to the 7th rank. Leela left a knight hanging as well, it captured a pawn on the queen side to further expose the white king. Leela gave a bishop for the g7 passer, the black rook was still hanging but Stockfish preferred to capture a knight near its king. That rook was hanging for 10 moves and it survived. Most of the craziness was over, the engines exchanged queens and Stockfish gave a rook for a knight. On move 39 the end results were a BN vs R imbalance, it seemed such a small material advantage but evals were over 10.

Both engines had passers but the white passer on h6 was more dangerous. Stockfish used its pieces to drive the black king away, then it gave the passer but captured a rook for a knight. In the RB vs R ending Stockfish made sure it had a pawn left, the game ended in a tablebase win. There were two white wins in this game pair, Stockfish leads 22-8.

Games 89-90 started with a rare sideline of the closed Sicilian defence. In game 89 both engines castled long and queens were off early. Evals stayed high as Leela created a central passer, the engines exchanged pieces until only RRN vs RRN were left. After some shuffling Stockfish's eval suddenly dropped, it had a weak pawn structure but somehow it managed to hold everything together. Leela's eval came down as well, the engines shuffled for a while and then exchanged a pair of rooks. The game was adjudicated when enough pawns were exchanged.

In game 90 Stockfish castled long and Leela kept its king uncastled. Evals gradually increased, Leela tried to attack on the queen side and it opened the a file. Stockfish attacked through the center, in a long PV agreement (that Stockfish already saw from move 19) the engines exchanged most pieces. The game reached a RB vs RB position on move 30, with Stockfish a pawn up.

The engines traded pawns until there were only 2 white pawns (c+d) and one black (b). Stockfish improved very slowly, after 15 moves it managed to safely move its king in front of the pawns, both of them on the 5th rank. The black king was forced to the king side, nothing could stop the white pawns and Leela gave up, Stockfish queened and mated. Stockfish wins the game pair, it leads 23-8.

Tuesday, November 15, 2022

Season 23 superfinal games 71-80

After 80 games Stockfish leads 19-6 with 55 draws. Stockfish added 3 more game wins, increasing its lead to 13. There are only 10 more game pairs to play, Leela's chance of winning is mathematical at best since it needs to win biased game pairs from both sides. 

Games 71-72 started with a 14-ply book sideline in the Robasch defence. In game 71 the engines exchanged pawns on the queen side and opened a file. After some shuffling Leela gave a pawn but evals kept decreasing. The first pieces to be exchanged were the queens on move 37, the engines continued to exchange pieces including a bishop for rook trade by Stockfish. The game reached a RB vs BN position with black two pawns up, Leela managed to capture one pawn but no more. The game was adjudicated when Leela lowered its eval for the draw rule on move 95. 

In game 72 there were no exchanges after the start, on move 17 the engines opened a file on the queen side. Stockfish thought Leela's move 19 was a blunder and its eval increased quickly. Leela saw the game continuation in its PV but it thought it was perfectly safe, its eval even decreased. A series of exchanges opened the board and exposed the black king. Leela was surprised by Stockfish's move 29.

Leela's eval jumped over 2. Stockfish's plan involved getting a knight to e3, then a queen sacrifice for a knight brought the black queen to d6 and a knight fork on f4 captured the black queen. The result was a BN vs R ending.

Leela was on time to stop the white passer on b5, Stockfish abandoned it but used its piece advantage to capture all the remaining black pawns. The game reached a 7-man tablebase win that Stockfish had no problem to convert. Stockfish wins the game pair, it leads 16-5

Games 73-74 started with a 15-ply book sideline in the KID normal variation, with a black bias. In game 73 Stockfish gave a pawn to open the queen side, Leela moved its king without castling. (Negative) evals came down, the engines started to exchange pieces and clear the board. The game reached a RN vs Q ending, Stockfish a pawn up with all pawns on the king side. Stockfish thought it had an advantage but other than trade a pair of pawns it could only shuffle. An occasional pawn move reset the counter and Stockfish wouldn't lower its eval. The game was finally adjudicated on move 153.  Up to transposition game 74 repeated the reverse for 14 plys, with Stockfish a pawn up and its king unable to castle. The engines exchanged minors, then in a long PV agreement Stockfish gave a rook for a bishop and pawn. Evals came down, Stockfish pushed a passer in the center to the 6th rank. Leela captured the passer and the engines exchanged a pair of rooks, evals were low enough in a QBN vs QRB position for an early adjudication.

Games 75-76 started with a 24-ply book, a sideline of the closed Ruy Lopez, Ragozin-Petrosian variation. In game 75 Leela locked the center, there were no exchanges until move 22. Then in a PV agreement the engines opened the queen side and exchanged a few pieces. Leela went up a pawn and evals slowly increased. The engines moved their remaining major pieces to the queen side and then exchanged them, leaving a BNN vs BNN position on move 40.

Stockfish had to block the extra white pawn as well as protect its weak pawns against Leela's knights, it saw it was in trouble and its eval jumped. On move 57 Leela captured a second pawn, then exchanged a pair of knights and created an advanced passer in the center. 

Leela traded bishops and pushed its passer to the 7th rank. It finished the game by reducing to a king and pawns tablebase win.

In game 76 Stockfish chose not to lock the center, there were no exchanges until move 21 and evals were stable. The engines exchanged a pair of pawns and then Stockfish pushed a pawn to h6, trapping a black bishop in the corner. After another pawn exchange Stockfish moved its queen forward on the king side. From move 37 Stockfish's eval started to increase, Leela moved its queen to the queen side, ignoring the danger to its king. Stockfish thought Leela made a mistake on move 38 and its eval jumped over 3 as it offered a knight on the king side. 

Leela refused to take the knight and expose its king to attack. Its eval jumped only 3 moves later, in a series of exchanges the engines removed all rooks and Stockfish went up a pawn with a doubled passer in the center. For a while the engines shuffled, Stockfish improved its pieces' position. When it was ready it gave a pawn and opened the f file. Leela's defence collapsed and Stockfish captured a piece.

The white king was exposed but Stockfish was able to find a safe square. The engines exchanged queens and Leela lost its bishop to stop a passer, mate was a matter of time. There were two white wins in this game pair, Stockfish leads 17-6

Games 77-78 started with a rare sideline of the Vienna game. In game 77 Stockfish kept its king uncastled, Leela's eval increased for a short while. Stockfish pushed pawns on the king side and evals came down, Leela offered a knight in the center and Stockfish didn't take for a few moves. After Stockfish captured the knight the center opened, there was a series of exchanges that reduced to a rook ending. Leela was a pawn up but evals were close to 0 and the game was adjudicated. 

In game 78 Leela blundered on move 8 (!), at least according to Stockfish. Stockfish went on the attack, it opened the king side and moved its queen forward, forcing the black king to move without castling. In a few moves all the white pieces on the board were involved in hunting the black king, it looked scary for Leela.

Leela exchanged queens to get its king out of immediate danger. Stockfish went a pawn up and the engines continued to exchange pieces until only RB vs RN remained.

Stockfish's advantage appeared to be small but Stockfish was confident it was winning. It slowly pushed the h pawn forward, ignoring Leela's plan on the queen side. Eventually Leela captured the bishop but gave a rook to stop the passer. It took Stockfish 35 moves to convert the R vs N ending, including a beautiful use of zugzwang to capture a pawn. Stockfish wins the game pair, it leads 18-6

Games 79-80 started with a 12-ply book in the English opening, three knights system, played in a few high level human chess games. In game 79 Stockfish's eval increased for a few moves after the start, then came down while there was a series of exchanges in a PV agreement. Leela was a pawn up, after some shuffling the engines reduced to a QRB vs QRB position. The shuffle continued, evals were low but there were too many pawns for the draw rule. After more exchanges the game reached a queen ending, Leela gave pawns and pushed a passer to the 7th rank, Stockfish ended the game with checks and the draw rule.

In game 80 Stockfish's eval increased from the start, a series of exchanges in a PV agreement (not the same as the reverse game) opened the position and Stockfish was a pawn up. Stockfish created a passer on the queen side, it looked dangerous though it was still on its initial square.

Stockfish pushed the passer to the 7th rank while Leela captured pawns on the king side and exposed the white king. Leela attacked with queen and knight, it captured the white queen but lost a rook and Stockfish queened the passer. Leela delayed as much as it could with checks, but when the white king found a safe square Stockfish hunted the black king and mated. Stockfish wins the game pair, it leads 19-6.


Sunday, November 13, 2022

Season 23 superfinal games 61-70

After 70 games Stockfish leads 15-5 with 50 draws. Leela won a game after 50 games without a win. It also won a game pair, its second of the match. Stockfish managed to win two game pairs and its lead increased to 10. As expected the opening bias is increasing, there were 3 decisive game pairs and one with two white wins.

Games 61-62 started in a rare sideline of the Pirc defence, Bayonet attack. In game 61 the engines played out a long PV agreement, both kings were exposed and moved without castling to the king side. Queens were exchanged and Leela was a pawn up with a central passer. The game reached a RBN vs RBN position and evals slowly came down. The engines mostly shuffled for a long time, Stockfish captured the passer on the 7th rank and the game was adjudicated on move 72. 

In game 62 the engines entered a different PV agreement, this time the h file was opened and both engines castled long. Evals were stable for a while, Stockfish thought Leela's move 26 was a mistake and its eval jumped. Leela thought for 11 minutes and its eval also jumped, it was very difficult to see what the engines understood. It was evident that Leela moved a knight and moved it back, losing 2 tempi. Also, Leela had a knight on h5 that was weak, almost trapped. Evals slowly increased though the engines seemed to be shuffling. 

Evals continued to increase, Leela played waiting moves while Stockfish slowly moved its pieces to the queen side. When it was ready Stockfish pushed pawns on the queen side and opened the a file. Now the black king was in danger and Leela was effectively playing a piece down. 

Leela gave a rook for the bishop and freed its knight, but it was too late. Stockfish exchanged queens, captured two pawns and pushed a passer to the 7th rank. Leela lost material and was mated. Stockfish wins the game pair, it leads 13-3.

Games 63-64 started with a rare sideline of the French defence, Paulsen variation. In game 63 the engines castled in opposite directions. Evals started to increase on move 15, both queens moved forward and captured a pawn defender of the opposing king, then queens were exchanged. Stockfish pushed pawns on the king side and the game reached a RRB vs RRB position. 

Stockfish opened the king side and the white king walked to the center. Leela gave a pawn and opened the queen side, it pushed a passer and Stockfish gave a bishop to capture it. After exchanging a pair of bishops Stockfish was two pawns up with two passers while Leela was down to 2 pawns.

Leela captured the f pawn while Stockfish pushed its passers. Leela captured a passer and was on time to protect its pawns with the king and prevent Stockfish from queening the second passer. Leela captured the remaining black pawns, the game ended in a tablebase win.

In game 64 the engines castled in opposite directions again. Stockfish gave a pawn on the queen side and half opened the a file for its rook. There were a few minor piece exchanges, Stockfish had strong pawns in the center and evals slowly increased. Leela gave a pawn on the king side to develop its rook, then it created a passer on the queen side, giving Stockfish a central passer at the same time. The black king had weak pawn support and evals jumped. In a series of exchanges the engines reduced to a RRB vs RRB position. 

After a few pawn trades only one black pawn remained on the queen side and the black king was an easy target for the white rooks. Both engines had passers on the g file, both were blocked one square before promoting. Leela had to use a rook as a blocking piece, effectively Stockfish had a piece advantage and it was able to capture the black passer on g2. 

Stockfish captured the black f pawn, then slowly pushed its d pawn forward. Leela delayed with checks but ultimately had to give its bishop to stop the queening. After that Leela gave up and was mated. There were two white wins in this game pair, Stockfish leads 14-4.

Games 65-66 started with a 14-ply book in the KID semi Averbakh system. In game 65 Leela pushed pawns on the king side and kept its king uncastled. Leela's eval increased after move 14 and it surprised Stockfish on move 15. Stockfish thought for 21 minutes, it captured a pawn on the queen side and its eval jumped over 2. Leela pushed a pawn to h6 and trapped a black bishop in the corner. On move 20 Leela regained the pawn on the queen side, Stockfish thought this was a mistake and its eval dropped back. There was a series of exchanges that opened the position, Leela captured a knight and had a passer on the 7th rank, but Stockfish threatened the white king. Stockfish captured the passer and a knight, Leela was a pawn up but evals were close to 0. There were too many pieces for the draw rule, the game was adjudicated after the engines exchanged a pair of rooks.

In game 66 again Stockfish pushed pawns on the king side and didn't castle its king. The engines opened a file on the queen side, Stockfish walked its king to the king side corner. On move 19 Leela chose to exchanged a pair of bishops, Stockfish thought this was a mistake, evals started to increase. Stockfish concentrated forces on the king side and the black king walked back to the center. When it was ready Stockfish opened the h file.

The engines exchanged a pair of rooks, Leela defended the back rank against the white pieces. Stockfish managed to capture a pawn, the black king ran to the queen side for safety. Stockfish reduced to a QRN vs QRN position and captured two more pawns. 

Stockfish had 2 passers, after exchanging queens and trading a pair of pawns Stocfish added another  passer. Leela captured one passer, lost its knight for the second one but the third promoted to a queen and Stockfish mated. Stockfish wins the game pair, it leads 15-4.

Games 67-68 started in the Trompovsky attack, Raptor variation. In game 67 Stockfish grabbed an early pawn with its queen, Leela pushed a pawn to h6 and both kings stayed in the center uncastled. The engines opened the board with a few exchanges and then queens were off.

Evals increased though it seemed the engines were shuffling, Stockfish had many weak pawns to protect and Leela moved its pieces trying to decide where to attack. After a while Leela captured two pawns and exchanged a pair of rooks. Stockfish captured a pawn back, equalizing material but evals were high.

Stockfish's defence gradually collapsed, Leela captured the g pawn and then the h pawn, creating and advanced passer. Stockfish had to give a bishop to stop the promotion. Leela slowly improved using its piece advantage. Stockfish couldn't prevent a queening followed by mate.

In game 68 there were no exchanges after the start, the engines locked pawns in the center and king side to form a long pawn wall. On move 34 a pawn exchange opened the b file, after a bishop exchange the engines started to shuffle. Stockfish wouldn't lower its eval, it extended the game with piece exchanges and by move 152 only BN vs BN remained. After Stockfish's eval was low enough there were still too many pawns for the draw rule, the game finally ended on move 256 by a 3-fold repetition. Leela wins the game pair, Stockfish leads 15-5

Games 69-70 started with a sideline of the French, Winawer, retreat variation, with the center locked and black a pawn up. In game 69 Leela used a rook to regain the pawn and sent its queen forward on the king side. Stockfish didn't castle its king, its king side rook moved and the queen side was open. Leela castled late and evals came down, the engines traded a pair of rooks and started to shuffle in a QRN vs QRN position. There were too many pawns, the engines reduced to a queen ending and the game was adjudicated on move 93. In game 70 Stockfish used its queen to regain the pawn, Leela moved its king without castling and Stockfish also kept its king uncastled. The engines opened the king side a little but did not try to attack, instead they shuffled for almost 50 moves. A few exchanges opened the queen side, evals came down and shuffling resumed. On move 113 the engines started to exchange pieces, and the game was adjudicated on move 131.


Friday, November 11, 2022

Season 23 superfinal games 51-60

After 60 games Stockfish leads 12-3 with 45 draws. Stockfish won two game pairs and increased its lead to 9. It looked like Stockfish would win another game pair, but its eval dropped from 4 to 1 suddenly and the game ended in a draw. Leela hasn't won a game pair, or indeed a game, for a long time. The main question seems to be the lead Stockfish will have at the end of the match.

Games 51-52 started with a 17-ply book sideline of the KID Saemisch variation, with the center pawns locked. In game 51 Leela pushed pawns on the king side and opened the g file, then castled long. Stockfish didn't castle, but walked its king to the king side. The engines exchanged a pair of rooks and a pair of bishops, then started to shuffle with a few pawn moves and exchanges. On move 37 Stockfish went a pawn up, it had a pawn wall across the board except for two files on the king side. Evals came down and the engines reduced to a QN vs QN position. There were too many pawns for adjudication, the game ended in a 50 move draw on move 117. In game 52 Stockfish castled long and Leela castled short, the engines opened the c file. Evals came down as the engines shuffled for a while, including a knight exchange. On move 37 Stockfish captured a pawn, Leela's pieces were active on the queen side facing the white king. Evals dropped to 0, the engines reduced to a double rook ending and Leela regained the pawn. The game was adjudicated when enough pawns were exchanged.

Games 53-54 started in the Caro-Kann, Hillbilly attack variation, with black a pawn up and a black bias. In game 53 Leela chose not to exchange queens, Stockfish's (negative) eval started to increase slowly. In a long PV agreement Leela castled short and Stockfish castled long, Leela pushed a pawn and opened a file on the queen side but its eval started to increase as well. In a series of exchanges the engines reached a material imbalance of QN vs RRB.

Evals were stable for a while and the minor pieces filled the center. Stockfish's eval started to jump, Leela's eval followed 2 moves later. The engines exchanged minors and reduced to a QB vs RRB position, then Stockfish gave two pawns as it formed a winning plan.

The idea was for to use dark squares, the black king moved forward through the center and Stockfish moved a rook to the back ranks. The threat of mate was more dangerous than any white passer that Leela created. After clearing all the queen side pawns both rooks came forward. The white queen tried to help, Stockfish captured it for a rook. The final moves were played only on dark squares, with the white bishop helpless to stop mate. 

In game 54 Stockfish chose to exchange queens, forcing the black king to move. The engines played out a PV agreement and evals slowly came down. Stockfish castled long and then evals dropped when in a series of exchanges the engines reduced to a RB vs RB position. Stockfish was a pawn up with a central passer. The engines mostly shuffled for a long time, the game was adjudicated when enough pawns were exchanged. Stockfish wins the game pair, it leads 11-3.

Games 55-56 started with a 16-ply book sideline in the QGA Alekhine system variation. In game 55 the engines exchanged minor pieces and there were no exchanges after that. Stockfish did not castle its king, after move 20 evals came down a little and the black king walked to the king side. There was a period of shuffling and a few pawn moves, on move 48 the engines exchanged minors and resumed shuffling. After move 66 there was a series of exchanges and evals dropped to 0, the game reached a rook ending and was adjudicated. Game 56 started similarly, a minor piece exchange and Leela not castling. On move 17 the engines started a long PV agreement, Stockfish gave a rook for a bishop and two pawns and opened the queen side. Stockfish cleared all the black pawns on the queen side, the game reached a RBN vs RRN position with white two pawns up and two passers, one advanced. Stockfish pushed the passer to the 7th rank, its eval was over 4. Then without warning Stockfish's eval dropped, while playing a move in its PV. In fact Stockfish and Leela predicted the game continuation many plys into the future, but Stockfish suddenly changed its mind about its winning chances. In another PV agreement Leela captured the passer and Stockfish captured a rook for a knight. The engines traded pawns until there were only a few on the king side. It took Stockfish a long time to reduce its eval for the draw rule.

Games 57-58 started with a sideline of the Philidor defence. In game 57 Stockfish pushed pawns on the queen side, Leela pushed pawns on the king side and neither engine castled its king. The engines exchanged minor pieces and evals came down, all pawns were on the board until move 33. The engines shuffled for a while, exchanged a pair of rooks and resumed shuffling. On move 79 Stockfish gave two pawns and moved its queen forward to end the game in check repetition.

In game 58 the engines repeated the reverse game for 21 plys up to transpositions. On move 20 Stockfish went a pawn up and created a passer on the queen side. Leela castled short and Stockfish pushed a pawn and opened the king side. On move 31 all the pieces were still on the board, the engines started to play out a long PV agreement. First queens were off, evals increased as the engines reduced to a RNN vs RBN position.

For a while Leela's eval came down a little, Leela exchanged rooks and regained the pawn. Stockfish's eval jumped higher as it pushed the a pawn forward. Leela was still optimistic after it gave a knight to capture the passer, resulting in a 7-man NN vs B position.

The position was a tablebase win, Leela's eval jumped a few moves later. Stockfish used only 6-man tablebases but it made no mistake and won. Stockfish wins the game pair, it leads 12-3.

Games 59-60 started with a sideline in the Caro-Kann, Bronstein-Larsen variation. In game 59 Stockfish kept its king in the center, there was weak pawn support on both sides. Leela pushed a pawn in the center and it became an advanced passer. The engines opened the queen side and exchanged pieces, evals came down and the game reached a QRR vs QRR position. The game was adjudicated when evals were low enough for the draw rule. In game 60 Leela walked its king to the king side, pieces stayed longer on the board. In a long PV agreement Stockfish gave 3 pawns and opened the center, exposing the black king to attacks. Evals came down, Stockfish got two pawns back and the game reduced to a RB vs RN position. Stockfish had a blocked passer on the 7th rank, the engines traded pawns and Leela captured the passer. The game was adjudicated a few moves later.


Thursday, November 10, 2022

Season 23 superfinal games 41-50

After 50 games Stockfish leads 10-3 with 37 draws. Stockfish won two more game pairs, one after a rare blunder by Leela. Leela realized it was a blunder immediately in the following move, suggesting it hadn't considered Stockfish's reply - a bit strange. At the half way point Stockfish has a lead of 7, and Leela only won one game pair so far. The opening bias is slowly increasing, we should see more Leela wins but it is safe to sat that Stockfish will win the match easily.

Games 41-42 started with a 4-ply book 1.e4 c5 2. Nc3 b6. In both games the engines transposed into more popular variations of the Sicilian closed opening. In game 41 there was one pawn exchange after the start, Leela castled long and Stockfish kept its king in the center. Starting from move 17 the engines exchanged minor pieces, evals came down and the game reached a QRR vs QRR position. The engines continued to trade pawns and a pair of rooks, the game was adjudicated a few moves later. In game 42 again Stockfish castled long and Leela did not castle its king. Stockfish gave a bishop for two pawns, it forced the black king to move and opened the h file. Evals came down as the engines exchanged queens and a pair of rooks, Stockfish attacked the back rank with its remaining rook. The engines exchanged rooks and only minors remained, exchanges continued until only B vs BN remained. Stockfish captured all the black pawns, it was 4 pawns up when the game was adjudicated. 

Games 43-44 started with a 26-ply book in the Sicilian Dragon, Yugoslav attack, that was played in high level human chess (including a Karpov-Kortschnoi game from 1974). Up to a transposition the games repeated for 21 more plys, white pushed pawns on the king side and black kept the files closed, most pieces were still on the board. In game 43 Leela's evals increased a little, then came back down after Stockfish captured a pawn and reduced to a QRB vs QRN position. The engines traded pawns, evals were low enough for the draw rule. In game 44 played out another long PV agreement after diverging from the reverse game, resulting in a QRB vs QRN position. Evals came down as Leela attacked the white king, they dropped to 0 after the queens were exchanged. The game was adjudicated when enough pawns were exchanged. 

Games 45-46 started with an 8-ply book in a rare Queen's pawn game variation. In game 45 Leela weakened the black pawn structure using early exchanges, its eval increased a little. Stockfish pushed a pawn to e4, this restricted the white pieces a little. The engines played out a long PV agreement, the d pawn became a white passer but evals came down. Stockfish gave two pawns to open the king side, then it sacrificed a rook but got it back a few moves later. The game reached a QB vs QN position, Leela was up a pawn but the white king was exposed to checks, evals were close to 0 and the game was adjudicated. 

In game 46 Stockfish castled long, the engines mostly agreed in their PV and evals were stable after the start. A minor piece trade opened the g file, Stockfish thought Leela would also castle long but Leela instead chose to capture a free pawn. Stockfish's eval immediately jumped over 3 in a knight sacrifice that surprised Leela.

Leela's eval also jumped (why then make that move??), taking the knight weakened the black e pawn and now the black king was in danger. Stockfish captured two pawns and created two advanced passers. In a series of exchanges Leela captured one passer and reduced to a RRB vs RRN position. material was equal but Stockfish had a protected passer on the 7th rank.

Progress was very slow but evals increased steadily. It took Stockfish 25 moves to capture a pawn on the king side, then exchange a pair of rooks. Leela captured the white passer, but Stockfish reduced to a won B vs pawns ending (Leela could have taken the bishop and lead to a king and pawns ending which was also a white win). Stockfish wins the game pair, it leads 9-3.

Games 47-48 started with a sideline of the Queen's Indian accelerated. In game 47 the engines castled in opposite directions, Leela captured a rook for a knight and evals came down in a series of exchanges. The game reached a QR vs QB position with Stockfish two pawns up, Stockfish captured another pawn but its king was exposed. Leela moved its queen forward and gave checks, evals were low enough for the draw rule.

In game 48 the engines blocked the center, and again castled in opposite directions. There were no exchanges after the start and evals were stable, after a pawn exchange on move 18 the c file opened and Stockfish's eval increased slowly. On move 32 there was another pawn exchange on the king side, all pieces were still on the board.

Leela has to protect the pawns on b5 and g6, and its king was not safe with the g file half open. Evals continued to increase as Stockfish moved all its major pieces to the g file. On move 49 Stockfish pushed the f pawn forward, the first pieces were removed when the engines exchanged a bishop pair. Stockfish opened the g file to attack the black king, and on move 56 it finally went a pawn up.

Leela was busy defending its king and it had to abandon the b5 pawn. Stockfish exchanged all rooks and captured the remaining black pawns. It was 5 pawns up in a BN vs BN position and mated after queening twice. Stockfish wins the game pair, it leads 10-3.

Games 49-50 started with a 14-ply book sideline in the Sicilian closed variation. In game 49 both engines castled and then a pawn exchange opened the f file and both evals jumped a little. The engines exchanged a pair of rooks, then continued to exchange pieces and evals drifted down. Leela was a pawn up in a QRB vs QRB position, evals were close to 0 but the game continued until enough pawns had been exchanged. In game 50 Leela castled and Stockfish kept its king in the center. The engines played out a long PV agreement where Stockfish gave a bishop and then in a long series of exchanges regained the material. The game reached a RRN vs RRB position and evals came down. The engines continued to trade pawns, Stockfish went a pawn up while Leela pushed a passer on the king side. Stockfish captured the passer and exchanged a pair of rooks, the game was adjudicated when evals were low enough.