After 20 games Stockfish leads 5-1 with 14 draws. Or is it 4-1?. Game 17's surprising outcome has caused the first heated fan argument of the superfinal, see details below. Houdini finally managed to win a game and Stockfish is no longer unbeaten in the season.
Game 17 started with a 13 move book of the Sicilian Dragon, Yugoslav line. Stockfish had an eval advantage and a strong king side attack, and it was willing to sacrifice a bishop to put pressure on Houdini's king.
In an amazing sequence of moves that both engines saw in advance Houdini sacrificed its queen and gained a rook, then Stockfish captured a rook, and then Houdini combined its rook and two bishops to force the white king to a8 (!!) and win back a rook.
As the game continued Houdini allowed Stockfish to capture its pawns while taking the king side pawn. Stockfish slowly advanced its queen side pawns and to prevent queening Houdini traded a rook for the two. Through all of this the evals gradually decreased, and for the last 10 moves both evals were 0. Then came the shock: the game ended in a table Q vs BB position and Stockfish was given the win. Inconceivable !! Both engines used 6-man tablebases and both evaluated the position at 0 in a 5-man tablebase win ???
It turned out that the reason was simple. The win is 72 moves long, with no pawns and no captures for more than 50 moves. Thus it was a tablebase win but a draw under conventional chess rules. TCEC rules use tablebase decisions without taking into account 50 move draws (so called 'cursed wins'). Both engines were playing with the 50 move rule logic and were unaware of this rule in TCEC.
What was to be done about this? Either keep the win for Stockfish or decide it was a draw after all. A replay was also suggested but that's a different question, since any game replayed will develop in new directions, so this should be done only if there was some technical problem with the game, not really the case here. Everyone had an opinion and for days people argued about this in the TCEC chat. Some of the arguments I saw:
a) Rules are rules. This was Martin's argument and it is very strong, especially when you consider the next argument.
b) This is a known TCEC incosistency. There were previous TCEC games that ended similarly and decided by tablebase despite violating the 50-move rule.
c) The 50-move rule should be ignored altogether in engine game endings, tablebases are the real truth and should always decide.
a) Conventional chess and TCEC use the 50 move draw rule, and this should also apply to tablebase decisions.
b) Both engines considered the position a draw since they were using the 50 move rule in their logic. It is likely that the game would have continued otherwise if they knew they were heading for a white win. It is very probable that other drawing lines exist without using cursed wins.
Personally I think this should be declared a draw. I can't argue with 'rules are rules' though. In any case the match result will probably not be affected by the results of this one game.
In game 18 Houdini played with more caution, instead of an all out attack it used slow strangulation. Stockfish's pieces hardly had room to move while Houdini could plan the best way to proceed.
Houdini exchanged a few pieces and went a pawn up. The evals gradually increased to over 2. After exchanging queens only RRN vs RRN remained, Houdini's central passer looked strong.
Stockfish's position deteriorated fast. Houdini used a nice combination to fork a knight and a rook. Stockfish gave the rook for a knight and together with the advanced passer this was enough to give the win to Houdini.
There were many exchanges after the opening of game 19 and after 26 moves only QRN vs QRN were left. The engines continued to capture pawns, evals were 0 after queens were exchanged. The game ended in a tablebase draw. In game 20 the engines kept their pieces on the board and started to shuffle behind pawn lines on move 27. Stockfish had a bishop trapped in Houdini territory, but it had enough defenders so that Houdini couldn't capture it. The evals dropped to 0 when Houdini broke the blockade to avoid a 50 move draw, and the game reached a draw in a RB vs RB ending.