Carlsen-Caruana WCC Game 6 – The actual forced-mate sequence that was missed
Game 6 of the 2018 World Chess Championship is perhaps most notable for an incredible winning sequence that was found by a computer named Sesse, but missed by the players. As widely reported, the Norwegian computer running the Stockfish engine announced a mate in 30 moves for Black after 68.Bc4, and even Carlsen was asked about it in the post-game press conference. Many GMs have since explained this very impressive winning manoeuvre starting with 68…Bh4!!, but they all finish their analysis with 84…Kxh7, because that capture leaves Black in a clearly won position. Like a lot of players, I was interested in seeing the full forced-mate sequence, but it wasn't easy to find online. And strangely, some sources claim that it's a mate in 36 moves rather than 30, while others that provide a specific mating line suggest that it takes even longer, such as the M42 given in a Chess.com news article, and a M63 from a Reddit post that apparently shows a Sesse screenshot. So what is the actual forced mate and how many moves does it involve? The answer is… none of the above!
I first noticed something was amiss when looking at the just-mentioned M42 in the Chess.com article. It commences with Sesse's 17-move study-like play that wins the h-pawn with 84…Kxh7, but note that this capture reduces the position to 7 pieces, meaning from here we can use endgame tablebases to determine the shortest mating line with best play by both sides. (If you're not familiar with tablebases, which effectively play perfect chess in light positions, check out my introductory blog, Adventures with endgame tablebases.) And the Lomonosov tablebases indicate that it takes another 41 moves for Black to force mate, after the pawn capture. Checking some random moves in the M42 sequence - as well as those in the Reddit post M63 - against Lomonosov's perfect analysis confirms that they include many sub-optimal choices. Now if the tablebases prove that after the first 17 moves, Black needs 41 more to mate, that's a total of 58 moves. So how could Sesse have announced a mate in 30 or 36? Perhaps it started with another manoeuvre that's quicker, but then all of the GMs would have been discussing a weaker line, which doesn't make sense.
After more digging around online, I eventually found the full variations for the purported M30 and M36. Both were posted by a user on Chessgames.com, and confirmed by other sites. Here's an image for each; the first (M30) is from a live coverage of the game on Chessdom.com, and the second (M36) is a Sesse screenshot posted on Imgur.com.
Both lines indeed begin with the familiar series of 17 moves, so the GMs were correct, but if they had examined Sesse's ensuing play once the h-pawn is lost, they would have noticed something quite bizarre. In the M30, White's play isn't merely sub-optimal, but outright suicidal! After leaving the bishop en prise, White obligingly marches the king to a corner to make it easy for Black to mate, as if this were some sort of helpmate problem. The M36 isn't as obviously erroneous, but consider the position after 98.Kh2 (=31.Kh2 below), for instance: the Stockfish on this site finds various mates-in-3 for Black here, but Sesse took 6 moves to mate.
How could the same Sesse/Stockfish that discovered the brilliant 68…Bh4!! and 70…Ng1!! make these terrible moves in its analysis? Sesse is even linked to the Lomonosov tablebases, according to their site. I'm no engine expert and could only hazard a guess that it relates to some incorrect or incomplete tablebases lookup. Tablebases provide two kinds of information: (1) the win-draw-loss outcome of a position and (2) the quickest route to a forced win/loss for any decisive move. Because the two types of data could be stored independently, perhaps Sesse had access to (1) and knew that 84…Kxh7 would result in a won position, but not (2) and so chose non-optimal moves based on a formula best known to itself.
Until this software issue is fixed, the Sesse team should really avoid announcing mate in X moves, and chess journalists should take the computer's pronouncements of such with a grain of salt. But while there was actually no forced mate in 30 or 36 moves in Game 6, there was one in 58 moves as mentioned, and Sesse could have declared such a daunting mate if it was able to consult the tablebases properly. For the record, here is the correct mating sequence that combines Sesse/Stockfish's winning manoeuvre with Lomonosov's perfect finish.
Congratulations to @Rocky64 and @ddtru for winning Chess.com's November Blog of the Month contest. Learn more here.
