I learned so much from the exercise in Evaluation Anomaly - Long Combination that I decided to repeat it on another position from TCEC Season 8 - Evaluation Anomalies. This time I chose game 18, Stockfish - Komodo.
The composite chart below shows four key positions from the game. I don't expect anyone to follow the moves mentally, but the game can be played on TCEC - Archive Mode, using the same instructions given in 'Long Combination'. All of the key move metrics are shown in the helpful interface used there.
The first position shows the game after the initial eight moves of the opening variation imposed on the engines. Stockfish's next move as White was 9.Bb3, which it evaluated as wv=0.17. Komodo's move as Black was 9...h6, with a value of wv=0.23. Note that both 'wv' values are close to the 0.20 predicted for the initial start position. For the other positions, I'll use a sum of the values from a White-Black move pair (0.17 + 0.23 = 0.40 here) to compare the evaluations through the game.
The game continued 10.Nf1 Re8 11.a4 b5 12.Ng3 Bd7 13.Bd2 b4 14.a5 Rb8 15.h3 bxc3 16.bxc3 Qc8 17.Bc2 Rb2, then 18.Qc1 (wv=0.54) and 18...Qb8 (wv=0.37), reaching the second position. Here the combined wv is 0.91, more than double the value in our first position. This was followed by 19.Nf5 Qb7 20.Ne3 Qb8 21.Nh4 Ne7 22.Nc4 Rb7 23.d4 Ng6.
Stockfish - Komodo, TCEC Season 8 Superfinal, game 18
The third diagram shows the position after 24.Nxg6 (wv=0.67) 24...fxg6 (wv=0.56), with a combined wv of 1.23. This is the highest evaluation reached in the game, which continued 25.Be3 Kh7 26.Qd2 Be6 27.Bd3 Bxc4 28.Bxc4 Nxe4 29.Qd3 c6 30.d5 cxd5 31.Bxd5 Nc5 32.Bxc5 Bxc5.
The next two moves -- 33.Bxb7 (wv=0.27) 33...Qxb7 (wv=0.09) -- reach the fourth diagram and involve an exchange sacrifice, where Black has one Pawn as compensation. The combined value of wv=0.36 is substantially below the combined wv from the third diagram.
What happened between the third and fourth diagrammed positions to cause such a dramatic decline in the evaluation? In the third position, a pair of Knights has just been exchanged, but all of the other pieces are still on the board. In the fourth position, three pieces remain for each side. Between the two positions, half of the pieces were swapped off. Although exchange sacrifices can be tricky to evaluate, the position in the fourth diagram looks harder to win for White than to draw for Black -- the Bishop is well placed for defense -- and the evaluation proves to be accurate.
The conclusion is that the evaluation in the third position is overly optimistic. Using the same calculation explained at the end of the post on 'Evaluation Anomalies', the ~0.60 advantage for White gives a 67% chance of winning the game. The third position might simply be in the 33% of positions that are more difficult to win. We are, after all, dealing with probabilities here. Only after more pieces are exchanged do we start to see the eventual outcome.