Jonathan Sarfati: More from the Computer Sees Everything files:
16...♕d7, missed by all annotators, threatens to win the ♘h4 with ... g5. 17. g4 to support a flight square on f5 is beaten quickly with h5.
Arcadios Feneridis suggested 19.♘e2 h4 20.d4 exd3 21.cxd3 on the grounds that it helps White develop, and maybe exchange ♕s on the c-file. But it wouldn't help after 21... g4, e.g. 22.♕c2 (22.hxg4 h3)
22...gxh3 23.d4 ♖h5 24.dxc5 ♖xc5 25.♘c3 h2+ 26.♔h1 h3 27.f3 ♘h5 28.♘f1 ♘g3+ 29.♘xg3 ♕xg3 30.♗e3 ♖g8
23... ♗a7 was good enough, over -3 evaluation, but even stronger and missed by me and all annotators was:
23...♖xg2+ 24.♔xg2
(24.♘xg2 ♘g4)
24...♖g8+ 25.♔h1 ♕g3 26.♘fg4 ♘xg4 27.♕g2 ♕xh4+ 28.♔g1 ♘h6
Zygmunt Frankel and Murray Chandler both suggested 28...♗h2 as stronger. Indeed it does have the highest engine evaluation of -10, e.g. 29.♔e1 ♕xg1+ 30.♕f1 ♘xe3 31.dxe3 ♖xg2 32.♖xg2 ♖d1+ 33.♔xd1 ♕xf1+
But I had already worked out that the game move 28... ♘f4 was a clear cut win. Indeed the engine evaluation is -7.
32...♖xg2 was mate in 6. No matter; I could see that the game move 32...♕xg2+ would win a whole ♕. The engine calculates mate in 9. I missed 32...♖xg2, as did all the annotators, I suspect because we weren't looking for it after seeing that 32...♕xg2+ won quickly. Same principle with missing 28...♗h2 because 28... ♘f4 was good enough.
But I admit that 23...♖xg2+ would have been a nice way to finish.