Long analysis of some opening positions

[OneTrickPony]

I am making this thread to dump long analysis of some opening positions I run. Maybe it's of some interest to practical players here. It might be fun to see if engines of the future change some assessments. No promises to post often but for now I am dumping a lot of those to my notes and I thought why not share publicly.
I will start with recently popular line which might become a practical refutation of classical 1.d4 setups:

[pgn]1.d4 Nf6 2.c4 e6 3.Nf3 d5 4.Nc3 c5 5.cxd5 cxd4 6.Qxd4 exd5 7.e4 dxe4 8.Qxd8+ Kxd8 9.Ng5 Be6 10.Nxe6+ fxe6[/pgn]

Lc0 v0.26.3, J96-28 net by JHortos:

1. 27 [+52.05] 11.Bg5 Nc6 12.Bc4 Ke7 13.Nxe4 Kf7 14.Bxf6 gxf6 15.O-O-O Rd8 16.Rxd8 Nxd8 17.Rd1 Be7 18.Nd6+ Kg6 19.b4 f5 20.a3 Kf6 21.Kb2 h5 22.h4 a6 23.Be2 b5 24.g3 Nc6 25.Nb7 e5 26.Rd7 Ke6 27.Rc7 Nd4 28.Bd1 f4 29.gxf4 Bxh4 30.Nc5+ Kf5 31.fxe5 Kxe5 32.Rf7 Ne6 33.Nd3+ Kd6 34.Rf5 Be7 35.Rxh5 Bf6+ 36.Ka2 Rxh5 37.Bxh5 Bd8 38.Bg4 Nd4
2. 27 [+51.77] 11.Bc4 Ke7 12.O-O Kf7 13.Nb5 Na6 14.Bf4 Rd8 15.Rac1 Nb4 16.Nxa7 Be7 17.Be3 Nd3 18.Rc2 Nd5 19.Nb5 Bf6 20.Bc1 N5b4 21.Re2 Nxc1 22.Rxc1 Nd3 23.Rcc2 Nxb2 24.g3 Rd1+ 25.Kg2 Nd3 26.Kh3 Nc5 27.Bb3 Nxb3 28.axb3 Rhd8 29.Rxe4 R1d2 30.Rc7+ Kg6 31.Kg2 Rb2 32.Rg4+ Kh5 33.Rf4
3. 27 [+51.93] 11.Bd2 Bb4 12.g3 Ke7 13.Bg2 Rd8 14.Nxe4 Nc6 15.Bxb4+ Nxb4 16.O-O Nd3 17.Rfd1 Rac8 18.Rd2 Ne5 19.Rxd8 Rxd8 20.f4 Nd3 21.Nxf6 gxf6 22.Bxb7 Nxb2 23.Rc1 Rd1+ 24.Rxd1 Nxd1 25.Ba6 Kd6 26.Kg2 Nc3 27.a3 e5
4. 27 [+51.87] 11.g3 Bb4 12.Bd2 Ke7 13.Bg2 Rd8 14.Nxe4 Nc6 15.Bxb4+ Nxb4 16.O-O Nd3 17.Rfd1 Rac8 18.Rd2 Ne5 19.Rxd8 Rxd8 20.f4 Nd3 21.Nxf6 gxf6 22.Bxb7 Nxb2 23.Rc1 Rd1+ 24.Rxd1 Nxd1 25.Ba6 Kd6 26.Kg2 Nc3 27.Bc4 f5 28.Kf3
5. 27 [+50.83] 11.a3 Bc5 12.Bg5 Ke7 13.Nxe4 Bb6 14.O-O-O h6 15.Nxf6 gxf6 16.Bd2 Nc6 17.f4 Rad8 18.Be2 h5 19.Kb1 h4 20.Bf3 Rd7 21.Bc3 Rhd8 22.Rxd7+ Rxd7 23.Be1
6. 27 [+49.94] 11.Bb5 Bb4 12.Bd2 Ke7 13.Nxe4 Bxd2+ 14.Nxd2 Rc8 15.Nb3 a6 16.Be2 a5 17.Bf3 Nc6 18.O-O a4 19.Nc5 Rc7 20.Rac1 Nd4 21.Nd3 Nxf3+ 22.gxf3 Rac8


228M nodes, 14315 seconds, around 56GB of RAM taken.

[Dann Corbit]

I have this in my database:
[d]rn1k1b1r/pp4pp/4pn2/8/4p3/2N5/PP3PPP/R1B1KB1R w KQ - acd 55; acs 1121; bm Bg5; cce 52; ce 10; id "18803"; pm Bg5 {42} Bc4 {5}; pv Bg5 Nc6 Bc4 Ke7 Nxe4 Ne5 Bb3 Kf7 Nc3 Rd8 Ke2 Bc5 Na4 Bd4 Rac1 h6 Bf4 Nc6 Rhe1 g5 Bg3 Ne4 Bc2 Nxg3+ hxg3 Rd7 Be4 Ne7 g4 b6 Nc3 Rhd8 Red1 Be5 Rxd7 Rxd7 a3 Ng6 g3 Ne7 Rh1 Kg7 Rd1 Rxd1 Kxd1 Bd4 f4 Ng6 f5 exf5 gxf5; white_wins 340; black_wins 140; draws 1216;

Here are the outcomes of some particular games:

Code: Select all

Player, move, outcome
"Gimeno D�az, Fernando",Bc4,0-1
"Gorovets, Andrey",Bc4,1-0
"Grinberg, Iosif P",Bc4,1-0
"L'Ami, Erwin",Bc4,1/2
"Navara, David",Bc4,1/2
"Spartinos, Pantelis",Bc4,1/2
"Holt, Conrad",Bg5,0-1
"Schulz, Klaus-J�rgen (Bendorf)",Bg5,0-1
Fire 7.1,Bg5,1-0
"Goodman, Conrad",Bg5,1-0
"Haugen, Arild",Bg5,1-0
"Jean, Pierre",Bg5,1-0
"Khairullin, Ildar",Bg5,1-0
"Kopeikin, Valentin Vasilievich",Bg5,1-0
"Parmet, Daniel Edward",Bg5,1-0
SlowChess Blitz Classic 2.25,Bg5,1-0
"St�l, Carl-Olof",Bg5,1-0
"Werner, J�rgen (Berlin)",Bg5,1-0
Booot 6.3.1,Bg5,1/2
"Budzyn, Roman",Bg5,1/2
"Cade, Steven",Bg5,1/2
Fire,Bg5,1/2
"Fressinet, L",Bg5,1/2
"Giannetto, Salvatore",Bg5,1/2
"Grayland, Stan J.",Bg5,1/2
"Grinkevicius, Alvydas",Bg5,1/2
"Hansen, Michael",Bg5,1/2
Houdini,Bg5,1/2
Houdini 6.03,Bg5,1/2
"Kaan, J. E. F",Bg5,1/2
"Kazmierczuk, Zenon",Bg5,1/2
"Khalifman, Alexander",Bg5,1/2
"Khodoskin, Dmitry Petrovich",Bg5,1/2
"Killer, Oliver",Bg5,1/2
Laser,Bg5,1/2
Lc0 0.18.1,Bg5,1/2
Lc0 0.19,Bg5,1/2
"Leiner, Raimund",Bg5,1/2
"Leiner, Reimond",Bg5,1/2
"Lins, Thomas",Bg5,1/2
"Morcin, Tom",Bg5,1/2
Nemorino 5.08,Bg5,1/2
"Perry, Richard M.",Bg5,1/2
Rofchade,Bg5,1/2
"Rotaru, Ion",Bg5,1/2
Rusty 101019,Bg5,1/2
"Sacerdotali, Sergio",Bg5,1/2
"Sherwood, Russell",Bg5,1/2
"Simeonov, Lyuben",Bg5,1/2
Stockfish,Bg5,1/2
Stockfish 18111122,Bg5,1/2
Stockfish 8,Bg5,1/2
TCEC,Bg5,1/2
"Tseng, Wilbur",Bg5,1/2
Tucano 7.05,Bg5,1/2
"Tyutyunnik, Roman",Bg5,1/2
"Wilhelmi, Dirk",Bg5,1/2
"Wojtaszek, R",Bg5,1/2
Xiphos,Bg5,1/2
Xiphos 0.4.3,Bg5,1/2
"Zatko, Gregor",Bg5,1/2
"Zeihser, Michael",Bg5,1/2

I think the conclusion is clear:
The best move is Bg5 (LC0, SF and players agree)
It is probably going to end in a draw, but there is a small advantage for white.

[OneTrickPony]

Pawn sac in Ba6/Qc2 Queen's Indian:

[pgn]1.d4 Nf6 2.c4 e6 3.Nf3 b6 4.g3 Ba6 5.Qc2 c5 6.d5 exd5 7.cxd5 Bb7 8.Bg2 Nxd5 9.O-O Nc6 10.Rd1 Be7 *[/pgn]

1. 37 [+51.96] 11.Qa4 Nf6 12.Nh4 O-O 13.Nf5 d5 14.Nc3 Nd4 15.Nxd4 cxd4 16.Rxd4 Bc5 17.Rd3 Qe8 18.Qxe8 Raxe8 19.e3 Ba6 20.Rd1 d4 21.exd4 Bxd4 22.Bf4 Rd8 23.h3 Bc4 24.Rac1 Bxc3 25.Rxd8 Rxd8 26.Rxc3 Bxa2 27.Ra3 Rd1+ 28.Kh2 Bd5 29.Rxa7 h5 30.Ra3 Rd4 31.Be5 Rd2 32.Kg1 Rd1+ 33.Kh2 Ne4 34.Bc7 Nxf2 35.Bxb6 Ne4 36.Be3 Nf6 37.Bxd5 Nxd5 38.Rb3
2. 37 [+51.96] 11.a3 Nc7 12.Bf4 Ne6 13.Nc3 O-O 14.Rd2 Nxf4 15.gxf4 d6 16.Rad1 Qd7 17.h3 Rad8 18.e3 Kh8 19.Kh2 f5 20.Nd5 b5 21.Ng1 a6 22.b3 h6 23.Nxe7 Qxe7 24.Ne2 Qc7 25.Ng3 Bc8 26.Nh5 Ne7 27.Qc3 Rf7 28.Rg1 Rg8 29.Rgd1 Rd8
3. 37 [+51.94] 11.Qf5 Nf6 12.e4 g6 13.Qf4 O-O 14.e5 Nh5 15.Qg4 Qb8 16.Nc3 Nxe5 17.Nxe5 Bxg2 18.Nxd7 Qb7 19.Nxf8 Nf6 20.Qh4 Nh5 21.Bg5 Bxf8 22.g4 Ng7 23.Bf6 Bh1 24.Qg3 Ne8 25.Bg5 h6 26.Bf4 g5 27.Be3 Bd6 28.Qh3 Qf3 29.Qxf3 Bxf3 30.Rd2 Bxg4 31.a4 Be5 32.a5 Rb8 33.Nb5 Bf3 34.Re1 f6 35.a6 Kf7 36.Rd3 Bc6 37.Nxa7 Be4 38.Rb3 Bd5 39.Rd3 Be4
4. 37 [+50.17] 11.h4 O-O 12.Nc3 Nxc3 13.bxc3 Na5 14.Ng5 Bxg5 15.Bxg5 Qc8 16.Bxb7 Nxb7 17.Rd5 d6 18.Bf4 Re8 19.Rad1 Qe6 20.e3 h6 21.f3 c4 22.e4 Nc5 23.Rxd6 Qe7 24.R1d4
5. 37 [+50.43] 11.Ne5 Nd4 12.Qd2 d6 13.Nc4 Qd7 14.e3 Nf4 15.exf4 Bxg2 16.Kxg2 Qc6+ 17.Kh3 Qd7+ 18.Kg2
6. 37 [+47.95] 11.Nc3 Nxc3 12.bxc3 Qc7 13.Bf4 d6 14.Nd2 Na5 15.Nb3 Bxg2 16.Nxa5 Bc6 17.Nxc6 Qxc6 18.c4 O-O 19.Rd5 a6 20.Rad1 Rad8 21.a4 Rfe8 22.Qb3 Bf8 23.R1d2 Re6 24.f3

Lc0 26.3, J96-28 net by JHortos, 377M nodes, 100GB of RAM, around 6 hours.

Leela finds refutations to 11.Qa4 and 11.Qf5 but because of MCTS still has those at ~52% (they are completely dead). It looks like either 11.a3 or 11.h4 are practical options here. It would be interesting to see what AlphaZero plays after 11.Qf5 but unfortunately weak Stockfish they played against didn't follow the main line.

[Stephen Ham]

Your message is unclear. Perhaps English isn't your primary language. If so, then thanks for making the effort.

For example, you wrote, "Leela finds refutations to 11.Qa4 and 11.Qf5..." What refutations? Where are they? Are you claiming that 11 Qa4 and 11 Qa5 have been refuted? Or do you mean the opposite, that those two moves refute Black? The confusion is compounded with "...they are completely dead". What?

Either way, all those claims are false as demonstrated by your own Leela output, showing equality in all lines. Basically a score of 52 represents equality.

Also, you claim Leela uses MCTS. That too is false. Leela instead uses PUCT.
http://lczero.org/dev/wiki/technical-ex ... hess-zero/

This is a Queen's Indian mainline with 200+ games in my database. The database also confirms equality based upon game scores. So, there's no merit analyzing this position on move 11 because established opening theory runs at least another 15 moves further.

Your first post also generated confusion due to claims of "refutation" when you wrote, "I will start with (sic) recently popular line which might become a practical refutation of classical 1.d4 setups." Where's the refutation? Are you claiming White, or Black "might become a practical refutation"? Instead, the line presented was harmless. As Dann stated, it's surely a draw.

Sincerely,
-Steve-

[Pi4Chess]

@OneTrickPony : are you confused about centipawn scoring ?

[OneTrickPony]

For example, you wrote, "Leela finds refutations to 11.Qa4 and 11.Qf5..." What refutations? Where are they? Are you claiming that 11 Qa4 and 11 Qa5 have been refuted? Or do you mean the opposite, that those two moves refute Black? The confusion is compounded with "...they are completely dead". What?

It's well known for few years now that both are dead ends. 11.Qf5 because of exchange sac (15...Qb8!) and Qa4 because of 17...Qe8!. Leela has found both (which is nice to see as not so long ago it wasn't able to find 15...Qb8 even analyzing a position after white's 15th move and here it sees it from distance.

Either way, all those claims are false as demonstrated by your own Leela output, showing equality in all lines. Basically a score of 52 represents equality.

It doesn't in general but the positions those variations end in are dead equal that's why I've written it's not 50% because of MCTS (it takes a lot of time for it to show 50% even when it's sees a dead position at the end).

Also, you claim Leela uses MCTS. That too is false. Leela instead uses PUCT.

This is a variation of MCTS, c'mon now.

I will start with (sic) recently popular line which might become a practical refutation of classical 1.d4 setups." Where's the refutation? Are you claiming White, or Black "might become a practical refutation"? Instead, the line presented was harmless. As Dann stated, it's surely a draw.

I am claiming that it might become a refutation of classical 1.d4 setups in a sense that it's just too comfortable for black. Obviously when you are "refuting classical 1.d4 setups" it must be a variation for black, not for white. It's the choice of black to go into that semi-Tarrash endgame and white can't do much about once Nc3 is played (there is also Bg5 but this isn't very promising either).

The purpose of this post is to show some very deep engine analysis of popular opening positions. I add some thoughts I have about it but I am assuming rather high level understanding of the history of given lines as well as their current status. It will probably not make much sense if you are not familiar with those.

@OneTrickPony : are you confused about centipawn scoring ?

I am not but I prefer % as I think centi-pawn is an artificial measure which isn't as good for chess.

[OneTrickPony]

Exchagne Grunfeld:

[pgn]1.d4 Nf6 2.c4 g6 3.Nc3 d5 4.cxd5 Nxd5 5.e4 Nxc3 6.bxc3 Bg7 [/pgn]

1. 28 [+54.22] 7.Bc4 c5 8.Ne2 Nc6 9.Be3 O-O 10.O-O b6 11.Rc1 Bb7 12.Bb5 e6 13.dxc5 Ne5 14.cxb6 axb6 15.Qxd8 Rfxd8 16.Bxb6 Rdc8 17.a4 Nc4 18.Bxc4 Rxc4 19.a5 Bxe4 20.Rfd1 h5 21.h3 Bb7 22.Nd4 Ba6 23.Rb1 Rxc3 24.Nb5 Bxb5 25.Rxb5 Ra3 26.Rc1 h4 27.Kf1 Bf6 28.Rbc5 Kg7 29.Rc8 R3xa5 30.Bxa5 Rxa5 31.R8c7 g5 32.Rd7 Be5
2. 28 [+54.16] 7.Be3 c5 8.Rc1 Qa5 9.Qd2 cxd4 10.cxd4 Qxd2+ 11.Bxd2 O-O 12.Nf3 Bg4 13.Be3 Rd8 14.Bb5 a6 15.Ba4 Nc6 16.Bxc6 bxc6 17.Ne5 Bxe5 18.dxe5 Be6 19.a3 Rd3 20.Bc5 Kf8 21.f3 Rb8 22.Kf2 Ke8 23.Kg3 Rb5 24.h4 h6 25.Rhe1 a5 26.a4 Rb2 27.Rb1 Rdb3
3. 28 [+53.93] 7.Nf3 c5 8.h3 O-O 9.Be2 cxd4 10.cxd4 Nc6 11.Be3 Qa5+ 12.Bd2 Qa3 13.d5 Ne5 14.O-O Bd7 15.Rb1 Ba4 16.Qe1 Nxf3+ 17.Bxf3 Bc2 18.Rb4 b6 19.e5 Rac8 20.Bg5 Bd3 21.Rb3 Qa6 22.Rxd3 Qxd3 23.Bxe7 Rfe8 24.d6 Bf8 25.e6 Rxe7 26.dxe7 Bxe7 27.exf7+ Kxf7 28.Bg4 Rd8 29.Qe6+ Kf8 30.Re1 Qd6 31.Qe4 Qc7 32.g3 Bc5 33.Kg2 Kg7 34.Re2 Rd4 35.Qc2 Qd8
4. 28 [+53.82] 7.Qa4+ Nd7 8.Nf3 O-O 9.Be2 Nb6 10.Qd1 c5 11.Be3 Bg4 12.O-O e6 13.a4 Bxf3 14.gxf3 Rc8 15.a5 cxd4 16.cxd4 Nc4 17.Bxc4 Rxc4 18.Ra4 Rc7 19.Kg2 Rd7 20.Qd2 f5 21.Bg5 Qe8 22.Rc4 fxe4
5. 28 [+54.09] 7.h3 O-O 8.Nf3 c5 9.Be2 cxd4 10.cxd4 Nc6 11.Be3 Qa5+ 12.Bd2 Qa3 13.d5 Ne5 14.Rb1 Qxa2 15.O-O Nxf3+ 16.Bxf3 a5 17.Qc1 b6 18.Bg5 Bd4 19.Be3 Be5 20.Rxb6 Ba6 21.Rd1 Rac8 22.Bc5 Be2 23.Bxe2 Qxe2 24.Rc6 Rxc6 25.dxc6 Rc8 26.Bxe7 Qxe4 27.Qc5 a4 28.Bd6 Qe2 29.Rc1 Bb2 30.Qc2 Qxc2 31.Rxc2 a3
6. 28 [+53.87] 7.Rb1 c5 8.Bb5+ Nc6 9.Nf3 O-O 10.d5 Ne5 11.O-O a6 12.Bd3 Nxf3+ 13.Qxf3 Bxc3 14.Be3 Qd6 15.Rfc1 Bd4 16.a4 b5 17.axb5 axb5 18.Bxb5 Ba6 19.Bf4 Be5 20.Bxe5 Qxe5 21.Bc6 Rab8 22.Ra1 Bb5 23.Qc3 Qxe4 24.Qxc5 Rfc8 25.Qa7 Qe5 26.Re1 Qd6 27.Rac1 Bxc6 28.dxc6 Ra8 29.Qb7 Rab8 30.Qxe7

211M nodes, almost 60GB of RAM, 12739 seconds, same JHortos net.

It's interesting that 7.Bc4 is back aa the top choice. Leela doesn't really want to play what Alpha-Zero preferred (7.Nf3 in combination with Be3) as this ends in equal endgame (although black needs to be somewhat precise). 7.Nf3/8.h3 setup is the 3rd choice and I've noticed it recently got some traction among GMs as well.

The endgame Leela wants to play in Be3/Rc1 line is imo interesting from practical point of view. It might not be that easy to hold at human GM level as white will be advancing K-side pawns. Interestingly even Stockfish doesn't show 0.00 there. There is also 13...Nc6 instead of 13...Rd8 and that might be easier to hold for black as the forced line with pawn sac (16..Bh3 f5 17.Kf1 etc.) ends in a draw, at least it did last time I've checked.

[OneTrickPony]

Nimzo Indian:
[pgn]1.d4 Nf6 2.c4 e6 3.Nc3 Bb4 [/pgn]

The deepest I've done so far:

1. 25 [+52.33] 4.Nf3 O-O 5.e3 d5 6.Bd3 dxc4 7.Bxc4 c5 8.a3 Ba5 9.O-O Nc6 10.Bd3 cxd4 11.exd4 h6 12.Bc2 Bc7 13.Re1 Re8 14.h3 e5 15.dxe5 Qxd1 16.Bxd1 Nxe5 17.Nxe5 Rxe5 18.Be3 Bb6 19.Bf3 Bxe3 20.Rxe3 Rxe3 21.fxe3 Bd7 22.Rd1 Rc8 23.e4 Be6 24.e5 Nd7 25.Bxb7 Rb8 26.Bd5 Nxe5 27.b4 Nc4 28.Bxc4 Bxc4 29.Ne4 Bb5 30.Nc5 Rc8 31.a4 Be8 32.Rd5 Kf8 33.b5 Ke7
2. 25 [+52.27] 4.e3 O-O 5.Nf3 d5 6.Bd3 dxc4 7.Bxc4 c5 8.O-O cxd4 9.exd4 b6 10.Qe2 Bb7 11.Bg5 Bxc3 12.bxc3 Nbd7 13.Rac1 Qc7 14.Nd2 Rfc8 15.Bb3 Qc6 16.f3 b5 17.Rfe1 h6 18.Bh4 a5 19.Ne4 Qb6 20.Qd2 Ne8 21.Rb1 Bd5 22.Ba4 Nc7 23.Bg3 Qc6 24.Bc2
3. 25 [+52.15] 4.Qc2 O-O 5.e4 d5 6.e5 Ne4 7.Bd3 c5 8.Nf3 cxd4 9.Nxd4 Nd7 10.Bf4 Qh4 11.g3 Qh3 12.O-O-O Nxc3 13.bxc3 Ba3+ 14.Kb1 Nb6 15.Bf1 Qh5 16.Be2 Qg6 17.Bd3 Qh5 18.h4 dxc4 19.Be4 f5 20.exf6 e5 21.Bf3 Qe8 22.Rhe1 gxf6 23.Bh6 Bd7 24.Bxf8 Bxf8 25.Nf5 Ba4 26.Qe4 Qb5+ 27.Ka1 Bxd1 28.Rxd1 Qa5 29.Qc2 e4 30.Bxe4 Re8 31.Ne3 f5 32.Bd5+ Kh8 33.Bf7 Rc8 34.Be6 Re8 35.Bf7
4. 25 [+51.89] 4.f3 d6 5.e4 c5 6.d5 b5 7.Bd2 Bxc3 8.Bxc3 b4 9.Bd2 O-O 10.Be3 exd5 11.cxd5 Ne8 12.Qd2 Qh4+ 13.Qf2 Qxf2+ 14.Bxf2 Nd7 15.Ne2 a5 16.a3 Ne5 17.Bg3 f5 18.axb4 cxb4 19.exf5 Nd3+ 20.Kd2 Nxb2 21.Nd4 a4 22.Nc6 Rf7 23.Nxb4 Rb7 24.Nc6 a3 25.Be2 Bxf5 26.Ra2 Ra4 27.Rha1 Nc4+
5. 25 [+51.33] 4.Bd2 c5 5.Nf3 cxd4 6.Nxd4 O-O 7.e3 Nc6 8.Bd3 d5 9.cxd5 Nxd4 10.exd4 Nxd5 11.O-O Nf6 12.Bg5 Be7 13.Qf3 Qxd4 14.Rfd1 Qe5 15.Bf4 Qh5 16.Be2 Qxf3 17.Bxf3 a5 18.Rac1 a4 19.a3 Rd8 20.Rxd8+
6. 25 [+50.91] 4.g3 O-O 5.Bg2 d5 6.a3 Bxc3+ 7.bxc3 dxc4 8.Nf3 Nc6 9.O-O Rb8 10.a4 Na5 11.Bg5 h6 12.Bh4 c5 13.e4 e5 14.Bxf6 Qxf6 15.dxe5 Qb6 16.Nh4 Rd8 17.Qh5 Nb3 18.Rad1 Nd2

Lc0 26,3, J96-28 JHortos net, 461M nodes, 125GB of RAM

Leela again recreating known human theory but with some twists which require further research.
The top choices are both classical Rubinstein e3 variations. It would be interesting to see why Leela chooses 8...Ba5 instead of the standard 8...Bxc3. From what I remember it was showing quick full equality in those lines.
4.Qc2 leads to the long forced variations quite popular recently as white discovered some ways to get better endgames there. 11...Qh3 there is considered to be a mistake. It's hard to believe Leela found a refutation of 11...Qh5 so it's likely too tactical for her even at this depth.

[Cornfed]

@OneTrickPony : are you confused about centipawn scoring ?

I am not but I prefer % as I think centi-pawn is an artificial measure which isn't as good for chess.

Ummm...what isn't 'artificial'? The only thing that is not is results without even opening books or table bases.

And if I remember correctly, Chessbase kind of ignores openings and endgames when it comes to their new 'Centipawn analysis'.

[OneTrickPony]

Ummm...what isn't 'artificial'? The only thing that is not is results without even opening books or table bases.

I don't know but "We will score 55% vs equal level opponent from here" sounds better to me than "we are ahead by 0.1 of a pawn".
I don't think it matters much. I like % so I use it. If you like centipawns use that. My understanding is that "centipawns" are scaled to represent % anyway these days (so like +1.0 is 80% or w/e). You may just as well skip that step.

And if I remember correctly, Chessbase kind of ignores openings and endgames when it comes to their new 'Centipawn analysis'.

Interesting but I am not really sure how what Chessbase does with one of their features is relevant to anything.
Why would you ignore openings anyway? There are better and worse moves in openings as well.