Scratch Chess Engine - Game of Kings

ArnoHu

waabooboo wrote:
ArnoHu wrote:
ArnoHu wrote:
waabooboo wrote:
ArnoHu wrote:
ArnoHu wrote:
GoK Classic vs. Wolverine 2, 92% vs. 87%:
https://lichess.org/study/owu9JlSd/P8sC3DOH

GoK NNUE vs. Black Crow, 96% vs. 92%:
https://lichess.org/study/owu9JlSd/ptIAootX

GoK Classic vs. Wolverine 2, 85% vs. 86%:
https://lichess.org/study/owu9JlSd/9tiRz0T2

GoK Classic vs. Wolverine 2, 93% vs. 92%:
https://lichess.org/study/owu9JlSd/dXV0RnpN

GoK Classic vs. White Dove 2, 91% vs. 88%:
https://lichess.org/study/owu9JlSd/o9n8lYYU

Very nice games! The second one GoK is winning with 53. h5, but neither engine knows that rook+minor vs rook is drawn without pawns. Wolverine used to have special knowledge for such endgames, but I scrapped it because it slowed things down…

Seems like a definite improvement for Wolverine – still some strategical misunderstandings, but he showed excellent defense to secure those two draws.

Wolverine's development is very impressive. Reminds me of the old GoK Classic, dominating by search depth. But I had to to change that, as White Dove started winning by superior evaluation. This in turn costed GoK quite some depth, which now is required against Wolverine… an endless cycle.

Seems GoK should play black more often, as that worked better:

GoK Classic vs. Wolverine 2, 96% vs. 92%:
https://lichess.org/study/owu9JlSd/K20Wc4xR

OK, now even a win as white, in 27 moves:

GoK Classic vs. Wolverine 2, 96% vs. 87%:
https://lichess.org/study/owu9JlSd/xkIf6rlk

Yeah, there's no king safety evaluation at all, besides pawn shield/storm… Everything else relies on Wolverine's calculation (check extensions usually resolve the most serious threats).

I haven't tested Wolverine vs White Dove under these conditions, but I feel like Wolverine would be a very heavy favorite at the 5 sec/move time control – the search depth advantage should be insurmountable.

The Nc3/e3 thing as white is a bit strange, but Wolverine's choice of opening often changes.

Yes that would be interesting.

I am still verifying if there are more GoK Classic eval issues, as the logic is very complex, hence error-prone. Search depth is not so far off from Wolverine (often with the exception of one or two moves, the same happened in last game's endgame), and GoK evaluation should pay off more than it does right now. Probably there are several minor bugs that add up. Very difficult to pin-point… I even asked Copilot, which found some evaluation bugs in the past, but not this time.

Last edited by ArnoHu (Oct. 31, 2025 06:39:50)

Destructor_chess

Rha, I don’t know what implementing for my new chess engine: mini NNUE or PSQT… because I ́ll lost the 1.2Mnps I get for legal moves gen into the NNUE if I use NNUE. I dob’t think a big NNUE is adapted to scratch btw

Last edited by Destructor_chess (Oct. 31, 2025 07:28:30)

waabooboo

ArnoHu wrote:
waabooboo wrote:
ArnoHu wrote:
ArnoHu wrote:
waabooboo wrote:
ArnoHu wrote:
ArnoHu wrote:
GoK Classic vs. Wolverine 2, 92% vs. 87%:
https://lichess.org/study/owu9JlSd/P8sC3DOH

GoK NNUE vs. Black Crow, 96% vs. 92%:
https://lichess.org/study/owu9JlSd/ptIAootX

GoK Classic vs. Wolverine 2, 85% vs. 86%:
https://lichess.org/study/owu9JlSd/9tiRz0T2

GoK Classic vs. Wolverine 2, 93% vs. 92%:
https://lichess.org/study/owu9JlSd/dXV0RnpN

GoK Classic vs. White Dove 2, 91% vs. 88%:
https://lichess.org/study/owu9JlSd/o9n8lYYU

Very nice games! The second one GoK is winning with 53. h5, but neither engine knows that rook+minor vs rook is drawn without pawns. Wolverine used to have special knowledge for such endgames, but I scrapped it because it slowed things down…

Seems like a definite improvement for Wolverine – still some strategical misunderstandings, but he showed excellent defense to secure those two draws.

Wolverine's development is very impressive. Reminds me of the old GoK Classic, dominating by search depth. But I had to to change that, as White Dove started winning by superior evaluation. This in turn costed GoK quite some depth, which now is required against Wolverine… an endless cycle.

Seems GoK should play black more often, as that worked better:

GoK Classic vs. Wolverine 2, 96% vs. 92%:
https://lichess.org/study/owu9JlSd/K20Wc4xR

OK, now even a win as white, in 27 moves:

GoK Classic vs. Wolverine 2, 96% vs. 87%:
https://lichess.org/study/owu9JlSd/xkIf6rlk

Yeah, there's no king safety evaluation at all, besides pawn shield/storm… Everything else relies on Wolverine's calculation (check extensions usually resolve the most serious threats).

I haven't tested Wolverine vs White Dove under these conditions, but I feel like Wolverine would be a very heavy favorite at the 5 sec/move time control – the search depth advantage should be insurmountable.

The Nc3/e3 thing as white is a bit strange, but Wolverine's choice of opening often changes.

Yes that would be interesting.

I am still verifying if there are more GoK Classic eval issues, as the logic is very complex, hence error-prone. Search depth is not so far off from Wolverine (often with the exception of one or two moves, the same happened in last game's endgame), and GoK evaluation should pay off more than it does right now. Probably there are several minor bugs that add up. Very difficult to pin-point… I even asked Copilot, which found some evaluation bugs in the past, but not this time.

It seems a bit hard to compare directly just from search depth. GoK still prunes much better than Wolverine, that's why depth is close despite Wolverine's huge speed advantage. But there must be at least some tradeoff because Wolverine is more “thorough” in his calculations.

ArnoHu

waabooboo wrote:
ArnoHu wrote:
ArnoHu wrote:
waabooboo wrote:
ArnoHu wrote:
ArnoHu wrote:
GoK Classic vs. Wolverine 2, 92% vs. 87%:
https://lichess.org/study/owu9JlSd/P8sC3DOH

GoK NNUE vs. Black Crow, 96% vs. 92%:
https://lichess.org/study/owu9JlSd/ptIAootX

GoK Classic vs. Wolverine 2, 85% vs. 86%:
https://lichess.org/study/owu9JlSd/9tiRz0T2

GoK Classic vs. Wolverine 2, 93% vs. 92%:
https://lichess.org/study/owu9JlSd/dXV0RnpN

GoK Classic vs. White Dove 2, 91% vs. 88%:
https://lichess.org/study/owu9JlSd/o9n8lYYU

Very nice games! The second one GoK is winning with 53. h5, but neither engine knows that rook+minor vs rook is drawn without pawns. Wolverine used to have special knowledge for such endgames, but I scrapped it because it slowed things down…

Seems like a definite improvement for Wolverine – still some strategical misunderstandings, but he showed excellent defense to secure those two draws.

Wolverine's development is very impressive. Reminds me of the old GoK Classic, dominating by search depth. But I had to to change that, as White Dove started winning by superior evaluation. This in turn costed GoK quite some depth, which now is required against Wolverine… an endless cycle.

Seems GoK should play black more often, as that worked better:

GoK Classic vs. Wolverine 2, 96% vs. 92%:
https://lichess.org/study/owu9JlSd/K20Wc4xR

OK, now even a win as white, in 27 moves:

GoK Classic vs. Wolverine 2, 96% vs. 87%:
https://lichess.org/study/owu9JlSd/xkIf6rlk

Yeah, there's no king safety evaluation at all, besides pawn shield/storm… Everything else relies on Wolverine's calculation (check extensions usually resolve the most serious threats).

I haven't tested Wolverine vs White Dove under these conditions, but I feel like Wolverine would be a very heavy favorite at the 5 sec/move time control – the search depth advantage should be insurmountable.

The Nc3/e3 thing as white is a bit strange, but Wolverine's choice of opening often changes.

Here is a tournament game between White Dove and Wolverine:
https://lichess.org/broadcast/halloween-2025-scratch-chess-engine-tournament/halloween-2025-scet-round-4/3YmDsfrX/CcwfDdUT

waabooboo

ArnoHu wrote:
waabooboo wrote:
ArnoHu wrote:
ArnoHu wrote:
waabooboo wrote:
ArnoHu wrote:
ArnoHu wrote:
GoK Classic vs. Wolverine 2, 92% vs. 87%:
https://lichess.org/study/owu9JlSd/P8sC3DOH

GoK NNUE vs. Black Crow, 96% vs. 92%:
https://lichess.org/study/owu9JlSd/ptIAootX

GoK Classic vs. Wolverine 2, 85% vs. 86%:
https://lichess.org/study/owu9JlSd/9tiRz0T2

GoK Classic vs. Wolverine 2, 93% vs. 92%:
https://lichess.org/study/owu9JlSd/dXV0RnpN

GoK Classic vs. White Dove 2, 91% vs. 88%:
https://lichess.org/study/owu9JlSd/o9n8lYYU

Very nice games! The second one GoK is winning with 53. h5, but neither engine knows that rook+minor vs rook is drawn without pawns. Wolverine used to have special knowledge for such endgames, but I scrapped it because it slowed things down…

Seems like a definite improvement for Wolverine – still some strategical misunderstandings, but he showed excellent defense to secure those two draws.

Wolverine's development is very impressive. Reminds me of the old GoK Classic, dominating by search depth. But I had to to change that, as White Dove started winning by superior evaluation. This in turn costed GoK quite some depth, which now is required against Wolverine… an endless cycle.

Seems GoK should play black more often, as that worked better:

GoK Classic vs. Wolverine 2, 96% vs. 92%:
https://lichess.org/study/owu9JlSd/K20Wc4xR

OK, now even a win as white, in 27 moves:

GoK Classic vs. Wolverine 2, 96% vs. 87%:
https://lichess.org/study/owu9JlSd/xkIf6rlk

Yeah, there's no king safety evaluation at all, besides pawn shield/storm… Everything else relies on Wolverine's calculation (check extensions usually resolve the most serious threats).

I haven't tested Wolverine vs White Dove under these conditions, but I feel like Wolverine would be a very heavy favorite at the 5 sec/move time control – the search depth advantage should be insurmountable.

The Nc3/e3 thing as white is a bit strange, but Wolverine's choice of opening often changes.

Here is a tournament game between White Dove and Wolverine:
https://lichess.org/broadcast/halloween-2025-scratch-chess-engine-tournament/halloween-2025-scet-round-4/3YmDsfrX/CcwfDdUT

Cool! Fast system helps White Dove for sure, but as expected it's not enough to overcome Wolverine's brutal tactical strength. Strange opening again, but good control by Wolverine.

Destructor_chess

@everyone I found the bug in TK. The corrected version is available on this link: https://experiments.turbowarp.org/faster-lists/1215422188

ArnoHu

Halloween 2025 Scratch Chess Engine Tournament - Results

   Engine           Score

1. GoK NNUE          7.0
2. Black Crow        5.5
3. GoK Classic       4.5
4. Wolverine 2       3.5
5. Shallow Blue 3    3.0
6. Delta             2.5
7. White Dove        1.5
8. TurboKnight       0.5

Results on lichess.org:
https://lichess.org/broadcast/halloween-2025-scratch-chess-engine-tournament/halloween-2025-scet-round-1/9eNvlNVb#players

Last edited by ArnoHu (Nov. 1, 2025 19:45:11)

ArnoHu

waabooboo wrote:
ArnoHu wrote:
waabooboo wrote:
ArnoHu wrote:
ArnoHu wrote:
waabooboo wrote:
ArnoHu wrote:
ArnoHu wrote:
GoK Classic vs. Wolverine 2, 92% vs. 87%:
https://lichess.org/study/owu9JlSd/P8sC3DOH

GoK NNUE vs. Black Crow, 96% vs. 92%:
https://lichess.org/study/owu9JlSd/ptIAootX

GoK Classic vs. Wolverine 2, 85% vs. 86%:
https://lichess.org/study/owu9JlSd/9tiRz0T2

GoK Classic vs. Wolverine 2, 93% vs. 92%:
https://lichess.org/study/owu9JlSd/dXV0RnpN

GoK Classic vs. White Dove 2, 91% vs. 88%:
https://lichess.org/study/owu9JlSd/o9n8lYYU

Very nice games! The second one GoK is winning with 53. h5, but neither engine knows that rook+minor vs rook is drawn without pawns. Wolverine used to have special knowledge for such endgames, but I scrapped it because it slowed things down…

Seems like a definite improvement for Wolverine – still some strategical misunderstandings, but he showed excellent defense to secure those two draws.

Wolverine's development is very impressive. Reminds me of the old GoK Classic, dominating by search depth. But I had to to change that, as White Dove started winning by superior evaluation. This in turn costed GoK quite some depth, which now is required against Wolverine… an endless cycle.

Seems GoK should play black more often, as that worked better:

GoK Classic vs. Wolverine 2, 96% vs. 92%:
https://lichess.org/study/owu9JlSd/K20Wc4xR

OK, now even a win as white, in 27 moves:

GoK Classic vs. Wolverine 2, 96% vs. 87%:
https://lichess.org/study/owu9JlSd/xkIf6rlk

Yeah, there's no king safety evaluation at all, besides pawn shield/storm… Everything else relies on Wolverine's calculation (check extensions usually resolve the most serious threats).

I haven't tested Wolverine vs White Dove under these conditions, but I feel like Wolverine would be a very heavy favorite at the 5 sec/move time control – the search depth advantage should be insurmountable.

The Nc3/e3 thing as white is a bit strange, but Wolverine's choice of opening often changes.

Here is a tournament game between White Dove and Wolverine:
https://lichess.org/broadcast/halloween-2025-scratch-chess-engine-tournament/halloween-2025-scet-round-4/3YmDsfrX/CcwfDdUT

Cool! Fast system helps White Dove for sure, but as expected it's not enough to overcome Wolverine's brutal tactical strength. Strange opening again, but good control by Wolverine.

White Dove had stability issues in the last couple of days, also during recent tournament.

Destructor_chess

TurboKnight too at the beginning of the tournament (until round 6)

HasiLover

I am running a new Tournament, I call it the Supersonic Scratch 2025, because of it's very fast think time(0.5 sec).
The format is as follows:
-Every match is played in a pair, meaning 2 games instead of 1, between 2 engines
-the first two rounds are a normal swiss
-the rest of the rounds, the two worst performing engines will be eliminated from the tournament.
-the last remaining engine wins.
https://lichess.org/broadcast/supersonic-scratch-2025/round-1/aAphgRkm#boards

ArnoHu

HasiLover wrote:
I am running a new Tournament, I call it the Supersonic Scratch 2025, because of it's very fast think time(0.5 sec).
The format is as follows:
-Every match is played in a pair, meaning 2 games instead of 1, between 2 engines
-the first two rounds are a normal swiss
-the rest of the rounds, the two worst performing engines will be eliminated from the tournament.
-the last remaining engine wins.
https://lichess.org/broadcast/supersonic-scratch-2025/round-1/aAphgRkm#boards

OK, the GoK's might face a challenge then, as they initialize larger lists on search init, which might take up to 0.1sec on moderate systems.

waabooboo

HasiLover wrote:
I am running a new Tournament, I call it the Supersonic Scratch 2025, because of it's very fast think time(0.5 sec).
The format is as follows:
-Every match is played in a pair, meaning 2 games instead of 1, between 2 engines
-the first two rounds are a normal swiss
-the rest of the rounds, the two worst performing engines will be eliminated from the tournament.
-the last remaining engine wins.
https://lichess.org/broadcast/supersonic-scratch-2025/round-1/aAphgRkm#boards

You're changing my code in some way? I only have options for think time in units of seconds. Also, Wolverine sees and prevents that mate in 2 in about 1/30 of a second for me…

Destructor_chess

waabooboo wrote:
HasiLover wrote:
I am running a new Tournament, I call it the Supersonic Scratch 2025, because of it's very fast think time(0.5 sec).
The format is as follows:
-Every match is played in a pair, meaning 2 games instead of 1, between 2 engines
-the first two rounds are a normal swiss
-the rest of the rounds, the two worst performing engines will be eliminated from the tournament.
-the last remaining engine wins.
https://lichess.org/broadcast/supersonic-scratch-2025/round-1/aAphgRkm#boards

You're changing my code in some way? I only have options for think time in units of seconds. Also, Wolverine sees and prevents that mate in 2 in about 1/30 of a second for me…

Changing a slider range to make it counting hundredth to get a 0.5s think time isn't modifying the code, sorry

waabooboo

Destructor_chess wrote:
waabooboo wrote:
HasiLover wrote:
I am running a new Tournament, I call it the Supersonic Scratch 2025, because of it's very fast think time(0.5 sec).
The format is as follows:
-Every match is played in a pair, meaning 2 games instead of 1, between 2 engines
-the first two rounds are a normal swiss
-the rest of the rounds, the two worst performing engines will be eliminated from the tournament.
-the last remaining engine wins.
https://lichess.org/broadcast/supersonic-scratch-2025/round-1/aAphgRkm#boards

You're changing my code in some way? I only have options for think time in units of seconds. Also, Wolverine sees and prevents that mate in 2 in about 1/30 of a second for me…

Changing a slider range to make it counting hundredth to get a 0.5s think time isn't modifying the code, sorry

Well it doesn't check the time on half second intervals, because I wasn't anticipating someone doing this. Not that it necessarily helps or hurts

ArnoHu

GoK Classic vs. Wolverine 2, 94% vs. 88%:
https://lichess.org/study/owu9JlSd/FJaNkCuO

GoK Classic vs. Wolverine 2, 87% vs. 85%:
https://lichess.org/study/owu9JlSd/MlgJ9fJL

GoK NNUE vs. Black Crow, 93% vs. 86%:
https://lichess.org/study/owu9JlSd/BxCtMfBt

Last edited by ArnoHu (Nov. 2, 2025 10:01:39)

Destructor_chess

Wolverine 2 vs. TK, 86% vs 88%, 1/2 - 1/2
https://lichess.org/study/5IpBFQUb/y8HxGgyu
I hate this bug

TK vs. SB3, 96% vs. 91%, 1-0
https://lichess.org/study/5IpBFQUb/HWNgERmY

Amarillo vs. TK, 87% vs. 95%, 0-1
https://lichess.org/study/5IpBFQUb/IK0YwWHz

Last edited by Destructor_chess (Nov. 2, 2025 13:49:29)

ArnoHu

ArnoHu wrote:
GoK Classic vs. Wolverine 2, 94% vs. 88%:
https://lichess.org/study/owu9JlSd/FJaNkCuO

GoK Classic vs. Wolverine 2, 87% vs. 85%:
https://lichess.org/study/owu9JlSd/MlgJ9fJL

GoK NNUE vs. Black Crow, 93% vs. 86%:
https://lichess.org/study/owu9JlSd/BxCtMfBt

GoK Classic vs. Wolverine 2, 93% vs. 86%:
https://lichess.org/study/owu9JlSd/WRM9Acxf

GoK NNUE vs. Black Crow, 92% vs. 88%:
https://lichess.org/study/owu9JlSd/4PyIlAsZ

coooolboy55

What is the effective branching factor of your engines? I want to know what I am aiming for when I get to making my chess bot (currently working on the chess program itself).

ArnoHu

ArnoHu wrote:
ArnoHu wrote:
GoK Classic vs. Wolverine 2, 94% vs. 88%:
https://lichess.org/study/owu9JlSd/FJaNkCuO

GoK Classic vs. Wolverine 2, 87% vs. 85%:
https://lichess.org/study/owu9JlSd/MlgJ9fJL

GoK NNUE vs. Black Crow, 93% vs. 86%:
https://lichess.org/study/owu9JlSd/BxCtMfBt

GoK Classic vs. Wolverine 2, 93% vs. 86%:
https://lichess.org/study/owu9JlSd/WRM9Acxf

GoK NNUE vs. Black Crow, 92% vs. 88%:
https://lichess.org/study/owu9JlSd/4PyIlAsZ

This is certainly a pinnacle point for Scratch HCE engines (at 5sec think time):

GoK Classic vs. White Dove, 99% vs. 99%:
https://lichess.org/study/owu9JlSd/3I15lM22

Last edited by ArnoHu (Nov. 2, 2025 19:46:59)

HasiLover

Running Round 2 of my current Tournament now: https://lichess.org/broadcast/supersonic-scratch-2025/round-2/lf4rnuHF/IjqOFdwe

Discuss Scratch