Scratch Chess Engine - Game of Kings

ArnoHu

New GoK versions have just been uploaded to https://scratch.mit.edu/projects/148769358/ and https://scratch.mit.edu/projects/828094886/

Test game results:

GoK Classic - W3 NNUE Leorik256 (1sec): 7.5 - 2.5
GoK Classic - W3 NNUE Leorik128 (1sec): 6.5 - 3.5
GoK Classic - White Dove (1sec): 7.5 - 2.5
GoK Classic - Black Crow (1sec): 5 - 5 (including an invalid opening by BC: https://lichess.org/study/Tv2tcUi7/CbSoO0jO#12 )

GoK NNUE - Black Crow (1sec): 8.5 - 1.5

Last edited by ArnoHu (Oct. 24, 2026 21:46:09)

ArnoHu

waabooboo wrote:
ArnoHu wrote:
ArnoHu wrote:
waabooboo wrote:
https://scratch.mit.edu/projects/1202302970

New Wolverine – some search improvements, and the little Leorik net has been replaced with the bigger one.

Cool, I will run some test series later.

Hmmm, it doesn't have DCI integration, does it?

Not yet, I'll make a DCI version later…

I merged the old W3 DCI logic into the new version, to me it the little Leorik net still seems to be slightly better. Which would be surprising given that the new net reaches similar search depth. Maybe there is different reason. Anyway, currently running a longer series for both to go sure.

Was also curious how the new GoK Classic version would do against W3 NNUE, as I found and fixed a crucial regression that happened some weeks ago (safe check threats applied wrong attack table data), among other improvements. I can say that much, Classic has reached W3 NNUE small net strength, and is currently at 52 - 48 (1sec).

Last edited by ArnoHu (Oct. 25, 2026 03:51:18)

waabooboo

ArnoHu wrote:
waabooboo wrote:
ArnoHu wrote:
ArnoHu wrote:
waabooboo wrote:
https://scratch.mit.edu/projects/1202302970

New Wolverine – some search improvements, and the little Leorik net has been replaced with the bigger one.

Cool, I will run some test series later.

Hmmm, it doesn't have DCI integration, does it?

Not yet, I'll make a DCI version later…

I merged the old W3 DCI logic into the new version, to me it the little Leorik net still seems to be slightly better. Which would be surprising given that the new net reaches similar search depth. Maybe there is different reason. Anyway, currently running a longer series for both to go sure.

Was also curious how the new GoK Classic version would do against W3 NNUE, as I found and fixed a crucial regression that happened some weeks ago (safe check threats applied wrong attack table data), among other improvements. I can say that much, Classic has reached W3 NNUE small net strength, and is currently at 52 - 48 (1sec).

Seems GoK Classic has improved quite a bit recently. Still I'd be surprised if it can beat the bigger net, but who knows… NPS is like 30% less with the bigger net, I'd expect maybe one ply less on average?

I wanted to change the DCI integration logic, that's why I haven't made the DCI version yet. Wolverine still needs a list of previous board states in DCI mode, without that it will allow draws by repetition in winning positions. I'm not sure how many games out of 100 that costs it, but it's probably at least a few – so that's on my to-do list.

ArnoHu

waabooboo wrote:
ArnoHu wrote:
waabooboo wrote:
ArnoHu wrote:
ArnoHu wrote:
waabooboo wrote:
https://scratch.mit.edu/projects/1202302970

New Wolverine – some search improvements, and the little Leorik net has been replaced with the bigger one.

Cool, I will run some test series later.

Hmmm, it doesn't have DCI integration, does it?

Not yet, I'll make a DCI version later…

I merged the old W3 DCI logic into the new version, to me it the little Leorik net still seems to be slightly better. Which would be surprising given that the new net reaches similar search depth. Maybe there is different reason. Anyway, currently running a longer series for both to go sure.

Was also curious how the new GoK Classic version would do against W3 NNUE, as I found and fixed a crucial regression that happened some weeks ago (safe check threats applied wrong attack table data), among other improvements. I can say that much, Classic has reached W3 NNUE small net strength, and is currently at 52 - 48 (1sec).

Seems GoK Classic has improved quite a bit recently. Still I'd be surprised if it can beat the bigger net, but who knows… NPS is like 30% less with the bigger net, I'd expect maybe one ply less on average?

I wanted to change the DCI integration logic, that's why I haven't made the DCI version yet. Wolverine still needs a list of previous board states in DCI mode, without that it will allow draws by repetition in winning positions. I'm not sure how many games out of 100 that costs it, but it's probably at least a few – so that's on my to-do list.

There was one such game among the 100. Well, technically they are at a draw either way.

I found out why Leorik256 seemed weaker. First of all the sample was too small. And more importantly: DCI integration claimed draws after 100 moves. Re-running now, and it indeed seems stronger than Leorik128 (and GoK Classic). It leads 11 - 9 right now. I will rather let it play against Black Crow and GoK NNUE.

Update:
GoK Classic - W3 NNUE Leorik256 (1sec): 30.5 - 29.5

GoK Classic - White Dove (3sec): 14 - 1
GoK NNUE - Black Crow (3sec): 12 - 3

Last edited by ArnoHu (Oct. 25, 2026 11:11:48)

Discuss Scratch