Stockfish can now win KRK ending?

[Jouni]

Latest prerelease can mate with rook :

Using ./stockfish_fix1 on KRvK_1000pv.epd with --nodes 10000000 --mate 0
Engine ID: Stockfish dev-20260505-1429d98e
Total FENs: 1000
Found mates: 1000

[towforce]

Checkmating a lone king with a rook is easy. Are you saying the current version of Stockfish cannot do that? That doesn't make sense!

[Paloma]

Current dev version: 20260517-f8aa78e0
Previously:
dev-20260510-dd321af5
So is dev-20260505 somewhat older

[peter]

There wasn't any 260505 at site of (pre)releases

https://github.com/official-stockfish/S ... h/releases

, 260504 was followed by 260510, regards

[Eelco de Groot]

Checkmating a lone king with a rook is easy. Are you saying the current version of Stockfish cannot do that? That doesn't make sense!

To provide some context for even more lazy or clueless people than we , simply Googling Jouni's text; It had, at least according to the developers something to do with aging of hashtable entries, secondary aging was thrown out and that did not mix totally with latest tunes and extensions. So the aging was reintroduced. Of course the tests are with low search depths or in this case limited nodes. I am just glancing over this reported SF issue, probably more on Discord: https://github.com/official-stockfish/S ... ssues/6742


Last entries on GitHub read:

A recent test event at TCEC had Reckless up against SF15-1G (fixed nodes player), and perhaps remarkably SF15-1G appeared to show considerably better "matefinding" ability than Reckless did despite a considerable time advantage to Reckless: https://tcec-chess.com/#div=fdts&game=3&season=29

Ciecke also commented "phrax has started doing the same thing after recent extension patches and vltc tunes"

vondele
added a commit that references this issue 14 hours ago

Reintroduce Secondary TT Aging
94beadf

vondele
closed this as completedin 94beadf14 hours ago

[syzygy]

Latest prerelease can mate with rook :

That's one small step for mankind, one giant leap for Stockfish.

[towforce]

Checkmating a lone king with a rook is easy. Are you saying the current version of Stockfish cannot do that? That doesn't make sense!

To provide some context for even more lazy or clueless people than we , simply Googling Jouni's text; It had, at least according to the developers something to do with aging of hashtable entries, secondary aging was thrown out and that did not mix totally with latest tunes and extensions. So the aging was reintroduced. Of course the tests are with low search depths or in this case limited nodes. I am just glancing over this reported SF issue, probably more on Discord: https://github.com/official-stockfish/S ... ssues/6742


Last entries on GitHub read:

A recent test event at TCEC had Reckless up against SF15-1G (fixed nodes player), and perhaps remarkably SF15-1G appeared to show considerably better "matefinding" ability than Reckless did despite a considerable time advantage to Reckless: https://tcec-chess.com/#div=fdts&game=3&season=29

Ciecke also commented "phrax has started doing the same thing after recent extension patches and vltc tunes"

vondele
added a commit that references this issue 14 hours ago

Reintroduce Secondary TT Aging
94beadf

vondele
closed this as completedin 94beadf14 hours ago

Thank you - that was helpful. I had assumed that I'd misunderstood something, or that something had been mistyped - but yes - I could have searched.

[Jouni]

Mating with tablebases is still not fixed! https://tcec-chess.com/#div=frdfl&game=51&season=29 . https://github.com/official-stockfish/S ... /pull/6843 . Really a lot problems to +2 elo TB gain .

[Uri Blass]

Mating with tablebases is still not fixed! https://tcec-chess.com/#div=frdfl&game=51&season=29 . https://github.com/official-stockfish/S ... /pull/6843 . Really a lot problems to +2 elo TB gain .

I see that stockfish could win that game and I thought they do not care about the shortest mate because they do not care about the fact that stockfish needed more moves to mate in my queen handicap matches in the past.

See the following thread.
viewtopic.php?p=987334&hilit=Wasp#p987334
If they do not care about the fact that Wasp needed less moves to mate in my queen handicap match against Stockfish I see no reason for them to care if stockfish does not play the optimal moves in tablebases positions when stockfish finally won the game.

Note that I used old Wasp that is even not NNUE and I think that it is reasonable to expect every engine that is more than 300 elo stronger than it to mate faster or in the same number of moves in a competition against it(I know that it does not happen not only with stockfish but if programmer consider not finding the best tablebase move as a bug they should also consider the games that I posted as a bug by the same logic(and note that I did not search special positions and special time control to make it happen and it is going to happen also with a different handicap and a different time control).

CCRL rating list:
Stockfish 18 64-bit 4CPU 3651 +14 −14 57.4% −46.0 84.6% 1017
Wasp 4.50 64-bit 4CPU 3282 +18 −18 46.4% +23.5 60.6% 634

[Uri Blass]

I can add that the solution can be simple to change the way they test stockfish so patches that help to mate faster with no reduction in playing strength are going to pass.

I suggest simply to change the scoring of the winner to be less than 1 in SPRT tests.

score can be 0.9+0.1*0.99^number of plies when if you are afraid of losing elo then after a patch pass SPRT you also verify that you had no reduction in playing strength when you consider winning as 1 and not by this formula(in part of the cases it can be verified without more games because with the same score you could pass no reduction SPRT test and in some other cases you may need additional games to show no reduction in playing strength with the normal formula(the same you do for simplifications).