Fortress Positions

[chrisw]

So, that comes to:

Fewer than 16 men: 163,493 positions
Screened: drawn game and |V| > 100: 19,805
Distinct episodes (consecutive screened positions from one game collapsed to one) 6,326

Sampling 240 of those and applying SF screening, 60 passed the screen.
Removing "drawn for other reasons", 24 were left

Over to humans for opinions of whether these could be fortresses or not (I haven't looked)

[peter]

Nice find and awesome problem!

Knew, you'd like it

How about this one?
[pgn][Event "Probleemblad 2023-2024"]
[Site "; Probleemblad#545 no.2 4-6/2"]
[Date "2023.??.??"]
[Round "?"]
[White "Dreelinck=J"]
[Black "(=0230.35f1c1)"]
[Result "1/2-1/2"]
[Annotator "SF19"]
[SetUp "1"]
[FEN "8/6R1/3b2p1/3R2P1/6P1/1p4p1/1p3pP1/2k2K2 w - - 0 1"]
[PlyCount "33"]
[GameId "2363031560606168"]
[EventDate "2023.??.??"]

{source: Probleemblad#545 no.2 4-6/2023} 1. Rb7 $1 (1. Ke2 $2 f1=Q+) (1. Rxd6 $2 b1=Q 2. Rgd7 Qe4 $1 3. Rd1+ (3. Rc7+ Kb2 4. Rd2+ Ka3 5. Ra7+ Kb4) 3... Kb2 4. R7d2+ Ka3 $1 5. Re2 Qb7 6. Ra1+ Kb4 7. Re4+ Kc3 $1 8. Re3+ Kb2 9. Rd1 Qb5+ 10. Re2+ Kc3 11. Rc1+ Kb4 12. Rb1 Qc6 $1 (12... Qc4 $2 13. Rxb3+ $1 Qxb3 (13... Kxb3) 14. Rb2 $1 Qxb2) 13. Rd1 Qc4 14. Rb1 Kc5 $1) (1. Rxg6 $2 b1=Q 2. Rgxd6 Kc2+ 3. Rd1 Qxd1+ 4. Rxd1 Kxd1 5. g6 b2 6. g7 b1=Q 7. g8=Q Kd2#) (1. Rc7+ $2 Bxc7 2. Rd1+ Kc2 $1 (2... Kxd1 $2) 3. Rd2+ Kc3 4. Rd3+ Kb4 5. Rd1 Ka3) (1. Ra7 $2 b1=Q 2. Ra1 Qxa1 3. Rd1+ Kb2) 1... b1=Q (1... Kc2 2. Rxb3 Kxb3 3. Rd1 Ka2 4. Rd2) 2. Rd1+ $1 (2. Rxb3 $2 Qxb3) 2... Kxd1 3. Rxb3 Qe4 (3... Qc2 4. Rb1+ $1 Kd2 (4... Qc1 5. Rxc1+ Kxc1 6. Ke2) 5. Rb2 $1) (3... Qa1 4. Rb1+ $1 Qxb1) (3... Qc1 4. Rb1 $1 (4. Rd3+ $2 Kc2+ 5. Ke2 f1=Q#) 4... Kc2+ (4... Qxb1) 5. Rxc1+ Kxc1) (3... Qa2 4. Rd3+ $1 (4. Rb1+ $2 Kc2 $1 5. Rc1+ Kb2) 4... Kc1 5. Rd1+ $1 Kb2 6. Rd2+ Kb3 7. Rxa2 Kxa2 8. Ke2) (3... Kc2+ 4. Rxb1 Kxb1 5. Ke2) 4. Rb1+ Kd2 (4... Kc2 5. Rb2+) 5. Rb2+ (5. Rd1+ $2 Ke3 $1) 5... Kc3 (5... Kd3 6. Rb3+) (5... Ke3 6. Re2+ Kf4 7. Rxe4+ Kxe4 8. Ke2) 6. Rb3+ $1 (6. Rc2+ $2 Kb4 7. Rb2+ Ka4 8. Ra2+ Ba3 $1 9. Rxa3+ Kb5 10. Ra5+ (10. Rb3+ Kc5 11. Rc3+ Kd5 12. Rd3+ Qxd3#) 10... Kc6 11. Ra6+ Kb7 12. Ra7+ Kb8) 6... Kc4 7. Rb4+ Kd5 8. Rxe4 Kxe4 9. Ke2 Kf4 {<eg>} 10. Kf1 Kxg4 11. Ke2 Kxg5 12. Kf1 Kf4 13. Ke2 Bc5 14. Kf1 Bb6 15. Ke2 f1=Q+ 16. Kxf1 Bf2 17. Ke2 1/2-1/2
[/pgn]

[towforce]

A suggestion for automation:

* with the help of a chatbot, write a chess database query that might find blocked positions

Repeat:

- get 10 matches to that query

- for each one that's not suitable, explain to the chatbot why it's not and have it amend the query to filter out that issue

Until all 10 found are suitable positions.

Use that final query to get a large number of suitable positions.

And then make the chatbot learn the difference between a blocked position and a game deciding fortress and make it use Stockfish to judge, if the remaining found fortresses are hard enough to be not too easy and not too difficult for test positions, especially if they are single best move ones

Seriously, I don't say, AI wouldn't be able to help here, I just doubt, it would spare much work of user's own, finding not again much to much positions out of bigger database that you have to judge with engines' output then yet still by yourself.
Fortresses I easily found 400 in HHdb out of first 1000 positions, 390 out of second 1000, 1 was usable as test position out of first 400, 4 out of second 390, so...
I'll have a talk with Claude about it, regards

I was actually thinking about having the chatbot help writing a complex database query rather than doing the actual deciding (though Chris's idea of training an NN to decide quickly might have merit). If it could be done via a database query, that would handily resolve the large numbers issue - but reading this thread, it might not be feasible:

1) with chess positions in a database, you could easily index metrics like numbers of each type of piece, but other aspects of the position would be much more difficult to index

2) in chess, a tiny difference in the position of a piece can make all the difference to the character of a position - like whether it's a fortress or not

So I'm going to modify the proposal: have the chatbot write a fast "hand-coded" type eval (yes - calling a chatbot's output "hand coded" IS contradictory ), have this program loop through the positions in the db, pick out a small sample, tell the chatbot about the ones that are wrong, and have it update the eval. Problem (2) above still applies

Some starter criteria:

1. No Pawn Breakthroughs: The attacking side cannot force a useful pawn breakthrough to create new passed pawns

2. Blocked Pawns: If the stronger side possesses pawns, they are firmly blockaded and unable to advance

3. No King Infiltration: The attacking king is completely cut off from entering the defender's territory or is safely trapped near the edge of the board

4. Immunity to Zugzwang: The defending side has plenty of passive "waiting moves" available

Obviously an NN trained against a set of good examples would be better, but then there's a chicken and egg situation*: you can't sample the data until you have a trained NN, and you can't train the NN until you have a large sample of good fortresses.

*science has resolved this question: something that wasn't quite a chicken laid a chicken egg

[Ajedrecista]

Hello Peter:

Knew, you'd like it

How about this one?
[pgn][Event "Probleemblad 2023-2024"]
[Site "; Probleemblad#545 no.2 4-6/2"]
[Date "2023.??.??"]
[Round "?"]
[White "Dreelinck=J"]
[Black "(=0230.35f1c1)"]
[Result "1/2-1/2"]
[Annotator "SF19"]
[SetUp "1"]
[FEN "8/6R1/3b2p1/3R2P1/6P1/1p4p1/1p3pP1/2k2K2 w - - 0 1"]
[PlyCount "33"]
[GameId "2363031560606168"]
[EventDate "2023.??.??"]

{source: Probleemblad#545 no.2 4-6/2023} 1. Rb7 $1 (1. Ke2 $2 f1=Q+) (1. Rxd6 $2 b1=Q 2. Rgd7 Qe4 $1 3. Rd1+ (3. Rc7+ Kb2 4. Rd2+ Ka3 5. Ra7+ Kb4) 3... Kb2 4. R7d2+ Ka3 $1 5. Re2 Qb7 6. Ra1+ Kb4 7. Re4+ Kc3 $1 8. Re3+ Kb2 9. Rd1 Qb5+ 10. Re2+ Kc3 11. Rc1+ Kb4 12. Rb1 Qc6 $1 (12... Qc4 $2 13. Rxb3+ $1 Qxb3 (13... Kxb3) 14. Rb2 $1 Qxb2) 13. Rd1 Qc4 14. Rb1 Kc5 $1) (1. Rxg6 $2 b1=Q 2. Rgxd6 Kc2+ 3. Rd1 Qxd1+ 4. Rxd1 Kxd1 5. g6 b2 6. g7 b1=Q 7. g8=Q Kd2#) (1. Rc7+ $2 Bxc7 2. Rd1+ Kc2 $1 (2... Kxd1 $2) 3. Rd2+ Kc3 4. Rd3+ Kb4 5. Rd1 Ka3) (1. Ra7 $2 b1=Q 2. Ra1 Qxa1 3. Rd1+ Kb2) 1... b1=Q (1... Kc2 2. Rxb3 Kxb3 3. Rd1 Ka2 4. Rd2) 2. Rd1+ $1 (2. Rxb3 $2 Qxb3) 2... Kxd1 3. Rxb3 Qe4 (3... Qc2 4. Rb1+ $1 Kd2 (4... Qc1 5. Rxc1+ Kxc1 6. Ke2) 5. Rb2 $1) (3... Qa1 4. Rb1+ $1 Qxb1) (3... Qc1 4. Rb1 $1 (4. Rd3+ $2 Kc2+ 5. Ke2 f1=Q#) 4... Kc2+ (4... Qxb1) 5. Rxc1+ Kxc1) (3... Qa2 4. Rd3+ $1 (4. Rb1+ $2 Kc2 $1 5. Rc1+ Kb2) 4... Kc1 5. Rd1+ $1 Kb2 6. Rd2+ Kb3 7. Rxa2 Kxa2 8. Ke2) (3... Kc2+ 4. Rxb1 Kxb1 5. Ke2) 4. Rb1+ Kd2 (4... Kc2 5. Rb2+) 5. Rb2+ (5. Rd1+ $2 Ke3 $1) 5... Kc3 (5... Kd3 6. Rb3+) (5... Ke3 6. Re2+ Kf4 7. Rxe4+ Kxe4 8. Ke2) 6. Rb3+ $1 (6. Rc2+ $2 Kb4 7. Rb2+ Ka4 8. Ra2+ Ba3 $1 9. Rxa3+ Kb5 10. Ra5+ (10. Rb3+ Kc5 11. Rc3+ Kd5 12. Rd3+ Qxd3#) 10... Kc6 11. Ra6+ Kb7 12. Ra7+ Kb8) 6... Kc4 7. Rb4+ Kd5 8. Rxe4 Kxe4 9. Ke2 Kf4 {<eg>} 10. Kf1 Kxg4 11. Ke2 Kxg5 12. Kf1 Kf4 13. Ke2 Bc5 14. Kf1 Bb6 15. Ke2 f1=Q+ 16. Kxf1 Bf2 17. Ke2 1/2-1/2
[/pgn]

I liked it too, unsurprisingly! I have said many times that I enjoy how people can be so creative. Look at SF 19 with Multi-PV disabled to be deterministic, but with access to the full set of 3, 4 and 5-man Syzygy EGTB in the final position of the PGN (6 pieces):

Code: Select all

FEN: 8/8/6p1/8/5k2/6p1/4KbP1/8 b - - 2 17 

Stockfish-19-windows-x86-64-universal:
Found 145 WDL and 145 DTZ tablebase files (up to 5-man).
Available processors: 0-3
Using 1 thread
NNUE evaluation using nn-1a298aa575a0.nnue (109MiB, (86896, 1024, 32, 32, 1))
Network replica 1: Shared memory.
[...]
 64/77	01:18	 112.060.108	1.457.502	+0,74	Kf4e4 Ke2f1 g6g5 Kf1e2 Bf2c5 Ke2f1 Ke4d3 Kf1e1 Bc5e3 Ke1f1 g5g4 Kf1e1 Be3c5 Ke1f1 Bc5a7 Kf1e1 Ba7e3 Ke1f1 Be3b6 Kf1e1 Bb6a5+ Ke1f1 Kd3e3 Kf1g1 Ke3e2 Kg1h1 Ba5c7 Kh1g1 Ke2e1 Kg1h1 Bc7a5 Kh1g1 Ba5d2 Kg1h1 Bd2c3 Kh1g1 Bc3b4 Kg1h1 Ke1d2 Kh1g1 Kd2e2 Kg1h1 Ke2e1 Kh1g1 Bb4a3 Kg1h1 Ke1e2 Kh1g1 Ba3c5+ Kg1h1 Bc5e7 Kh1g1 Be7a3 Kg1h1 Ke2e1 Kh1g1 Ke1d2 Kg1h1 Ba3b2 Kh1g1 Kd2e2 Kg1h1 Bb2a1 Kh1g1 Ke2e1 Kg1h1 Ba1g7 Kh1g1 Bg7f6
 65/105	01:26	 123.732.628	1.469.962	+0,74	g6g5 Ke2d3 Bf2c5 Kd3d2 Kf4e4 Kd2e2 Bc5e3 Ke2f1 Ke4d3 Kf1e1 Be3d4 Ke1f1 Bd4b6 Kf1e1 Bb6c7 Ke1f1 Bc7a5 Kf1g1 Kd3e2 Kg1h1 Ba5b4 Kh1g1 Ke2e1 Kg1h1 Bb4c3 Kh1g1 Ke1e2 Kg1h1 Bc3a5 Kh1g1 Ba5b4 Kg1h1 Bb4d2 Kh1g1 Bd2f4 Kg1h1 Bf4c7 Kh1g1 Ke2e1 Kg1h1 Bc7b8 Kh1g1 Bb8d6 Kg1h1 Bd6c7 Kh1g1 Ke1e2 Kg1h1 Bc7d8 Kh1g1 Ke2e1 Kg1h1 Bd8e7 Kh1g1 Ke1e2 Kg1h1 Ke2d3 Kh1g1 Be7d6 Kg1f1 Bd6c5 Kf1e1 Bc5b4+ Ke1f1 Bb4d6 Kf1e1 Bd6e5 Ke1f1 Kd3d2 Kf1g1
 66/105+	01:30	 130.631.806	1.472.488	+0,75	Kf4e4
 66/105	01:32	 134.468.485	1.476.410	+0,74	Kf4e4 Ke2f1 Bf2e3 Kf1e2 Ke4d4 Ke2e1 Kd4d3 Ke1f1 Be3a7 Kf1e1 g6g5 Ke1f1 Ba7b6 Kf1e1 Kd3e3 Ke1f1 Ke3e4 Kf1e2 Ke4d4 Ke2d2 Bb6c5 Kd2d1 Kd4d3 Kd1e1 Bc5d4 Ke1f1 Kd3c2 Kf1e2 Bd4b6 Ke2f1 Bb6c5 Kf1e2 Bc5b4 Ke2f1 Kc2d3 Kf1g1 Kd3e2 Kg1h1 Bb4c3 Kh1g1 Ke2e1 Kg1h1 Bc3a5 Kh1g1 Ba5b4 Kg1h1 Bb4d2 Kh1g1 Bd2a5 Kg1h1 Ke1d1 Kh1g1 Kd1e2 Kg1h1 Ba5b4 Kh1g1 Bb4c3 Kg1h1 Bc3d2 Kh1g1 Ke2e1 Kg1h1 Bd2b4
 67/122	01:49	 163.338.550	1.515.720	+0,74	Kf4e4 Ke2f1 Bf2e3 Kf1e2 Ke4d4 Ke2e1 Kd4d3 Ke1f1 Be3a7 Kf1e1 g6g5 Ke1f1 Ba7c5 Kf1e1 Kd3e4 Ke1d2 Bc5b4+ Kd2e2 Bb4c3 Ke2f1 Ke4e3 Kf1g1 Ke3e2 Kg1h1 Bc3h8 Kh1g1 Ke2e1 Kg1h1 Bh8b2 Kh1g1 Bb2d4+ Kg1h1 Bd4g7 Kh1g1 Bg7c3 Kg1h1 Bc3a5 Kh1g1 g5g4 Kg1h1 Ba5d2 Kh1g1 Bd2h6 Kg1h1 Ke1e2 Kh1g1 Bh6c1 Kg1h1 Ke2d1 Kh1g1 Bc1e3+ Kg1f1 Kd1c2 Kf1e2 Be3d2 Ke2f1 Kc2d3 Kf1g1 Kd3e2 Kg1h1 Bd2c1 Kh1g1 Bc1e3+ Kg1h1 Be3f4 Kh1g1 Ke2e1 Kg1h1 Bf4e5 Kh1g1 Be5c3 Kg1h1 Bc3a5
 68/122	01:51	 165.873.864	1.518.212	+0,74	Kf4e4 Ke2f1 Bf2e3 Kf1e2 Ke4d4 Ke2e1 Kd4d3 Ke1f1 Be3a7 Kf1e1 g6g5 Ke1f1 Ba7c5 Kf1e1 Bc5d6 Ke1f1 Bd6e5 Kf1g1 Kd3e2 Kg1h1 Be5b8 Kh1g1 Ke2e1 Kg1h1 Ke1d2 Kh1g1 Bb8f4 Kg1h1 Bf4e5 Kh1g1 Kd2e1 Kg1h1 Be5c3 Kh1g1 Bc3b2 Kg1h1 Ke1e2 Kh1g1 Bb2c3 Kg1h1 Bc3d2 Kh1g1 Bd2f4 Kg1h1 Ke2e3 Kh1g1 Bf4d6 Kg1f1 Bd6e5 Kf1e1 Ke3d3 Ke1f1 Be5d4 Kf1e1 Bd4e3 Ke1f1 Be3b6 Kf1e1 Bb6d4 Ke1f1 Bd4e3 Kf1e1 Be3f2+ Ke1f1 Bf2a7 Kf1e1 Ba7c5 Ke1f1 Bc5d6 Kf1g1 Bd6f8 Kg1f1

There were 8,833 TB hits and SF insits to see a plus eval for black without any advances. I guess that 0.74 falls in the error margin in this kind of endgames.

Regards from Spain.

Ajedrecista.

[peter]

Code: Select all

FEN: 8/8/6p1/8/5k2/6p1/4KbP1/8 b - - 2 17 

Stockfish-19-windows-x86-64-universal:
Found 145 WDL and 145 DTZ tablebase files (up to 5-man).
Available processors: 0-3
Using 1 thread
NNUE evaluation using nn-1a298aa575a0.nnue (109MiB, (86896, 1024, 32, 32, 1))
Network replica 1: Shared memory.
[...]
 64/77	01:18	 112.060.108	1.457.502	+0,74	Kf4e4 Ke2f1 g6g5 Kf1e2 Bf2c5 Ke2f1 Ke4d3 Kf1e1 Bc5e3 Ke1f1 g5g4 Kf1e1 Be3c5 Ke1f1 Bc5a7 Kf1e1 Ba7e3 Ke1f1 Be3b6 Kf1e1 Bb6a5+ Ke1f1 Kd3e3 Kf1g1 Ke3e2 Kg1h1 Ba5c7 Kh1g1 Ke2e1 Kg1h1 Bc7a5 Kh1g1 Ba5d2 Kg1h1 Bd2c3 Kh1g1 Bc3b4 Kg1h1 Ke1d2 Kh1g1 Kd2e2 Kg1h1 Ke2e1 Kh1g1 Bb4a3 Kg1h1 Ke1e2 Kh1g1 Ba3c5+ Kg1h1 Bc5e7 Kh1g1 Be7a3 Kg1h1 Ke2e1 Kh1g1 Ke1d2 Kg1h1 Ba3b2 Kh1g1 Kd2e2 Kg1h1 Bb2a1 Kh1g1 Ke2e1 Kg1h1 Ba1g7 Kh1g1 Bg7f6
 65/105	01:26	 123.732.628	1.469.962	+0,74	g6g5 Ke2d3 Bf2c5 Kd3d2 Kf4e4 Kd2e2 Bc5e3 Ke2f1 Ke4d3 Kf1e1 Be3d4 Ke1f1 Bd4b6 Kf1e1 Bb6c7 Ke1f1 Bc7a5 Kf1g1 Kd3e2 Kg1h1 Ba5b4 Kh1g1 Ke2e1 Kg1h1 Bb4c3 Kh1g1 Ke1e2 Kg1h1 Bc3a5 Kh1g1 Ba5b4 Kg1h1 Bb4d2 Kh1g1 Bd2f4 Kg1h1 Bf4c7 Kh1g1 Ke2e1 Kg1h1 Bc7b8 Kh1g1 Bb8d6 Kg1h1 Bd6c7 Kh1g1 Ke1e2 Kg1h1 Bc7d8 Kh1g1 Ke2e1 Kg1h1 Bd8e7 Kh1g1 Ke1e2 Kg1h1 Ke2d3 Kh1g1 Be7d6 Kg1f1 Bd6c5 Kf1e1 Bc5b4+ Ke1f1 Bb4d6 Kf1e1 Bd6e5 Ke1f1 Kd3d2 Kf1g1
 66/105+	01:30	 130.631.806	1.472.488	+0,75	Kf4e4
 66/105	01:32	 134.468.485	1.476.410	+0,74	Kf4e4 Ke2f1 Bf2e3 Kf1e2 Ke4d4 Ke2e1 Kd4d3 Ke1f1 Be3a7 Kf1e1 g6g5 Ke1f1 Ba7b6 Kf1e1 Kd3e3 Ke1f1 Ke3e4 Kf1e2 Ke4d4 Ke2d2 Bb6c5 Kd2d1 Kd4d3 Kd1e1 Bc5d4 Ke1f1 Kd3c2 Kf1e2 Bd4b6 Ke2f1 Bb6c5 Kf1e2 Bc5b4 Ke2f1 Kc2d3 Kf1g1 Kd3e2 Kg1h1 Bb4c3 Kh1g1 Ke2e1 Kg1h1 Bc3a5 Kh1g1 Ba5b4 Kg1h1 Bb4d2 Kh1g1 Bd2a5 Kg1h1 Ke1d1 Kh1g1 Kd1e2 Kg1h1 Ba5b4 Kh1g1 Bb4c3 Kg1h1 Bc3d2 Kh1g1 Ke2e1 Kg1h1 Bd2b4
 67/122	01:49	 163.338.550	1.515.720	+0,74	Kf4e4 Ke2f1 Bf2e3 Kf1e2 Ke4d4 Ke2e1 Kd4d3 Ke1f1 Be3a7 Kf1e1 g6g5 Ke1f1 Ba7c5 Kf1e1 Kd3e4 Ke1d2 Bc5b4+ Kd2e2 Bb4c3 Ke2f1 Ke4e3 Kf1g1 Ke3e2 Kg1h1 Bc3h8 Kh1g1 Ke2e1 Kg1h1 Bh8b2 Kh1g1 Bb2d4+ Kg1h1 Bd4g7 Kh1g1 Bg7c3 Kg1h1 Bc3a5 Kh1g1 g5g4 Kg1h1 Ba5d2 Kh1g1 Bd2h6 Kg1h1 Ke1e2 Kh1g1 Bh6c1 Kg1h1 Ke2d1 Kh1g1 Bc1e3+ Kg1f1 Kd1c2 Kf1e2 Be3d2 Ke2f1 Kc2d3 Kf1g1 Kd3e2 Kg1h1 Bd2c1 Kh1g1 Bc1e3+ Kg1h1 Be3f4 Kh1g1 Ke2e1 Kg1h1 Bf4e5 Kh1g1 Be5c3 Kg1h1 Bc3a5
 68/122	01:51	 165.873.864	1.518.212	+0,74	Kf4e4 Ke2f1 Bf2e3 Kf1e2 Ke4d4 Ke2e1 Kd4d3 Ke1f1 Be3a7 Kf1e1 g6g5 Ke1f1 Ba7c5 Kf1e1 Bc5d6 Ke1f1 Bd6e5 Kf1g1 Kd3e2 Kg1h1 Be5b8 Kh1g1 Ke2e1 Kg1h1 Ke1d2 Kh1g1 Bb8f4 Kg1h1 Bf4e5 Kh1g1 Kd2e1 Kg1h1 Be5c3 Kh1g1 Bc3b2 Kg1h1 Ke1e2 Kh1g1 Bb2c3 Kg1h1 Bc3d2 Kh1g1 Bd2f4 Kg1h1 Ke2e3 Kh1g1 Bf4d6 Kg1f1 Bd6e5 Kf1e1 Ke3d3 Ke1f1 Be5d4 Kf1e1 Bd4e3 Ke1f1 Be3b6 Kf1e1 Bb6d4 Ke1f1 Bd4e3 Kf1e1 Be3f2+ Ke1f1 Bf2a7 Kf1e1 Ba7c5 Ke1f1 Bc5d6 Kf1g1 Bd6f8 Kg1f1

There were 8,833 TB hits and SF insits to see a plus eval for black without any advances. I guess that 0.74 falls in the error margin in this kind of endgames.

Ok, Jesus, but let's have a look at a position only a few moves before. Here after 11...Kxg5, so already within the 7 men, and SF dev. 20260913 had 6men Syzygys loaded:

Code: Select all

 
8/8/3b2p1/6k1/8/6p1/4KpP1/8 w - - 0 1

Analysis by Stockfish dev-20260913-m-031dfeb4:

12.Kf1 Lf4 13.Ke2 Kf5 14.Kf1 Ke6 15.Ke2 Lg5 16.Kf1 Ld8 17.Ke2 Kd5 18.Kf1 
  Schwarz steht deutlich besser.  
  -+ (-2.31)  Tiefe: 10/22   00:00:00
...
12.Kf1 Lf4 13.Ke2 Lb8 14.Kf1 Kf6 15.Ke2 Ld6 16.Kf1 Kg7 17.Ke2 Lb4 18.Kf1 Lc3 19.Ke2 Kf6 20.Kf1 Ld2 21.Ke2 Lh6 22.Kf1 Ke6 23.Ke2 Kd5 24.Kf1 Lf4 25.Ke2 Ke6 26.Kf1 Ld6 27.Ke2 Lb8 28.Kf1 Ke7 29.Ke2 Kf8 30.Kf1 Ke8 31.Ke2 Ld6 32.Kf1 Le5 33.Ke2 La1 34.Kf1 Kf8 
  Schwarz steht deutlich besser.  
  -+ (-1.88)  Tiefe: 35/94   00:00:11  64125kN, tb=22880
...
12.Kf1 Lf4 13.Ke2 Lb8 14.Kf1 Kf6 15.Ke2 Ld6 16.Kf1 Kg7 17.Ke2 Lb4 18.Kf1 Lc3 19.Ke2 Kf6 20.Kf1 Ld2 21.Ke2 Lh6 22.Kf1 Ke5 23.Ke2 Kd5 24.Kf1 Lf4 25.Ke2 Ke6 26.Kf1 Ld6 27.Ke2 Lb8 28.Kf1 Ke7 29.Ke2 Kf8 30.Kf1 Ke8 31.Ke2 Kd8 32.Kf1 Ld6 33.Ke2 Lf8 34.Kf1 Ke8 
  Schwarz steht deutlich besser.  
  -+ (-1.93)  Tiefe: 41/102   00:00:16  96960kN, tb=27924
12.Kf1 Lf4 13.Ke2 Lb8 14.Kf1 Kf6 15.Ke2 Ld6 16.Kf1 Kg7 17.Ke2 Lb4 18.Kf1 Lc3 19.Ke2 Kf6 20.Kf1 Ld2 21.Ke2 Lh6 22.Kf1 Ke5 23.Ke2 Kd5 24.Kf1 Lf4 25.Ke2 Ke6 26.Kf1 Ld6 27.Ke2 Lb8 28.Kf1 Ke7 29.Ke2 Kf8 30.Kf1 Ke8 31.Ke2 Kd8 32.Kf1 g5 
  Schwarz steht deutlich besser.  
  -+ (-1.82)  Tiefe: 42/114   00:00:17  105MN, tb=28971
...
12.Kf1 Kg4 13.Ke2 Kf5 14.Kf1 Lc7 15.Ke2 Kf6 16.Kf1 Ld6 17.Ke2 Le7 18.Kf1 Lf8 19.Ke2 Lh6 20.Kf1 Lc1 21.Ke2 Kf5 22.Kf1 Ke5 23.Ke2 Lg5 24.Kf1 Lh6 25.Ke2 Kf5 26.Kf1 Kg4 27.Ke2 Lg5 28.Kf1 Lf6 29.Ke2 Ld4 30.Kf1 Lc5 31.Ke2 La3 32.Kf1 Kg5 33.Ke2 Kh4 34.Kf1 g5 35.Ke2 Lb4 36.Kf1 Lc5 
  Schwarz steht deutlich besser.  
  -+ (-2.12)  Tiefe: 50/95   00:00:34  214MN, tb=39366
12.Kf1 Kg4 13.Ke2 Kf5 14.Kf1 Lc7 15.Ke2 Kf6 16.Kf1 Ld6 17.Ke2 Le7 18.Kf1 Lf8 19.Ke2 Lh6 20.Kf1 Lc1 21.Ke2 Kf5 22.Kf1 Ke5 23.Ke2 Lg5 24.Kf1 Lh6 25.Ke2 Kf5 26.Kf1 Kg4 27.Ke2 Lg5 28.Kf1 Lf6 29.Ke2 Ld4 30.Kf1 Lc5 31.Ke2 La3 32.Kf1 Kg5 33.Ke2 Kh4 34.Kf1 g5 35.Ke2 Lb4 36.Kf1 Le7 
  Schwarz steht deutlich besser.  
  -+ (-1.93)  Tiefe: 51/111   00:00:35  220MN, tb=40259
...
12.Kf1 Kg4 13.Ke2 Lb8 14.Kf1 Lf4 15.Ke2 Kh5 16.Kf1 Le3 17.Ke2 Kg4 18.Kf1 Kh4 19.Ke2 Lh6 20.Kf1 Kg5 21.Ke2 Kf4 22.Kf1 Ke5 23.Ke2 Kd5 24.Kf1 Lf8 25.Ke2 Ke4 26.Kf1 Lc5 27.Ke2 Kd5 28.Kf1 Kc6 29.Ke2 Ld6 30.Kf1 La3 31.Ke2 Kd7 32.Kf1 Kc7 33.Ke2 Kb8 34.Kf1 Lb2 35.Ke2 Lf6 36.Kf1 Lh4 37.Ke2 Kc7 38.Kf1 Kb6 39.Ke2 Lf6 40.Kf1 Kc7 41.Ke2 La1 42.Kf1 Lb2 43.Ke2 Lf6 44.Kf1 Kd8 45.Ke2 Lh4 46.Kf1 Kc7 47.Ke2 Kb6 48.Kf1 Lf6 49.Ke2 La1 50.Kf1 Lb2 51.Ke2 Kc7 52.Kf1 La1 53.Ke2 
  Schwarz steht deutlich besser.  
  -+ (-2.19 ++)  Tiefe: 57/86   00:01:01  390MN, tb=53497
...
12.Kf1 Kh6 13.Ke2 Kh5 14.Kf1 Kg4 15.Ke2 Lb8 16.Kf1 Lf4 17.Ke2 Kh5 18.Kf1 Le3 19.Ke2 Kg4 20.Kf1 Kf5 21.Ke2 Lc1 22.Kf1 Ke5 23.Ke2 Lg5 24.Kf1 Kf5 25.Ke2 Ld2 26.Kf1 Lb4 27.Ke2 Kf6 28.Kf1 Lf8 29.Ke2 Kf5 30.Kf1 Ke6 31.Ke2 Ld6 32.Kf1 Kd7 33.Ke2 Kc7 34.Kf1 Lf4 35.Ke2 Kd7 36.Kf1 Kc6 37.Ke2 Lh6 38.Kf1 Lg7 39.Ke2 Kd6 40.Kf1 Lh6 41.Ke2 Kc7 42.Kf1 Ld2 43.Ke2 Kd6 44.Kf1 Ke6 45.Ke2 Le3 46.Kf1 Ld4 47.Ke2 Lc5 48.Kf1 Kd6 49.Ke2 Kc7 50.Kf1 Le3 51.Ke2 Ld4 52.Kf1 La1 53.Ke2 Kd8 54.Kf1 Kd7 55.Ke2 Lf6 
  Schwarz steht deutlich besser.  
  -+ (-1.96)  Tiefe: 60/131   00:01:13  469MN, tb=60197

That was single thread too (4.3GHz), 1G hash, with 30 threads you just don't need to try, if it's similiar with your hardware, no output after 5 minutes at all, depth 7 still, massive search explosion.

Critter 9:

Code: Select all

 
8/8/3b2p1/6k1/8/6p1/4KpP1/8 w - - 0 1

Analysis by Crystal 9:

12.Kf1 Lc5 13.Ke2 Ld4 14.Kf1 Kh6 15.Ke2 Kg7 16.Kf1 Kf8 17.Ke2 Ke8 18.Kf1 La7 19.Ke2 Kd7 20.Kf1 
  Schwarz hat entscheidenden Vorteil.  
  -+ (-2.70)  Tiefe: 7/23   00:00:00
12.Kf1 Lc5 13.Ke2 Ld4 14.Kf1 Kh6 15.Ke2 Lg7 16.Kf1 Lb2 
  Schwarz steht deutlich besser.  
  -+ (-2.45)  Tiefe: 8/30   00:00:00
12.Kf1 Lc5 13.Ke2 Ld4 14.Kf1 Kh6 15.Ke2 Lb6 16.Kf1 Le3 17.Ke2 Kg5 18.Kf1 Kf5 19.Ke2 Ke6 20.Kf1 Kd5 21.Ke2 La7 22.Kf1 Kc4 23.Ke2 Lb8 24.Kf1 
  Schwarz hat entscheidenden Vorteil.  
  -+ (-2.70)  Tiefe: 9/34   00:00:01  22191kN, tb=72573
12.Kf1 Lc5 13.Ke2 Ld4 14.Kf1 Kh6 15.Ke2 Lb6 16.Kf1 Le3 17.Ke2 Kg5 18.Kf1 Kf5 19.Ke2 Ke6 20.Kf1 Kd5 21.Ke2 La7 22.Kf1 Kc4 23.Ke2 Lb8 24.Kf1 
  Schwarz hat entscheidenden Vorteil.  
  -+ (-2.55 ++)  Tiefe: 10/39   00:00:08  131MN, tb=183803
12.Kf1 Lc5 13.Ke2 Ld4 14.Kf1 Kh6 15.Ke2 Lb6 16.Kf1 Kh5 17.Ke2 Lc7 18.Kf1 Kh6 19.Ke2 Ld6 20.Kf1 Kh5 21.Ke2 Le5 22.Kf1 Kg4 23.Ke2 g5 24.Kf1 Lb8 25.Ke2 
  Schwarz hat entscheidenden Vorteil.  
  -+ (-2.75)  Tiefe: 10/39   00:00:09  147MN, tb=202905
12.Kf1 Lc5 13.Ke2 Ld4 14.Kf1 Kh6 15.Ke2 Lb6 16.Kf1 Kh5 17.Ke2 Lc7 18.Kf1 Kh4 
  Schwarz hat entscheidenden Vorteil.  
  -+ (-2.70)  Tiefe: 11/37   00:00:09  152MN, tb=210665
12.Kf1 Lc5 13.Ke2 Ld4 14.Kf1 Kh6 15.Ke2 Lb6 16.Kf1 Kh5 17.Ke2 Lc7 18.Kf1 Kh4 
  Schwarz hat entscheidenden Vorteil.  
  -+ (-2.65 ++)  Tiefe: 12/27   00:01:11  1180MN, tb=817541

And Critter with compile of my own from latest github- sources:

Code: Select all

8/8/3b2p1/6k1/8/6p1/4KpP1/8 w - - 0 1

Analysis by Crystal dev-20260930-c81c0c7d:

12.Kf1 Kf4 13.Ke2 Ke5 14.Kf1 Kf5 15.Ke2 Lb4 16.Kf1 g5 17.Ke2 Le1 
  Schwarz steht besser.  
  -/+ (-0.85)  Tiefe: 6/21   00:00:00
12.Kf1 Lb4 13.Ke2 Lc3 14.Kf1 Lg7 15.Ke2 Lh6 16.Kf1 Kh5 17.Ke2 Kh4 18.Kf1 Lg7 19.Ke2 Le5 20.Kf1 Lb8 21.Ke2 Lf4 22.Kf1 Le5 23.Ke2 Kg4 
  Schwarz steht etwas besser.  
  =/+ (-0.65)  Tiefe: 7/61   00:00:01
12.Kf1 Kh6 13.Ke2 Le7 14.Kf1 Lf6 15.Ke2 Kh7 16.Kf1 Kg8 17.Ke2 Lh4 18.Kf1 Kh8 
  Schwarz steht etwas besser.  
  =/+ (-0.65)  Tiefe: 8/77   00:00:04  24158kN, tb=1546246
12.Kf1 La3 
  Schwarz steht etwas besser.  
  =/+ (-0.70 --)  Tiefe: 9/63   00:00:04  25745kN, tb=1563866
12.Kf1 La3 
  Schwarz steht etwas besser.  
  =/+ (-0.70 --)  Tiefe: 9/63   00:00:04  26454kN, tb=1568200
12.Kf1 La3 
  Schwarz steht besser.  
  -/+ (-0.75 --)  Tiefe: 9/63   00:00:04  28881kN, tb=1578130

The output was copied after 3 minutes, after these first 4 seconds simply no further iteration within that time.

Lc0 with BT4 6147500PT332 and RTX5070 GPU:

Code: Select all

8/8/3b2p1/6k1/8/6p1/4KpP1/8 w - - 0 1

Analysis by Lc0 v0.32.1trt:

12.Kf1 Kf4 
  Schwarz steht etwas besser.  
  =/+ (-0.34)  Tiefe: 1/2   00:00:05  2Knoten
...
12.Kf1 Kf4 13.Ke2 Ke4 14.Kf1 Kd4 15.Ke2 Kc3 16.Kf1 
  Schwarz steht etwas besser.  
  =/+ (-0.55)  Tiefe: 5/9   00:00:05  197Knoten, tb=1
...
12.Kf1 Kf4 13.Ke2 Ke4 14.Kf1 Lf4 15.Ke2 Kd4 16.Kf1 Kc3 17.Ke2 Le5 18.Kf1 Kd4 19.Ke2 
  Schwarz steht etwas besser.  
  =/+ (-0.52)  Tiefe: 12/19   00:00:10  660kN, tb=13136
...
12.Kf1 Kf5 13.Ke2 Ke4 14.Kf1 Le5 15.Ke2 Ld4 16.Kf1 Lc5 17.Ke2 Le3 18.Kf1 Kd4 19.Ke2 f1D+ 
  Schwarz steht etwas besser.  
  =/+ (-0.40)  Tiefe: 13/20   00:00:31  9197kN, tb=154783
...
12.Kf1 Lf8 13.Ke2 Kf5 14.Kf1 Ke4 15.Ke2 Ld6 16.Kf1 Kd4 17.Ke2 Lf4 18.Kf1 Kc3 19.Ke2 Lh6 20.Kf1 Kc2 21.Ke2 
  Schwarz steht etwas besser.  
  =/+ (-0.37)  Tiefe: 15/23   00:01:02  26889kN, tb=540521

[peter]

Thinking Lc0 to do best here, disregarding the missing 0.00- eval, I then tried Texel at the very start position of the puzzle.
Single thread (4.3GHz), 1G hash, no tbs:

Code: Select all

8/6R1/3b2p1/3R2P1/6P1/1p4p1/1p3pP1/2k2K2 w - -

Engine: Texel 1.12 (1024 MB)
von Peter Osterlund

21.01  0:01   -9.93    1.Txd6 b1D 2.Tgd7 Kb2+ 3.Td1 De4 
                       4.T7d2+ Kc3 5.Td3+ Kc4 6.Txg3 b2 
                       7.Kxf2 Dc2+ 8.Ke1 Dxd1+ 9.Kxd1 b1D+ 
                       10.Ke2 Dg1 11.Tf3 Kd4 12.Tf4+ Ke5 
                       13.Tf2 Ke4 14.Tf6 (4.943.712) 3304 

...

25.01  0:09   -13.44   1.Txd6 b1D 2.Tgd7 Kb2+ 3.Td1 De4 
                       4.T7d2+ Kc3 5.Td3+ Kc4 6.Txg3 b2 
                       7.Kxf2 Dc2+ 8.Ke1 Dxd1+ 9.Kxd1 b1D+ 
                       10.Ke2 Dg1 11.Tf3 Dxg2+ 12.Ke3 Kd5 
                       13.Kf4 Dh2+ 14.Ke3 (32.352.514) 3317 

25.09  0:10   -13.43++ 1.Tb7 b1D 2.Td1+ Kxd1 3.Txb3 De4 
                       4.Tb1+ Kd2 5.Tb2+ Kc1 6.Tc2+ Kd1 
                       7.Td2+ Kxd2 (34.160.041) 3294 
...

25.09  0:10   -12.31++ 1.Tb7 b1D 2.Td1+ Kxd1 3.Txb3 De4 
                       4.Tb1+ Kd2 5.Tb2+ Kc1 6.Tc2+ Kd1 
                       7.Td2+ Kxd2 (35.185.139) 3302 

...

25.09  0:11   -4.06++  1.Tb7 b1D 2.Td1+ Kxd1 3.Txb3 De4 
                       4.Tb1+ Kd2 5.Tb2+ Kc1 6.Tc2+ Kd1 
                       7.Td2+ Kxd2 (37.546.878) 3325 

25.09  0:11   0.00     1.Tb7 b1D 2.Td1+ Kxd1 3.Txb3 De4 
                       4.Tb1+ Kd2 5.Tb2+ Kc1 6.Tc2+ Kd1 
                       7.Td2+ Kxd2 (38.812.626) 3338 

...
30.01  0:16   0.00     1.Tb7 b1D 2.Td1+ Kxd1 3.Txb3 De4 
                       4.Tb1+ Kd2 5.Tb2+ Kc1 6.Tc2+ Kd1 
                       7.Td2+ Kxd2 (54.223.959) 3315 

31.01  0:18   0.00     1.Tb7 b1D 2.Td1+ Kxd1 3.Txb3 De4 
                       4.Tb1+ Kd2 5.Tb2+ Kc1 6.Tc2+ Kd1 
                       7.Td2+ Kxd2 (61.781.171) 3306 

Claude tells me, it's this code- sniplet of Texel's, that's responsible for solving this very special one position so fast:

EndGameEval::isBishopPawnDraw() in lib/texellib/endGameEval.cpp.

Hiarcs (15.4) might have something similiar, it does about as well as Texel does with the puzzle, regards

[Dann Corbit]

You can go to YACPDB or the PDB Problem Chess Database server and put in the keyword "Fortress" and get lists of Fortress positions.