I wonder if chess is a draw with less ranks

[Uri Blass]

I tried to analyze e3 g3 and Wasp claims mate in 11 for white after these moves

[pgn][Event "Computer chess game"]
[Site "DESKTOP-7QE6S12"]
[Date "2026.01.15"]
[Round "?"]
[White "àåøé"]
[Black "àåøé"]
[Result "1-0"]
[BlackElo "2400"]
[Time "22:33:05"]
[WhiteElo "2400"]
[TimeControl "300+3"]
[SetUp "1"]
[FEN "BBBBBBBB/BBBBBBBB/PPPPPPPP/rnbqkbnr/pppppppp/8/PPPPPPPP/RNBQKBNR w KQ - 0 1"]
[Termination "normal"]
[PlyCount "23"]
[WhiteType "human"]
[BlackType "human"]

1. e3 g3 2. exf4+ Kxf4 3. d3+ Ke5 4. f4+ exf3 5. gxf3 Bxd3 6. f4+ Kf5 7.
Bh3+ Nxh3 8. Qxh5+ Ke4 9. Nd2+ Ke3 10. Nxc4+ Ke4 11. cxd3+ Kxd3 12. Qe2#
1-0
[/pgn]



I know it is not my position and it seems the site does not show it but it is possible to try to check if the same idea works without the pawns and bishops that I added that their target is to prvent black to go backward.

[pgn][Event "Computer chess game"]
[Site "DESKTOP-7QE6S12"]
[Date "2026.01.15"]
[Round "?"]
[White "àåøé"]
[Black "àåøé"]
[Result "*"]
[BlackElo "2400"]
[Time "22:48:31"]
[WhiteElo "2400"]
[TimeControl "300+3"]
[SetUp "1"]
[FEN "8/8/8/rnbqkbnr/pppppppp/8/PPPPPPPP/RNBQKBNR w KQ - 0 1"]
[Termination "normal"]
[PlyCount "23"]
[WhiteType "human"]
[BlackType "human"]

1. e3 g3 2. exf4+ Kxf4 3. d3+ Ke5 4. f4+ exf3 5. gxf3 Bxd3 6. f4+ Kf5 7.
Bh3+ Nxh3 8. Qxh5+ Ke4 9. Nd2+ Ke3 10. Nxc4+ Ke4 11. cxd3+ Kxd3 12. Qe2# *[/pgn]

Now except Bd3 the only way to prvent fast mate are Bh3 Be4 Bg4 Qe4+ Qxf3 Nh3 Nxf3 cxd3
I checked that after 6.f4+ the same line works only in a smaller board

[lkaufman]

A different but related question that I find interesting and potentially important, is this: What is the smallest rule change from standard chess that would put the result near the win-draw line, so that with Armageddon rule it would be fully playable for the best engines? We determined some years ago that "Black cannot castle short" is near that line; also "White plays first with 1.e4 on the board" is fairly near the line, though probably on the draw side. "Black must avoid repetition" might also be near the line, though that is harder to test and I know of no data. Many other options are possible, not all are easy to evaluate.

part of the options that you give are simply changing the initial position and not changing the rules so engines can play it.

If we talk about changing the rules.
For black must avoid repetition we need to change chess engines to know repetition is a loss for black and I guess that black can draw by stalemate or by fifty move rule.

Other options of changing the target of the game without changing the intial position may be that white is winning by mate or by something else.
Examples:mate or promote a pawn.
If it is still a draw maybe mate or push a pawn to the 7th and if it is still a draw maybe mate or push a pawn to the 6th.

Mate or capturing the oppoent queen and rooks so black need to draw but also not to allow trades of too many pieces.

White win by mate or push a specific piece to specific square( a lot of options and I do not know what will be 50%)
I think that we need engines tthat allow us to change the target of the game to test it.

The rule change should be something that would make sense to chess players, not change the game too much from the goal of catching the king. Obviously making stalemate a win is logical but not sufficient. Similarly making endgames like King and minor vs king a win make some sense since at least one side can theoretically stalemate the other against bad play. But these two are still not sufficient to make the game close to a win. Making rep a win for White might be fair but doesn't seem quite like normal chess; better to forbid reps by some rule, but then fifty move rule would surely still leave Black winning, so maybe with fifty move rule adjudicated by some simple rules (material, possibly pawn advancement, the rules must only involve material and pawn advancement to be consistent with fifty move rule) the game could be almost totally fair with Armageddon rule. We would need a suitable strong engine to prove so.

[jkominek]

With 4 ranks it is obviously mate in 1 for white who win by e2xf3#
What about 5 ranks or 6 rankss?

Is there a software to analyze it?

rules are the same as normal chess when pawns can go 2 squares forward only in the first move and promote when they go to their last rank.

I guess white is going to win with 5 ranks but I know no software to make a serious analysis.

Not sure about 6 ranks or 7 ranks

I'd go about it this way. One, make more money. Two, buy a high-end GPU computer. Three, modify the Leela training code to support non-standard board sizes. Four, start training. A strong indication of the answers will emerge from the wash.

I'm stuck on step one, but oh well.

[jkominek]

A different but related question that I find interesting and potentially important, is this: What is the smallest rule change from standard chess that would put the result near the win-draw line, so that with Armageddon rule it would be fully playable for the best engines? We determined some years ago that "Black cannot castle short" is near that line; also "White plays first with 1.e4 on the board" is fairly near the line, though probably on the draw side. "Black must avoid repetition" might also be near the line, though that is harder to test and I know of no data. Many other options are possible, not all are easy to evaluate.

part of the options that you give are simply changing the initial position and not changing the rules so engines can play it.

If we talk about changing the rules.
For black must avoid repetition we need to change chess engines to know repetition is a loss for black and I guess that black can draw by stalemate or by fifty move rule.

Other options of changing the target of the game without changing the intial position may be that white is winning by mate or by something else.
Examples:mate or promote a pawn.
If it is still a draw maybe mate or push a pawn to the 7th and if it is still a draw maybe mate or push a pawn to the 6th.

Mate or capturing the oppoent queen and rooks so black need to draw but also not to allow trades of too many pieces.

White win by mate or push a specific piece to specific square( a lot of options and I do not know what will be 50%)
I think that we need engines tthat allow us to change the target of the game to test it.

The rule change should be something that would make sense to chess players, not change the game too much from the goal of catching the king. Obviously making stalemate a win is logical but not sufficient. Similarly making endgames like King and minor vs king a win make some sense since at least one side can theoretically stalemate the other against bad play. But these two are still not sufficient to make the game close to a win. Making rep a win for White might be fair but doesn't seem quite like normal chess; better to forbid reps by some rule, but then fifty move rule would surely still leave Black winning, so maybe with fifty move rule adjudicated by some simple rules (material, possibly pawn advancement, the rules must only involve material and pawn advancement to be consistent with fifty move rule) the game could be almost totally fair with Armageddon rule. We would need a suitable strong engine to prove so.

This talk of rule changes hurts my head. From Larry's list of possibilities I favor the idea of modified castling rights at the start of the game. In particular, I lean towards the absolute minimal change offered by "Black Cannot Castle" rules, with an equal five minutes per side in Armageddon games. Unequal clock time always struck me as an unwelcome kludge.

Curious about the benefits of castling about a year ago I generated some Chess324 data with variable castling rights. Each match is 16 x 324 = 5184 game pairs. I used fast time controls (low node budgets) to get about 2.5 million games. Noticing this thread I have dusted it off for analysis.

I split the games into 16 camps and let ordo compute the respective Elo advantage for White. The numbers make sense to me.

Code: Select all

      Black 
White    kq      k      q      -
----- -----  -----  -----  -----
KQ    23.18  37.59  67.84  95.85
K      8.79  21.72  52.68  80.30  
Q    -20.17  -5.20  24.05  53.07
-    -54.44 -39.06  -9.96  18.88

Next I reassigned the draws as wins for Black and recalculated White's advantage.

Code: Select all

      Black 
White    kq      k      q       -
----- -----  -----  -----   -----
KQ   -67.66 -56.61 -18.96   13.08
K    -87.03 -75.61 -34.06   -4.81
Q   -110.35 -93.89 -68.67  -35.27
-   -142.76 -128.5 -101.32 -74.58

In this data the tightest match-up is Black Cannot Castle, White Can Castle Short, i.e. K/-. The second most evenly matched conditions is Black Cannot Castle. These results are averaged over all of the competitions. To look at a particular case I selected Stockfish 17.1 and had it self-play with a node budget of 64k nodes/mv.

With draws assigned as wins for Black, again K/- castling rights proved to provide even conditions for the full suite of Chess324. As before the difference is represented as White's advantage in terms of Elo points.

Code: Select all

       Black 
White     kq       k       q       -
----- ------  ------  ------  ------
KQ   -211.48 -176.83  -25.78   43.11
K    -243.52 -208.46  -80.78    4.29
Q    -275.76 -163.33 -176.83  -89.88
-    -431.09 -303.87 -308.15 -205.46

The standard opening position is in general more drawish than Chess324. So for Armageddon tie breaking in human matches where the pressure is on White for a must-win, Black Cannot Castle rules looks as good and as simple a choice as any. It's worth a try at least.

[jkominek]

Part 2. Having advocated above for Black Cannot Castle rules to serve as an Armageddon tie-breaker game, a natural question follows. Does disallowing Black from castling leave the game of chess a draw, or is it a theoretical win for White?

If you go to cdb.cn the position has a backed-up Stockfish evaluation of 121 centipawns, albeit without nearly the same amount of exploration as the standard starting position with full castling rights. 121 maps onto a win probability of 75%. Too close to call but it does suggest that it could be a win.

So I let Stockfish 17.1 cogitate on the position for a couple hours. It answered: 119. Enticed, I handed the job over to stockfish-dev-20260111 for a half day, letting it take a one trillion node think. One trillion nodes said: 121. How about that. Looking for a third option I asked Komodo Dragon 3.3 to also give it a one trillion nodes contemplate. Komodo split the difference with: 120. Finally, I put Komodo Dragon 3.3 into Monte Carlo mode and let it tree build until it reached 25 million tree nodes, an effort that took 17 hours. The Komodo Monte Carlo search replied with: suspicious crap that I can't take seriously.

To see it visually here are a couple graphs showing the evolving evaluations. The first has the y-axis in centipawns and the second in win probability. The deep search of stockfish-dev-20260111 - labeled as SF17.9 - gave a final answer of 84%.



To trace the moves with cdb.cn evaluations:
Stockfish "17.9": https://www.chessdb.cn/queryc_en/?rnbqk ... _c6d5_d2d5

Komodo 3.3, alpha-beta search: https://www.chessdb.cn/queryc_en/?rnbq ... d7e6_b1a1

Komodo 3.3, Monte Carlo search: https://www.chessdb.cn/queryc_en/?rnbq ... b3b6_g4h3

As is typical of principal variations the last eight or ten ply tend to lose the trail. I found it odd that Stockfish-dev locked onto 1.e4 as the best opening move while both variants of Komodo as well as the Stockfish-based cdb.cn says 1.d4 is far stronger.

[pgn]
[Event "Komodo 3.3 Alpha-Beta 1 trillion Node Analysis"]
[Site "2x Intel Xeon Gold 6248R CPU @ 3.00GHz"]
[Date "2026.01.17"]
[Round "1"]
[White "Komodo 3.3 Hash 512GiB Threads 48"]
[Black "n/a"]
[Result "*"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQ - 0 1"]

1. d2d4 d7d5 2. c2c4 e7e6 3. g1f3 a7a6 4. e2e3 c7c6 5. d1c2 b8d7 6. c1d2 f7f5 7. f1e2 g8h6 8. b1c3 h6f7 9. c3a4 f8d6 10. c4c5 d6c7 11. e1c1 d7f6 12. h2h4 g7g6 13. d2c3 h7h5 14. f3g1 f6e4 15. c3e1 c8d7 16. c1b1 d8e7 17. g1h3 e8f8 18. h3f4 f8g7 19. f2f3 e4f6 20. e1c3 e6e5 20. d4e5 c7e5 21. h1e1 a8e8 22. c3e5 e7e5 23. e2d3 h8g8 24. a4b6 d7e6 25. b1a1 *
[/pgn]

[jkominek]

P.S. I forgot to add the kicker line. The Komodo 3.3 alpha-beta search exhibits a fascinating strategy that I wonder is workable in general. By move eight White had cleared out the back rank queen, knights, and bishops to enable the option of castling on either side. On move eleven it castled long.

[lkaufman]

Part 2. Having advocated above for Black Cannot Castle rules to serve as an Armageddon tie-breaker game, a natural question follows. Does disallowing Black from castling leave the game of chess a draw, or is it a theoretical win for White?

If you go to cdb.cn the position has a backed-up Stockfish evaluation of 121 centipawns, albeit without nearly the same amount of exploration as the standard starting position with full castling rights. 121 maps onto a win probability of 75%. Too close to call but it does suggest that it could be a win.

So I let Stockfish 17.1 cogitate on the position for a couple hours. It answered: 119. Enticed, I handed the job over to stockfish-dev-20260111 for a half day, letting it take a one trillion node think. One trillion nodes said: 121. How about that. Looking for a third option I asked Komodo Dragon 3.3 to also give it a one trillion nodes contemplate. Komodo split the difference with: 120. Finally, I put Komodo Dragon 3.3 into Monte Carlo mode and let it tree build until it reached 25 million tree nodes, an effort that took 17 hours. The Komodo Monte Carlo search replied with: suspicious crap that I can't take seriously.

To see it visually here are a couple graphs showing the evolving evaluations. The first has the y-axis in centipawns and the second in win probability. The deep search of stockfish-dev-20260111 - labeled as SF17.9 - gave a final answer of 84%.



To trace the moves with cdb.cn evaluations:
Stockfish "17.9": https://www.chessdb.cn/queryc_en/?rnbqk ... _c6d5_d2d5

Komodo 3.3, alpha-beta search: https://www.chessdb.cn/queryc_en/?rnbq ... d7e6_b1a1

Komodo 3.3, Monte Carlo search: https://www.chessdb.cn/queryc_en/?rnbq ... b3b6_g4h3

As is typical of principal variations the last eight or ten ply tend to lose the trail. I found it odd that Stockfish-dev locked onto 1.e4 as the best opening move while both variants of Komodo as well as the Stockfish-based cdb.cn says 1.d4 is far stronger.

[pgn]
[Event "Komodo 3.3 Alpha-Beta 1 trillion Node Analysis"]
[Site "2x Intel Xeon Gold 6248R CPU @ 3.00GHz"]
[Date "2026.01.17"]
[Round "1"]
[White "Komodo 3.3 Hash 512GiB Threads 48"]
[Black "n/a"]
[Result "*"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQ - 0 1"]

1. d2d4 d7d5 2. c2c4 e7e6 3. g1f3 a7a6 4. e2e3 c7c6 5. d1c2 b8d7 6. c1d2 f7f5 7. f1e2 g8h6 8. b1c3 h6f7 9. c3a4 f8d6 10. c4c5 d6c7 11. e1c1 d7f6 12. h2h4 g7g6 13. d2c3 h7h5 14. f3g1 f6e4 15. c3e1 c8d7 16. c1b1 d8e7 17. g1h3 e8f8 18. h3f4 f8g7 19. f2f3 e4f6 20. e1c3 e6e5 20. d4e5 c7e5 21. h1e1 a8e8 22. c3e5 e7e5 23. e2d3 h8g8 24. a4b6 d7e6 25. b1a1 *
[/pgn]

If Black cannot castle gives evals around 1.2 it is clearly unsuitable for Armageddon, a 75 to 25 expected score is totally unacceptable. That's why I proposed Black cannot castle short, that should be much closer to 1.0. Even if your fast stats suggest it's not as good a fit as no Black castling, I would trust the eval more.

[Uri Blass]

Part 2. Having advocated above for Black Cannot Castle rules to serve as an Armageddon tie-breaker game, a natural question follows. Does disallowing Black from castling leave the game of chess a draw, or is it a theoretical win for White?

If you go to cdb.cn the position has a backed-up Stockfish evaluation of 121 centipawns, albeit without nearly the same amount of exploration as the standard starting position with full castling rights. 121 maps onto a win probability of 75%. Too close to call but it does suggest that it could be a win.

So I let Stockfish 17.1 cogitate on the position for a couple hours. It answered: 119. Enticed, I handed the job over to stockfish-dev-20260111 for a half day, letting it take a one trillion node think. One trillion nodes said: 121. How about that. Looking for a third option I asked Komodo Dragon 3.3 to also give it a one trillion nodes contemplate. Komodo split the difference with: 120. Finally, I put Komodo Dragon 3.3 into Monte Carlo mode and let it tree build until it reached 25 million tree nodes, an effort that took 17 hours. The Komodo Monte Carlo search replied with: suspicious crap that I can't take seriously.

To see it visually here are a couple graphs showing the evolving evaluations. The first has the y-axis in centipawns and the second in win probability. The deep search of stockfish-dev-20260111 - labeled as SF17.9 - gave a final answer of 84%.



To trace the moves with cdb.cn evaluations:
Stockfish "17.9": https://www.chessdb.cn/queryc_en/?rnbqk ... _c6d5_d2d5

Komodo 3.3, alpha-beta search: https://www.chessdb.cn/queryc_en/?rnbq ... d7e6_b1a1

Komodo 3.3, Monte Carlo search: https://www.chessdb.cn/queryc_en/?rnbq ... b3b6_g4h3

As is typical of principal variations the last eight or ten ply tend to lose the trail. I found it odd that Stockfish-dev locked onto 1.e4 as the best opening move while both variants of Komodo as well as the Stockfish-based cdb.cn says 1.d4 is far stronger.

[pgn]
[Event "Komodo 3.3 Alpha-Beta 1 trillion Node Analysis"]
[Site "2x Intel Xeon Gold 6248R CPU @ 3.00GHz"]
[Date "2026.01.17"]
[Round "1"]
[White "Komodo 3.3 Hash 512GiB Threads 48"]
[Black "n/a"]
[Result "*"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQ - 0 1"]

1. d2d4 d7d5 2. c2c4 e7e6 3. g1f3 a7a6 4. e2e3 c7c6 5. d1c2 b8d7 6. c1d2 f7f5 7. f1e2 g8h6 8. b1c3 h6f7 9. c3a4 f8d6 10. c4c5 d6c7 11. e1c1 d7f6 12. h2h4 g7g6 13. d2c3 h7h5 14. f3g1 f6e4 15. c3e1 c8d7 16. c1b1 d8e7 17. g1h3 e8f8 18. h3f4 f8g7 19. f2f3 e4f6 20. e1c3 e6e5 20. d4e5 c7e5 21. h1e1 a8e8 22. c3e5 e7e5 23. e2d3 h8g8 24. a4b6 d7e6 25. b1a1 *
[/pgn]

If Black cannot castle gives evals around 1.2 it is clearly unsuitable for Armageddon, a 75 to 25 expected score is totally unacceptable. That's why I proposed Black cannot castle short, that should be much closer to 1.0. Even if your fast stats suggest it's not as good a fit as no Black castling, I would trust the eval more.

I think that we need games to find the expected result and we have no data what is the probability of win with 1.2 evaluation.
Iwould like somebody to gnerate a lot of random chess positions find the position when the evaluation is 1.2 after 1 seconds of search and see what is the result of stockfish-stockfish games ar 1 second per move later find all the position when stockfrish evaluates as +1.2 after 10 seconds of search and find the result with 10 seconds per move.

I am not sure results are going to be the same with 1 second per move and 10 seconds per move and it may be interesting to try to tanslate results to expected probability based on data at different time controls.

People claim that +1 is translated to 50% win for white but I do not trust that it is the same at all time controls and we cannot know based on TCEC becauase they test different engines and not stockfish against itself.

[lkaufman]

Part 2. Having advocated above for Black Cannot Castle rules to serve as an Armageddon tie-breaker game, a natural question follows. Does disallowing Black from castling leave the game of chess a draw, or is it a theoretical win for White?

If you go to cdb.cn the position has a backed-up Stockfish evaluation of 121 centipawns, albeit without nearly the same amount of exploration as the standard starting position with full castling rights. 121 maps onto a win probability of 75%. Too close to call but it does suggest that it could be a win.

So I let Stockfish 17.1 cogitate on the position for a couple hours. It answered: 119. Enticed, I handed the job over to stockfish-dev-20260111 for a half day, letting it take a one trillion node think. One trillion nodes said: 121. How about that. Looking for a third option I asked Komodo Dragon 3.3 to also give it a one trillion nodes contemplate. Komodo split the difference with: 120. Finally, I put Komodo Dragon 3.3 into Monte Carlo mode and let it tree build until it reached 25 million tree nodes, an effort that took 17 hours. The Komodo Monte Carlo search replied with: suspicious crap that I can't take seriously.

To see it visually here are a couple graphs showing the evolving evaluations. The first has the y-axis in centipawns and the second in win probability. The deep search of stockfish-dev-20260111 - labeled as SF17.9 - gave a final answer of 84%.



To trace the moves with cdb.cn evaluations:
Stockfish "17.9": https://www.chessdb.cn/queryc_en/?rnbqk ... _c6d5_d2d5

Komodo 3.3, alpha-beta search: https://www.chessdb.cn/queryc_en/?rnbq ... d7e6_b1a1

Komodo 3.3, Monte Carlo search: https://www.chessdb.cn/queryc_en/?rnbq ... b3b6_g4h3

As is typical of principal variations the last eight or ten ply tend to lose the trail. I found it odd that Stockfish-dev locked onto 1.e4 as the best opening move while both variants of Komodo as well as the Stockfish-based cdb.cn says 1.d4 is far stronger.

[pgn]
[Event "Komodo 3.3 Alpha-Beta 1 trillion Node Analysis"]
[Site "2x Intel Xeon Gold 6248R CPU @ 3.00GHz"]
[Date "2026.01.17"]
[Round "1"]
[White "Komodo 3.3 Hash 512GiB Threads 48"]
[Black "n/a"]
[Result "*"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/RNBQKBNR w KQ - 0 1"]

1. d2d4 d7d5 2. c2c4 e7e6 3. g1f3 a7a6 4. e2e3 c7c6 5. d1c2 b8d7 6. c1d2 f7f5 7. f1e2 g8h6 8. b1c3 h6f7 9. c3a4 f8d6 10. c4c5 d6c7 11. e1c1 d7f6 12. h2h4 g7g6 13. d2c3 h7h5 14. f3g1 f6e4 15. c3e1 c8d7 16. c1b1 d8e7 17. g1h3 e8f8 18. h3f4 f8g7 19. f2f3 e4f6 20. e1c3 e6e5 20. d4e5 c7e5 21. h1e1 a8e8 22. c3e5 e7e5 23. e2d3 h8g8 24. a4b6 d7e6 25. b1a1 *
[/pgn]

If Black cannot castle gives evals around 1.2 it is clearly unsuitable for Armageddon, a 75 to 25 expected score is totally unacceptable. That's why I proposed Black cannot castle short, that should be much closer to 1.0. Even if your fast stats suggest it's not as good a fit as no Black castling, I would trust the eval more.

I think that we need games to find the expected result and we have no data what is the probability of win with 1.2 evaluation.
Iwould like somebody to gnerate a lot of random chess positions find the position when the evaluation is 1.2 after 1 seconds of search and see what is the result of stockfish-stockfish games ar 1 second per move later find all the position when stockfrish evaluates as +1.2 after 10 seconds of search and find the result with 10 seconds per move.

I am not sure results are going to be the same with 1 second per move and 10 seconds per move and it may be interesting to try to tanslate results to expected probability based on data at different time controls.

People claim that +1 is translated to 50% win for white but I do not trust that it is the same at all time controls and we cannot know based on TCEC becauase they test different engines and not stockfish against itself.

I investigated this issue a while ago and while time control does make some difference, the differences are tiny, not enough to be interesting or to change any other conclusions. The eval of a position must be pretty close to 1 (at least within the 0.9 to 1.1 range) at almost any reasonable time limit (ten seconds seems reasonable on good hardware) to be a candidate for use in Armageddon.

[jkominek]

If Black cannot castle gives evals around 1.2 it is clearly unsuitable for Armageddon, a 75 to 25 expected score is totally unacceptable. That's why I proposed Black cannot castle short, that should be much closer to 1.0. Even if your fast stats suggest it's not as good a fit as no Black castling, I would trust the eval more.

The WDL model of Stockfish, Leela, and others including I presume Komodo Dragon, are based on computer-computer self-play measurements at bullet time controls. It is a guide but not one we should over-reach.

The question in my mind, and one I am interested in your opinion on, is How do you know? By that I mean how do you know that computer results translate directly to equivalent human results without further empirical studies? Absent abundant evidence from human play I treat the measurements I've made (and yours) as suggestive but not definitive.

What stands out to me as the confounding variable is the ability of human grandmasters to semi-reliably secure a draw when needed, in this case by Black. We also know that computer ratings lists are dilated with respect to FIDE ratings, e.g. as expressed in Komodo Dragon's UCI Elo settings, a fact I attribute to human's greater tendency to play for draws, especially when the players have a large rating separation. (Versus engines fighting it out for the win for many more moves.)

Breaking it down,
i) Under Armageddon rules the ideal opening position score 50% between equally rated players.
ii) Modern engines are calibrated such that an evaluation of 100 centipawns equates to a 50% win/draw ratio - in computer-computer play. This puts us in the ballpark but should not be considered definitive.
iii) Observation: human grandmasters have the ability to successfully play for a draw when needed. Also, placing White in a must-win situation adds clutch-play pressure.
iv) Together this tilts the advantage further in Black's favor.
v) To compensate the opening position needs to more strongly favor White that what computer WDL models consider balanced.
vi) This argues for a value somewhat above 100 centipawns being the target point. A value of 120 may seem too high but should not be ruled out as suitable in human play.
vii) Likewise I would not rule out a position with an evaluation of 80, either. Because maybe I'm wrong about statement iv).

In short I advocate for a broader range of consideration (appreciation of uncertainty) until thoroughly put to the test.

I'm limited by not having a database of human Armageddon games at my disposal. Do you have a collection of games to study from your connections to either chess.com or Lichess? Maybe you have analysis that you could share. Even if only for the standard opening position it would provide a useful point of comparison.

More ideal of course a large experiment on either of those playing sites to study variant castling rights.