Add Human-SL even-game ELO rank ladder configs (gtp_human<rank>.cfg, 8d→25k)

[ChinChangYang]

Summary

This PR now ships a direct even-game ELO ladder — 33 GTP configs (gtp_human<rank>.cfg, 8d → 25k) that make KataGo + the Human-SL net play at a chosen amateur strength — and is slimmed to configs + documentation + one 13-line rank-parsing extension.

@lightvector — you were right that the komi-0.5 pairing is only half a stone. Rather than re-derive a rank spacing, this revision abandons rank-spacing entirely: each rung is tuned/measured against its stronger neighbour in a normal even game (komi 6.5, colours alternated), so the ladder no longer depends on any KGS handicap convention.

What changed since the previous revision

• Calibration target: komi-0.5 handicap @ 50% winrate → even game @ +100 ELO per rung, each adjacent gap held to a 95% CI ⊂ [70, 130].
• Configs: maxVisits 400 → 40, winLossUtilityFactor 1.0 → 0 (the old wLU=1 made the AI excessively strong even for DDK profiles), main net kata1-b28c512nbt-s8326494464-d4628051565 — every header names this exact net, and the calibrated rungs additionally warn that a different main net invalidates the calibration.
• Scope: the tunehuman subcommand, tests, and tuning scripts are removed from this PR — the engine should focus on playing and analysis, not offline tuning. The calibration harness lives in my fork (tunehuman-mlx). What remains: 33 configs + docs/HumanSL_Rank_Ladder.md + a 13-line extension of SGFMetadata::getProfile (rank strings 21k–30k; upstream parsing stops at 20k, so preaz_21k…preaz_25k would otherwise throw at startup — the rank input encoding itself already saturates at inverse-rank 34 = 25k).
• 9d dropped: the old gtp_human9d.cfg was a 400-visit, wLU=1 legacy reference measured roughly +655 ELO above 8d — not a rung of this ladder.

Method (brief)

Chained calibration anchored at a hand-set 8d top rung (λ=0.06): each rung is tuned via humanSLChosenMovePiklLambda so its already-locked stronger neighbour beats it by 100 ELO in even games (Japanese rules, komi 6.5, alternating colours, 40 visits). The lock criterion is the Wilson 95% CI of the gap, measured on games concentrated at the single shipped λ (binomial, φ=1). Full method, findings, and reproduction: docs/HumanSL_Rank_Ladder.md.

Results

7d → 14k — the 100-ELO staircase

Config	Profile	Baseline (stronger)	Even-game gap (95% CI)	Games	piklLambda
gtp_human8d.cfg	preaz_8d	— (hand-set anchor)	anchor, not calibrated	—	0.06
gtp_human7d.cfg	preaz_7d	gtp_human8d.cfg	+100 [72, 128] ✅	644	0.07760
gtp_human6d.cfg	preaz_6d	gtp_human7d.cfg	+100 [71, 130] ✅	576	0.09940
gtp_human5d.cfg	preaz_5d	gtp_human6d.cfg	+91 [70, 111] ✅	1208	0.13240
gtp_human4d.cfg	preaz_4d	gtp_human5d.cfg	+76 [57, 96] ⚠ below +100	1256	0.15750
gtp_human3d.cfg	preaz_3d	gtp_human4d.cfg	+112 [89, 134] ⚠ above +100	1039	0.18960
gtp_human2d.cfg	preaz_2d	gtp_human3d.cfg	+101 [74, 127] ✅	719	0.21300
gtp_human1d.cfg	preaz_1d	gtp_human2d.cfg	+106 [85, 128] ✅	1087	0.19170
gtp_human1k.cfg	preaz_1k	gtp_human1d.cfg	+105 [82, 129] ✅	896	0.20150
gtp_human2k.cfg	preaz_2k	gtp_human1k.cfg	+96 [71, 121] ✅	800	0.19950
gtp_human3k.cfg	preaz_3k	gtp_human2k.cfg	+100 [72, 127] ✅	664	0.20760
gtp_human4k.cfg	preaz_4k	gtp_human3k.cfg	+93 [71, 114] ✅	1088	0.21180
gtp_human5k.cfg	preaz_5k	gtp_human4k.cfg	+108 [86, 129] ✅	1104	0.21600
gtp_human6k.cfg	preaz_6k	gtp_human5k.cfg	+93 [71, 115] ✅	1000	0.22480
gtp_human7k.cfg	preaz_7k	gtp_human6k.cfg	+100 [71, 129] ✅	608	0.24590
gtp_human8k.cfg	preaz_8k	gtp_human7k.cfg	+102 [75, 129] ✅	688	0.25840
gtp_human9k.cfg	preaz_9k	gtp_human8k.cfg	+107 [85, 129] ✅	1032	0.30620
gtp_human10k.cfg	preaz_10k	gtp_human9k.cfg	+101 [74, 128] ✅	688	0.37250
gtp_human11k.cfg	preaz_11k	gtp_human10k.cfg	+103 [76, 129] ✅	704	0.40810
gtp_human12k.cfg	preaz_12k	gtp_human11k.cfg	+100 [71, 129] ✅	608	0.46300
gtp_human13k.cfg	preaz_13k	gtp_human12k.cfg	+104 [79, 129] ✅	800	0.83000
gtp_human14k.cfg	preaz_14k	gtp_human13k.cfg	+100 [71, 129] ✅	592	3.40040

All rungs run at maxVisits = 40; strength is set by λ only.

Honest exceptions, stated up front: 4d (+76) and 3d (+112) were re-measured at their exact shipped λ after an optimistic pooled estimator was found and removed (it had averaged a λ-gradient into a biased, too-tight CI on the steep dan rungs). Their honest gaps fall just outside [70, 130]; they are documented as measured, not re-tuned. All other dan/low-kyu rungs certify in-band.

15k → 25k — pure-human tail (λ = 1e8)

At this depth adjacent Human-SL ranks are near-tied: the even-game gap is non-monotonic in λ and peaks below +100 ELO, so a 100-ELO step is not reachable by the λ lever alone. These rungs ship at pure-human imitation and their natural gap vs the rung above is measured, not forced:

Config	Profile	Baseline (stronger)	Even-game gap (95% CI)	Games	piklLambda
gtp_human15k.cfg	preaz_15k	gtp_human14k.cfg	+29 [-1, 59]	528	1e8
gtp_human16k.cfg	preaz_16k	gtp_human15k.cfg	+24 [-6, 55]	512	1e8
gtp_human17k.cfg	preaz_17k	gtp_human16k.cfg	+44 [14, 75]	504	1e8
gtp_human18k.cfg	preaz_18k	gtp_human17k.cfg	+70 [40, 101]	520	1e8
gtp_human19k.cfg	preaz_19k	gtp_human18k.cfg	+32 [2, 62]	520	1e8
gtp_human20k.cfg	preaz_20k	gtp_human19k.cfg	+8 [-22, 39]	504	1e8
gtp_human21k.cfg	preaz_21k	gtp_human20k.cfg	+23 [-7, 54]	504	1e8
gtp_human22k.cfg	preaz_22k	gtp_human21k.cfg	+29 [-1, 59]	504	1e8
gtp_human23k.cfg	preaz_23k	gtp_human22k.cfg	−1 [-32, 29]	496	1e8
gtp_human24k.cfg	preaz_24k	gtp_human23k.cfg	+55 [25, 85]	536	1e8
gtp_human25k.cfg	preaz_25k	gtp_human24k.cfg	+86 [55, 116]	528	1e8

The small deep-kyu gaps are an honest property of the Human-SL net (adjacent deep-kyu rank inputs encode nearly the same strength), consistent with the earlier evaluation of the old rank-spaced ladder, which found several deep-kyu pairs tied or inverted. A full 100-ELO deep-kyu staircase would need a different weakening lever (fewer visits, temperature, or a weaker human-model rank) — possible future work.

Notes

• The rank calibration is realized at move selection (GTP genmove). The analysis engine can load the human model and return humanPolicy under these settings via overrideSettings, but it does not itself perform the calibrated move choice.
• Configs use humanSLProfile = preaz_<rank> (pre-AlphaZero KGS-rank profiles) with b18c384nbt-humanv0 as the human model.

🤖 Generated with Claude Code

https://claude.ai/code/session_014DKwH3TAhhGUSWqsMMoQcH

[lightvector]

29 configs, gtp_human{9d…1d,1k…20k}.cfg. The Human-SL net is conditioned on KGS rank, so a
1-rank gap is well-defined: per KGS, 1 rank = 1 stone, realized as an even game where the
stronger side (White) gets no komi compensation (komi 0.5). Each rung is tuned so the weaker
rank (Black, komi 0.5) is an even game (50%) vs the rung above, with the 95% CI ⊂
[40%, 60%], under Japanese rules.

Isn't that only half a stone?

KGS's match system by default matches players 1 rank apart with no-komi games, but that's to conform to convention. I'm pretty sure the ranking system understands this too.

[lightvector]

i.e. I'm pretty sure the ranking system understands and is calibrated for the fact that for players exactly 1 rank apart (in internal full-precision ratings) who play a no-komi game, white is expected to win more often than 50%. The true 1-rank difference would be a reverse-komi game.

[ChinChangYang]

I’m retuning Lambda with winLossUtilityFactor=0, 28b main network + maxVisits=40, and an ELO gap set to +100. The estimated time to completion is around 20 days. I’ll update this PR once it’s ready.

The retuning aims to address the following issues:

• The winLossUtilityFactor=1 makes the AI excessively strong, even for DDK profiles.
• The 6b network is weaker compared to the 18b human SL network for high-dan profiles.
• Tuning by 0.5 or 1 stone gaps results in unstable ELO gaps between ranks. Some ranks have ELO gaps exceeding 100, while others have nearly 0 ELO gaps.

[glittering-snow]

Thanks for working on this.
The rank-as-strength calibration is exactly the missing piece for my educational use case. When the retuning is complete, would the resulting per-rank settings also be intended for the analysis engine via overrideSettings, or only for GTP configs? I’d also be happy to test a few ranks in real user-facing play once candidate configs are available.

[ChinChangYang]

The ladder is now calibrated for normal even games (komi 6.5, alternating colors), with each rung tuned against its stronger neighbor to achieve a +100 ELO gain per rung. Additionally, every adjacent gap is directly measured to a 95% confidence interval (per-rung λ / gap / CI / games), as described in the updated description and docs/HumanSL_Rank_Ladder.md.

I also reduced the PR to configurations, documentation, and a 13-line SGFMetadata::getProfile extension for parsing preaz_21k and preaz_25k. The tunehuman calibration harness remains in my fork because the engine should focus on playing and analysis. Two honest caveats are documented rather than overlooked: 4d (+76 [57,96]) and 3d (+112 [89,134]) measurements are just outside the [70,130] range at their shipped λ, and the deep-kyu tail (15k→25k) is shipped as a pure-human imitation. (λ=1e8) because the ranks of adjacent deep-kyu Human-SL are nearly tied.