OpenBench Testing Framework

Finished
ClaudeCodasearch/fut-base-clock-formdiff10.0+0.10LLR: -2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 54920 W: 13447 L: 13627 D: 27846
Ptnml(0-2): 148, 6618, 14097, 6460, 137
thread11 clock form STC
ClaudeCodamaindiff10.0+0.00LLR: -2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 35408 W: 8958 L: 9177 D: 17273
Ptnml(0-2): 253, 4195, 8962, 4106, 188
no-inc divisor base 34->30 at 10+0 (no-inc path live). Reverse of the forfeit-fix trade: read time forfeits beside Elo. Registered: see coda-research SPEC_level_vs_variance 2026-09-25
ClaudeCodamaindiff10.0+0.00LLR: -2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 9050 W: 2122 L: 2387 D: 4541
Ptnml(0-2): 93, 1153, 2251, 982, 46
no-inc divisor base 34->26 at 10+0 (no-inc path live). Reverse of the forfeit-fix trade: read time forfeits beside Elo. Registered: see coda-research SPEC_level_vs_variance 2026-09-25
ClaudeCodarazor/allnode-rebalanced2diff10.0+0.10LLR: -2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 140836 W: 34177 L: 34232 D: 72427
Ptnml(0-2): 366, 16951, 35803, 16968, 330
allNode razoring + margin 220 (cutoff-count parity); rebalanced arm for 3736
ClaudeCodannue/shared-weightsdiff10.0+0.10LLR: -1.33 (-2.94, 2.94) [0.00, 2.00]
Games: 12644 W: 3069 L: 3163 D: 6412
Ptnml(0-2): 23, 1431, 3515, 1323, 30
share NNUE weights across processes (bench-identical)
ClaudeCodarazor/allnode-220-v2diff10.0+0.10LLR: -2.97 (-2.94, 2.94) [0.00, 2.00]
Games: 31970 W: 7714 L: 7931 D: 16325
Ptnml(0-2): 106, 3896, 8179, 3717, 87
allNode razoring + margin 220, cutoff parity recalibrated on this base
ClaudeCodarazor/tt-pv-guarddiff10.0+0.10LLR: -2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 161560 W: 39657 L: 39681 D: 82222
Ptnml(0-2): 440, 18730, 42464, 18706, 440
Razoring declines at formerly-PV nodes (tt_pv), matching RFP
ClaudeCodarazor/allnode-rebalanced-200diff10.0+0.10LLR: -1.99 (-2.94, 2.94) [0.00, 2.00]
Games: 26376 W: 6385 L: 6523 D: 13468
Ptnml(0-2): 79, 3210, 6738, 3092, 69
allNode razoring + margin 200 (tree-size parity); rebalanced arm for 3736
ClaudeCodatune/main-3725diff40.0+0.40LLR: -2.14 (-2.94, 2.94) [0.00, 2.00]
Games: 82782 W: 19745 L: 19810 D: 43227
Ptnml(0-2): 58, 9687, 21959, 9636, 51
tune #3725 applied (corrhist cluster), LTC — decides
ClaudeCodatune/universal-3701-3709diff40.0+0.40LLR: 0.15 (-2.94, 2.94) [0.00, 2.00]
Games: 89448 W: 21431 L: 21293 D: 46724
Ptnml(0-2): 53, 10281, 23924, 10407, 59
universal 15-knob subset of tunes 3701+3709, LTC
ClaudeCodatune/main-3709diff80.0+0.80LLR: 2.05 (-2.94, 2.94) [0.00, 2.00]
Games: 90966 W: 21658 L: 21360 D: 47948
Ptnml(0-2): 28, 10296, 24535, 10598, 26
tune #3709 (LTC-tuned pruning) at 2xLTC: STC -1.7 / LTC +1.8 — does it grow with clock? Hash 256 (fleet ceiling)
ClaudeCodatm/churn-memory-t1diff10.0+0.10LLR: -2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 22434 W: 5350 L: 5576 D: 11508
Ptnml(0-2): 86, 2676, 5879, 2530, 46
churn-memory T=1 factor, ungated, 350 (fail-low factor's own step/cap). Predicted -1..+1.5
ClaudeCodatm/churn-memory-t1diff10.0+0.10LLR: -2.97 (-2.94, 2.94) [0.00, 2.00]
Games: 18820 W: 4587 L: 4821 D: 9412
Ptnml(0-2): 72, 2281, 4909, 2105, 43
churn-memory T=1 factor, gated on settled move (stable>=1), 350. Predicted +0.5..+2
ClaudeCodarazor/tt-quiet-guarddiff10.0+0.10LLR: -2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 161872 W: 39399 L: 39422 D: 83051
Ptnml(0-2): 412, 18804, 42512, 18811, 397
Razoring declines when TT names a quiet best move; qsearch is blind to quiet moves
ClaudeCodarazor/hygiene-and-tunablesdiff10.0+0.10Tuning 3 Parameters
2500/2500 Iterations
40000/40000 Games Played
Short tune of the 3 newly-promoted razoring constants (Thor 627ede1)
ClaudeCodarazor/allnode-onlydiff10.0+0.10LLR: -2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 71264 W: 17388 L: 17545 D: 36331
Ptnml(0-2): 181, 8650, 18147, 8453, 201
Razoring restricted to allNodes (Thor a521755). Margin deliberately unchanged: one variable. Own measurement: 83.1% allNode hit rate vs 45.3% cutNode.
ClaudeCodatune/universal-3701-3709diff10.0+0.10LLR: 2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 95972 W: 23581 L: 23179 D: 49212
Ptnml(0-2): 236, 11292, 24562, 11626, 270
universal 15-knob subset of tunes 3701+3709 (both clocks agreed), STC
ClaudeCodatune/main-3725diff10.0+0.10LLR: 2.94 (-2.94, 2.94) [0.00, 2.00]
Games: 155450 W: 38205 L: 37717 D: 79528
Ptnml(0-2): 401, 18283, 39864, 18781, 396
tune #3725 applied (corrhist cluster, 10 knobs, 25k x 2), STC
ClaudeCodatune/main-3709diff40.0+0.40LLR: 2.96 (-2.94, 2.94) [0.00, 2.00]
Games: 65190 W: 15870 L: 15525 D: 33795
Ptnml(0-2): 42, 7516, 17137, 7855, 45
tune #3709 (LTC-tuned pruning), LTC
ClaudeCodamaindiff10.0+0.10Tuning 10 Parameters
25002/25000 Iterations
100010/100000 Games Played
Correction-history cluster deep tune (10 knobs: 4 source weights, hist div, update weight max, bonus cap, grain, err div, DEXT corr margin). 25k x 2 pairs, #3683 shape. Adam: narrow cluster, full length.
ClaudeCodaabl/rfp-qs-verifydiff10.0+0.10LLR: 0.37 (-2.94, 2.94) [-1.50, 0.50]
Games: 264948 W: 65184 L: 65338 D: 134426
Ptnml(0-2): 585, 29051, 73406, 28797, 635
remove no-op RFP qs-verify (tree-identical)
ClaudeCodatune/main-3709diff10.0+0.10LLR: -2.97 (-2.94, 2.94) [-1.50, 0.50]
Games: 64038 W: 15538 L: 15850 D: 32650
Ptnml(0-2): 169, 7901, 16183, 7605, 161
tune #3709 (LTC-tuned pruning, 49 knobs), STC non-regression
ClaudeCodasearch/tt-refined-prune-eval-tuneddiff40.0+0.40LLR: -2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 10686 W: 2431 L: 2666 D: 5589
Ptnml(0-2): 11, 1354, 2840, 1135, 3
TT-refined prune eval + #3723 tune, LTC
ClaudeCodasearch/tt-refined-prune-eval-tuneddiff10.0+0.10LLR: -2.96 (-2.94, 2.94) [0.00, 2.00]
Games: 8148 W: 1871 L: 2125 D: 4152
Ptnml(0-2): 35, 1096, 2051, 872, 20
TT-refined prune eval + #3723 tune, STC
ClaudeCodatune/main-3723-values-controldiff10.0+0.10LLR: -2.95 (-2.94, 2.94) [-1.50, 1.50]
Games: 11474 W: 2684 L: 2857 D: 5933
Ptnml(0-2): 34, 1435, 2958, 1290, 20
control: #3723 values on main, no TT refine