OpenBench Testing Framework

Finished
ClaudeCodatb/two-tier-probediff10.0+0.10LLR: 2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 161840 W: 39962 L: 39478 D: 82400
Ptnml(0-2): 418, 17914, 43786, 18370, 432
STC companion to 3707. Two-tier TB probe (Thor 2c19a98): pure NPS change, search bit-identical at fixed nodes, so STC is the correct instrument not LTC. TB REQUIRED.
ClaudeCodamaindiff40.0+0.40Tuning 49 Parameters
25000/25000 Iterations
100002/100000 Games Played
pruning clusters at LTC (3701 split STC +1.2 / LTC -2.4)
ClaudeCodasearch/tt-refined-prune-evaldiff10.0+0.10Tuning 9 Parameters
25002/25000 Iterations
100008/100000 Games Played
RFP/razor retune on TT-refined prune eval, #3683 shape
ClaudeCodatb/two-tier-probediff40.0+0.40LLR: 0.43 (-2.94, 2.94) [0.00, 2.00]
Games: 77854 W: 18729 L: 18592 D: 40533
Ptnml(0-2): 45, 7951, 22805, 8074, 52
Two-tier TB probe: skip the DTZ lookup on interior probes that cannot need it (Thor 2c19a98). Search bit-identical at fixed nodes, -20% probe time in TB range. TB REQUIRED. Resubmit of 3706 which lost its increment to + decoding.
ClaudeCodasearch/rfp-lesser-threatsdiff10.0+0.10LLR: -2.97 (-2.94, 2.94) [0.00, 2.00]
Games: 74490 W: 18128 L: 18282 D: 38080
Ptnml(0-2): 216, 8988, 18970, 8876, 195
RFP widen on any cheaper-attacker threat
ClaudeCodatm/phase-entry-creditdiff10.0+0.10LLR: -2.95 (-2.94, 2.94) [-1.50, 0.50]
Games: 71772 W: 17462 L: 17772 D: 36538
Ptnml(0-2): 197, 8478, 18844, 8172, 195
TM phase-damp late-entry credit=10, non-regression. Caveats: cap binds in 4.0% of RR games (RR validates the credit, not the capped form); RR Elo unmeasurable at 94-99% draws, so this is the only Elo evidence
ClaudeCodatb/dtz-hash-cachediff10.0+0.10LLR: -2.98 (-2.94, 2.94) [0.00, 2.00]
Games: 49762 W: 12106 L: 12290 D: 25366
Ptnml(0-2): 148, 5547, 13649, 5415, 122
Hash-keyed DTZ cache ON TOP OF two-tier (Thor 0d10c8b). Base is tb/two-tier-probe, so this isolates the increment. Search bit-identical at fixed nodes; -26.3% time in won endgames, -0.8% drawn. TB REQUIRED.
ClaudeCodasearch/rfp-return-half-margindiff10.0+0.10LLR: -2.97 (-2.94, 2.94) [0.00, 2.00]
Games: 21256 W: 5112 L: 5344 D: 10800
Ptnml(0-2): 63, 2641, 5439, 2435, 50
RFP return eval - margin/2
ClaudeCodasearch/rfp-opp-worseningdiff10.0+0.10LLR: -2.96 (-2.94, 2.94) [0.00, 2.00]
Games: 13242 W: 3082 L: 3329 D: 6831
Ptnml(0-2): 48, 1728, 3304, 1505, 36
RFP opp-worsening margin relief (retry)
ClaudeCodaresearch/corr-apply-fadediff10.0+0.10LLR: -2.96 (-2.94, 2.94) [0.00, 2.00]
Games: 4870 W: 1034 L: 1290 D: 2546
Ptnml(0-2): 26, 683, 1255, 463, 8
corrhist apply-by-ply fade (default 16): offline closes 5/5 verified-drawn gap positions to corrhist-OFF, middlegame control at noise floor. EGTB off: draw-recognition mechanism, don't let adjudication mask it.
ClaudeCodasearch/tt-refined-prune-evaldiff10.0+0.10Tuning 9 Parameters
0/1500 Iterations
0/24000 Games Played
RFP/razor retune on TT-refined prune eval
ClaudeCodamaindiff80.0+0.80LLR: 2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 3840 W: 1024 L: 767 D: 2049
Ptnml(0-2): 2, 324, 1012, 579, 3
HEAD vs v0.9.4, 2xLTC 80+0.8 (Hash 256)
ClaudeCodaresearch/corr-draw-guarddiff10.0+0.10LLR: 0.24 (-2.94, 2.94) [0.00, 2.00]
Games: 10340 W: 2594 L: 2557 D: 5189
Ptnml(0-2): 29, 1220, 2635, 1257, 29
corrhist: don't train on exact draw returns (repetition poisons the tables)
ClaudeCodamaindiff40.0+0.40LLR: 2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 4258 W: 1147 L: 878 D: 2233
Ptnml(0-2): 4, 406, 1041, 673, 5
HEAD vs v0.9.4, LTC
ClaudeCodamaindiff10.0+0.10LLR: 2.96 (-2.94, 2.94) [0.00, 2.00]
Games: 2324 W: 727 L: 437 D: 1160
Ptnml(0-2): 6, 158, 548, 440, 10
HEAD vs v0.9.4, STC T=4
ClaudeCodamaindiff10.0+0.10LLR: 2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 4378 W: 1268 L: 991 D: 2119
Ptnml(0-2): 16, 392, 1102, 657, 22
HEAD vs v0.9.4, STC
ClaudeCodatune/main-3701diff10.0+0.10LLR: 2.97 (-2.94, 2.94) [0.00, 2.00]
Games: 133466 W: 32966 L: 32504 D: 67996
Ptnml(0-2): 347, 15941, 33724, 16345, 376
tune #3701 applied (RFP cluster post-verify), STC
ClaudeCodatune/main-3701diff40.0+0.40LLR: -2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 30506 W: 7243 L: 7450 D: 15813
Ptnml(0-2): 19, 3645, 8136, 3430, 23
tune #3701 applied, LTC
ClaudeCodasmp/shared-root-instabilitydiff10.0+0.10LLR: -2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 14864 W: 3467 L: 3696 D: 7701
Ptnml(0-2): 17, 1813, 3995, 1596, 11
main asp window widened by helpers' root fails, q=2, T=4
ClaudeCodasmp/shared-root-instabilitydiff10.0+0.10LLR: -2.95 (-2.94, 2.94) [0.00, 2.00]
Games: 7438 W: 1704 L: 1946 D: 3788
Ptnml(0-2): 11, 982, 1970, 750, 6
main asp window widened by helpers' root fails, q=4, T=4
ClaudeCodatb/two-tier-probediff40.0+0.00LLR: 0.00 (-2.94, 2.94) [0.00, 2.00]
Games: 0 W: 0 L: 0 D: 0
Ptnml(0-2): 0, 0, 0, 0, 0
Two-tier TB probe: skip the DTZ lookup on interior probes that cannot need it (Thor 2c19a98). Search bit-identical at fixed nodes, -20% probe time. TB REQUIRED.
ClaudeCodamaindiff10.0+0.10Tuning 49 Parameters
25000/25000 Iterations
100000/100000 Games Played
pruning clusters (LMR/FUT/LMP/NMP/SEE/PC, 49 knobs) retuned against the new RFP regime
ClaudeCodaresearch/corr-cycle-drawdiff40.0+0.40LLR: -2.96 (-2.94, 2.94) [0.00, 2.00]
Games: 166584 W: 39856 L: 39872 D: 86856
Ptnml(0-2): 99, 16915, 49279, 16901, 98
LTC re-test of STC #3669 (-0.2)
AtlasCodanet-20B09859diff40.0+0.40LLR: -2.95 (-2.94, 2.94) [0.00, 3.00]
Games: 86266 W: 20610 L: 20595 D: 45061
Ptnml(0-2): 60, 10051, 22888, 10082, 52
Third resumed decay tail on the pawn-pair lineage, testing whether the tail gain stacks a third time or the recipe settles at two stages.
AtlasCodanet-20B09859diff10.0+0.10LLR: -2.95 (-2.94, 2.94) [0.00, 3.00]
Games: 303330 W: 74345 L: 73849 D: 155136
Ptnml(0-2): 851, 36178, 77088, 36720, 828
Third resumed decay tail on the pawn-pair lineage, testing whether the tail gain stacks a third time or the recipe settles at two stages.