| Finished | ||||
|---|---|---|---|---|
| Test | TC | Engine | Author | Result |
History pruning: non-PV, depth<=4, quiet non-killer with hist < -2048*depth | 4.0+0.04 | Solstice | Dragon | LLR: 0.00 (-2.94, 2.94) [0.00, 5.00] Games: 0 W: 0 L: 0 D: 0 Ptnml(0-2): 0, 0, 0, 0, 0 |
Singular extensions (depth>=8, sb = tt - 2d, (d-1)/2 search) + double ext (non-PV, 20 margin, max 6) + multicut + negative ext | 4.0+0.04 | Solstice | Dragon | LLR: 3.09 (-2.94, 2.94) [0.00, 5.00] Games: 1264 W: 379 L: 237 D: 648 Ptnml(0-2): 12, 108, 274, 202, 36 |
NMP dynamic R = 4 + depth/4 + min((eval-beta)/200, 3) | 4.0+0.04 | Solstice | Dragon | LLR: -3.06 (-2.94, 2.94) [0.00, 5.00] Games: 4344 W: 973 L: 1085 D: 2286 Ptnml(0-2): 91, 552, 970, 496, 63 |
net gen34_h512_cos40 vs gen4_h256_cos60: gen3+gen4 data (193.9M), hidden 512, 40-epoch cosine | 4.0+0.04 | Solstice | Dragon | LLR: -2.98 (-2.94, 2.94) [0.00, 5.00] Games: 1486 W: 318 L: 449 D: 719 Ptnml(0-2): 55, 210, 311, 145, 22 |
net gen4_h256_cos60 vs gen3_h256_cos60: gen4 data (104.5M, cos60 + current search), same arch/recipe | 4.0+0.04 | Solstice | Dragon | LLR: 3.00 (-2.94, 2.94) [0.00, 5.00] Games: 4818 W: 1351 L: 1193 D: 2274 Ptnml(0-2): 91, 541, 1024, 625, 128 |
net: cos60 + 10 epochs quantisation-aware fine-tuning (rounded-net val loss 0.46947 -> 0.46937) | 4.0+0.04 | Solstice | Dragon | LLR: -1.23 (-2.94, 2.94) [0.00, 5.00] Games: 1902 W: 449 L: 493 D: 960 Ptnml(0-2): 34, 244, 435, 208, 30 |
audit 2: no repetitions across null moves (pliesFromNull), 50-move rule vs checkmate, mate stop only when depth >= mate length | 4.0+0.04 | Solstice | Dragon | LLR: -0.39 (-2.94, 2.94) [-5.00, 0.00] Games: 9146 W: 2177 L: 2234 D: 4735 Ptnml(0-2): 153, 1130, 2062, 1077, 151 |
net gen3_h256_cos60: same data/arch, 60 epochs warmup+cosine vs 15 epochs step decay (val 0.4703 -> 0.4692) | 4.0+0.04 | Solstice | Dragon | LLR: 3.09 (-2.94, 2.94) [0.00, 5.00] Games: 926 W: 327 L: 173 D: 426 Ptnml(0-2): 10, 70, 177, 168, 38 |
speed: SIMD SCReLU output + SIMD accumulator updates, aligned accumulators, single-SEE qsearch, plain-net make, TT prefetch, in-check reuse, LMR log table, -DNDEBUG. Same search (bench identical); +49% nps Xeon, +31% laptop (non-regression) | 4.0+0.04 | Solstice | Dragon | LLR: 3.01 (-2.94, 2.94) [-5.00, 0.00] Games: 820 W: 265 L: 134 D: 421 Ptnml(0-2): 4, 63, 165, 154, 24 |
staged move picker (lazy TT/captures/killers/quiets/bad captures); ~10% nps at equal ordering (non-regression) | 4.0+0.04 | Solstice | Dragon | LLR: 2.97 (-2.94, 2.94) [-5.00, 0.00] Games: 3036 W: 770 L: 657 D: 1609 Ptnml(0-2): 43, 327, 674, 422, 52 |
Time management tune (7 params: base/inc/soft/hard constants + best-move stability + node-fraction scaling, both starting neutral) at 8+0.08 | 8.0+0.08 | Solstice | Dragon | Tuning 7 Parameters 488/1000 Iterations 15646/32000 Games Played |
continuation history 1-ply + 2-ply (ordering + LMR), retest of v17_conthist | 4.0+0.04 | Solstice | Dragon | LLR: -1.89 (-2.94, 2.94) [0.00, 5.00] Games: 5476 W: 1258 L: 1313 D: 2905 Ptnml(0-2): 80, 691, 1254, 630, 83 |
qsearch TT probe/cutoff + store (depth 0, never evicting main-search entries) | 4.0+0.04 | Solstice | Dragon | LLR: 0.00 (-2.94, 2.94) [0.00, 5.00] Games: 0 W: 0 L: 0 D: 0 Ptnml(0-2): 0, 0, 0, 0, 0 |
capture history (ordering within MVV-LVA tier, bonus/malus), retest of v24_capthist | 4.0+0.04 | Solstice | Dragon | LLR: 0.00 (-2.94, 2.94) [0.00, 5.00] Games: 0 W: 0 L: 0 D: 0 Ptnml(0-2): 0, 0, 0, 0, 0 |
SPSA tune #13 values (33 params, 128k games; first tune with LMR running) | 4.0+0.04 | Solstice | Dragon | LLR: 3.29 (-2.94, 2.94) [0.00, 5.00] Games: 2158 W: 634 L: 473 D: 1051 Ptnml(0-2): 31, 227, 444, 304, 73 |
Full retune after the 2026-09-28 fixes (LMR live, twofold, hist malus, x255 net): 17 search params + SEE/ordering piece values, history scale/cap, LMR killer, corrhist, aspiration growth, qsearch SEE (33 total) | 2.0+0.02 | Solstice | Dragon | Tuning 33 Parameters 2000/2000 Iterations 128028/128000 Games Played |
history malus retest (first tried ddd4e6f when LMR was dead) | 4.0+0.04 | Solstice | Dragon | LLR: 3.02 (-2.94, 2.94) [0.00, 5.00] Games: 7612 W: 2029 L: 1860 D: 3723 Ptnml(0-2): 141, 879, 1625, 992, 169 |
qsearch: no stand pat in check, mate when no evasion | 4.0+0.04 | Solstice | Dragon | LLR: -2.96 (-2.94, 2.94) [0.00, 5.00] Games: 24802 W: 6578 L: 6596 D: 11628 Ptnml(0-2): 584, 2998, 5255, 2980, 584 |
ply-adjusted TT mate scores + negamax ply guard (non-regression) | 4.0+0.04 | Solstice | Dragon | LLR: 2.97 (-2.94, 2.94) [-5.00, 0.00] Games: 27796 W: 7033 L: 7029 D: 13734 Ptnml(0-2): 601, 3267, 6090, 3407, 533 |
gen3-256 re-exported at ft_scale 255 (quantization loss 24 -> 11 raw) | 4.0+0.04 | Solstice | Dragon | LLR: 2.98 (-2.94, 2.94) [0.00, 5.00] Games: 3140 W: 909 L: 755 D: 1476 Ptnml(0-2): 59, 352, 640, 414, 105 |
twofold repetition draw in search (isRepetition(1)) | 4.0+0.04 | Solstice | Dragon | LLR: 3.08 (-2.94, 2.94) [0.00, 5.00] Games: 3040 W: 870 L: 716 D: 1454 Ptnml(0-2): 58, 315, 638, 433, 76 |
LMR actually runs now (was dead since ee65682: is_capture after make) + full-depth verification of reduced fail highs | 4.0+0.04 | Solstice | Dragon | LLR: 3.40 (-2.94, 2.94) [0.00, 10.00] Games: 336 W: 155 L: 54 D: 127 Ptnml(0-2): 3, 11, 66, 58, 30 |
NET: gen3_h256_l075 (256 wide) vs gen3_l075 (128), same data and search, equal time | 4.0+0.04 | Solstice | Dragon | LLR: 3.05 (-2.94, 2.94) [0.00, 10.00] Games: 4762 W: 1686 L: 1548 D: 1528 Ptnml(0-2): 219, 501, 846, 553, 262 |
NET: gen3_l075 (128, 89.3M gen3 positions) vs gen2 -- same v24 search | 4.0+0.04 | Solstice | Dragon | LLR: 3.04 (-2.94, 2.94) [0.00, 10.00] Games: 684 W: 271 L: 187 D: 226 Ptnml(0-2): 18, 45, 148, 97, 34 |
SPSA-tuned search defaults (tunes #1 + #3) vs master | 4.0+0.04 | Solstice | Dragon | LLR: 2.96 (-2.94, 2.94) [0.00, 10.00] Games: 3580 W: 1220 L: 1106 D: 1254 Ptnml(0-2): 118, 393, 715, 385, 179 |