Files
natural-memory-nm21/router_verdict_prod.json
WpyQwq 643e22ecb9 Natural Memory NM2.1: 记忆路由器分叉、数据集缺陷修复与全轴评测证据
- 引入 MemoryRouterXL 与 v5/v6 流式多线程训练/编码管线
- 修复 prepare_memory_router_dataset 候选池重建缺陷(mega 家族 3568x 加速,输出逐字节相同)
- 修复 v5 被破坏的拒答与多跳标签(train 未知样本 319 -> 16319,multi_hop 平均正例 1.00 -> 2.00)
- 同存储预算下 V2-128 v6 逐轴 22/22 通过:Top-1 41.12% -> 94.62%,未知拒答 0.00% -> 100.00%
- 记录三条被实测推翻的显然优化(logits_to_keep=1 反而慢 55%、XL 容量未带来收益)
- 记忆手术跨架构可移植性 14/14,读写关闭时与原生模型逐位相同
2026-09-19 11:11:31 +08:00

1107 lines
34 KiB
JSON

{
"candidates": [
"REPLAY-128 v7",
"PROD-128 step20000",
"PROD-128 step10000",
"V2-128 v6 best"
],
"baselines": [
"V2-128 deployed(v3)"
],
"passed_axes": 82,
"failed_axes": 6,
"skipped_axes": 0,
"dominates": false,
"per_candidate": {
"REPLAY-128 v7": {
"passed": 22,
"failed": 0,
"dominates": true,
"failed_axes": []
},
"PROD-128 step20000": {
"passed": 22,
"failed": 0,
"dominates": true,
"failed_axes": []
},
"PROD-128 step10000": {
"passed": 18,
"failed": 4,
"dominates": false,
"failed_axes": [
"单查询延迟 ms (GPU)",
"路由 QPS (GPU, 单查询)",
"批量 QPS (GPU, batch=64)",
"批量 QPS (GPU, batch=256)"
]
},
"V2-128 v6 best": {
"passed": 20,
"failed": 2,
"dominates": false,
"failed_axes": [
"单查询延迟 ms (GPU)",
"路由 QPS (GPU, 单查询)"
]
}
},
"rows": [
{
"axis": "Top-1 正确率",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4112495789828225,
"candidate_value": 0.9413944088918828,
"delta": 0.5301448299090603,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 128.91072891072892
},
{
"axis": "Top-1 正确率",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4112495789828225,
"candidate_value": 0.9378017289772089,
"delta": 0.5265521499943864,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 128.03712803712804
},
{
"axis": "Top-1 正确率",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4112495789828225,
"candidate_value": 0.9380824070955428,
"delta": 0.5268328281127204,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 128.10537810537812
},
{
"axis": "Top-1 正确率",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4112495789828225,
"candidate_value": 0.946165936903559,
"delta": 0.5349163579207366,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 130.07098007098008
},
{
"axis": "Recall@1",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.37027057370607386,
"candidate_value": 0.8857920736499383,
"delta": 0.5155214999438644,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 139.22832019405703
},
{
"axis": "Recall@1",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.37027057370607386,
"candidate_value": 0.8834343774559336,
"delta": 0.5131638037498597,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 138.59157064887813
},
{
"axis": "Recall@1",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.37027057370607386,
"candidate_value": 0.8837711911979342,
"delta": 0.5135006174918604,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 138.68253486961797
},
{
"axis": "Recall@1",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.37027057370607386,
"candidate_value": 0.8919108566296171,
"delta": 0.5216402829235433,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 140.8808368708308
},
{
"axis": "Recall@3",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4673009992141013,
"candidate_value": 0.9647748961490962,
"delta": 0.4974738969349949,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 106.45684425491018
},
{
"axis": "Recall@3",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4673009992141013,
"candidate_value": 0.9658976086224318,
"delta": 0.4985966094083305,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 106.69709892486034
},
{
"axis": "Recall@3",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4673009992141013,
"candidate_value": 0.9677500842034354,
"delta": 0.5004490849893342,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 107.09351913027807
},
{
"axis": "Recall@3",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4673009992141013,
"candidate_value": 0.9780509711462895,
"delta": 0.5107499719321882,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 109.2978557270707
},
{
"axis": "Recall@5",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4891377568204783,
"candidate_value": 0.9714550353654429,
"delta": 0.4823172785449646,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 98.60561198140815
},
{
"axis": "Recall@5",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4891377568204783,
"candidate_value": 0.9732794431346132,
"delta": 0.48414168631413496,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 98.97859643082573
},
{
"axis": "Recall@5",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4891377568204783,
"candidate_value": 0.9777702930279556,
"delta": 0.4886325362074773,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 99.89671199862283
},
{
"axis": "Recall@5",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4891377568204783,
"candidate_value": 0.984001347254968,
"delta": 0.49486359043448974,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 101.1705973489413
},
{
"axis": "MRR",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.476908110965675,
"candidate_value": 0.957689277532011,
"delta": 0.48078116656633607,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 100.81211778781002
},
{
"axis": "MRR",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.476908110965675,
"candidate_value": 0.9559179074946493,
"delta": 0.47900979652897435,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 100.44068983415772
},
{
"axis": "MRR",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.476908110965675,
"candidate_value": 0.9570269878646712,
"delta": 0.48011887689899624,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 100.67324624167533
},
{
"axis": "MRR",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.476908110965675,
"candidate_value": 0.9648813461092199,
"delta": 0.48797323514354496,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 102.32017949023022
},
{
"axis": "nDCG@3",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.44220238482445295,
"candidate_value": 0.9536787031890833,
"delta": 0.5114763183646304,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 115.6656625829094
},
{
"axis": "nDCG@3",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.44220238482445295,
"candidate_value": 0.9538607332074894,
"delta": 0.5116583483830365,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 115.70682699645694
},
{
"axis": "nDCG@3",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.44220238482445295,
"candidate_value": 0.9552694474506009,
"delta": 0.513067062626148,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 116.02539475896927
},
{
"axis": "nDCG@3",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.44220238482445295,
"candidate_value": 0.9649494672503449,
"delta": 0.522747082425892,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 118.21444215716157
},
{
"axis": "多跳证据全中(Top-3)",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.43426518468620184,
"candidate_value": 0.9622207252722578,
"delta": 0.5279555405860559,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 121.57445708376422
},
{
"axis": "多跳证据全中(Top-3)",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.43426518468620184,
"candidate_value": 0.96497137083193,
"delta": 0.5307061861457281,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 122.20785935884177
},
{
"axis": "多跳证据全中(Top-3)",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.43426518468620184,
"candidate_value": 0.9670483889076008,
"delta": 0.5327832042213989,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 122.68614270941056
},
{
"axis": "多跳证据全中(Top-3)",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.43426518468620184,
"candidate_value": 0.9774896149096216,
"delta": 0.5432244302234197,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 125.09048603929678
},
{
"axis": "多跳证据全中(仅多正例)",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.3715,
"candidate_value": 0.9545,
"delta": 0.583,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 156.93135935397038
},
{
"axis": "多跳证据全中(仅多正例)",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.3715,
"candidate_value": 0.9835,
"delta": 0.6120000000000001,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 164.7375504710633
},
{
"axis": "多跳证据全中(仅多正例)",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.3715,
"candidate_value": 0.9875,
"delta": 0.6160000000000001,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 165.814266487214
},
{
"axis": "多跳证据全中(仅多正例)",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.3715,
"candidate_value": 0.99,
"delta": 0.6185,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 166.4872139973082
},
{
"axis": "hop 正确率",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.7322992700729927,
"candidate_value": 1.0,
"delta": 0.26770072992700733,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 36.556192374781965
},
{
"axis": "hop 正确率",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.7322992700729927,
"candidate_value": 1.0,
"delta": 0.26770072992700733,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 36.556192374781965
},
{
"axis": "hop 正确率",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.7322992700729927,
"candidate_value": 1.0,
"delta": 0.26770072992700733,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 36.556192374781965
},
{
"axis": "hop 正确率",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.7322992700729927,
"candidate_value": 1.0,
"delta": 0.26770072992700733,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 36.556192374781965
},
{
"axis": "hop 欠预测率",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.046817110138093636,
"candidate_value": 0.0,
"delta": -0.046817110138093636,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -100.0
},
{
"axis": "hop 欠预测率",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.046817110138093636,
"candidate_value": 0.0,
"delta": -0.046817110138093636,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -100.0
},
{
"axis": "hop 欠预测率",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.046817110138093636,
"candidate_value": 0.0,
"delta": -0.046817110138093636,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -100.0
},
{
"axis": "hop 欠预测率",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.046817110138093636,
"candidate_value": 0.0,
"delta": -0.046817110138093636,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -100.0
},
{
"axis": "need F1 (门槛 0.50)",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.895773512669387,
"candidate_value": 1.0,
"delta": 0.10422648733061302,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 11.635361601619609
},
{
"axis": "need F1 (门槛 0.50)",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.895773512669387,
"candidate_value": 1.0,
"delta": 0.10422648733061302,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 11.635361601619609
},
{
"axis": "need F1 (门槛 0.50)",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.895773512669387,
"candidate_value": 1.0,
"delta": 0.10422648733061302,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 11.635361601619609
},
{
"axis": "need F1 (门槛 0.50)",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.895773512669387,
"candidate_value": 1.0,
"delta": 0.10422648733061302,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 11.635361601619609
},
{
"axis": "need 召回 (门槛 0.50)",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.998203660042663,
"candidate_value": 1.0,
"delta": 0.0017963399573369676,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 0.1799572601507184
},
{
"axis": "need 召回 (门槛 0.50)",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.998203660042663,
"candidate_value": 1.0,
"delta": 0.0017963399573369676,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 0.1799572601507184
},
{
"axis": "need 召回 (门槛 0.50)",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.998203660042663,
"candidate_value": 1.0,
"delta": 0.0017963399573369676,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 0.1799572601507184
},
{
"axis": "need 召回 (门槛 0.50)",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.998203660042663,
"candidate_value": 1.0,
"delta": 0.0017963399573369676,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 0.1799572601507184
},
{
"axis": "未知拒答率 (门槛 0.50)",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.0,
"candidate_value": 1.0,
"delta": 1.0,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": null
},
{
"axis": "未知拒答率 (门槛 0.50)",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.0,
"candidate_value": 1.0,
"delta": 1.0,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": null
},
{
"axis": "未知拒答率 (门槛 0.50)",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.0,
"candidate_value": 1.0,
"delta": 1.0,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": null
},
{
"axis": "未知拒答率 (门槛 0.50)",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.0,
"candidate_value": 1.0,
"delta": 1.0,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": null
},
{
"axis": "已知问题被误拒率",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.001796339957336926,
"candidate_value": 0.0,
"delta": -0.001796339957336926,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -100.0
},
{
"axis": "已知问题被误拒率",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.001796339957336926,
"candidate_value": 0.0,
"delta": -0.001796339957336926,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -100.0
},
{
"axis": "已知问题被误拒率",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.001796339957336926,
"candidate_value": 0.0,
"delta": -0.001796339957336926,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -100.0
},
{
"axis": "已知问题被误拒率",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.001796339957336926,
"candidate_value": 0.0,
"delta": -0.001796339957336926,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -100.0
},
{
"axis": "未知问题被误读率",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 1.0,
"candidate_value": 0.0,
"delta": -1.0,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -100.0
},
{
"axis": "未知问题被误读率",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 1.0,
"candidate_value": 0.0,
"delta": -1.0,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -100.0
},
{
"axis": "未知问题被误读率",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 1.0,
"candidate_value": 0.0,
"delta": -1.0,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -100.0
},
{
"axis": "未知问题被误读率",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 1.0,
"candidate_value": 0.0,
"delta": -1.0,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -100.0
},
{
"axis": "仲裁准确率",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.8112226277372263,
"candidate_value": 1.0,
"delta": 0.1887773722627737,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 23.27072320323923
},
{
"axis": "仲裁准确率",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.8112226277372263,
"candidate_value": 1.0,
"delta": 0.1887773722627737,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 23.27072320323923
},
{
"axis": "仲裁准确率",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.8112226277372263,
"candidate_value": 1.0,
"delta": 0.1887773722627737,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 23.27072320323923
},
{
"axis": "仲裁准确率",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.8112226277372263,
"candidate_value": 1.0,
"delta": 0.1887773722627737,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 23.27072320323923
},
{
"axis": "平均分数余量",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": -9.29892828391031,
"candidate_value": 40.14557690149532,
"delta": 49.44450518540563,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -531.7226208847976
},
{
"axis": "平均分数余量",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": -9.29892828391031,
"candidate_value": 21.892463866317613,
"delta": 31.191392150227923,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -335.4299678189535
},
{
"axis": "平均分数余量",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": -9.29892828391031,
"candidate_value": 18.42186796360512,
"delta": 27.72079624751543,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -298.1074313206608
},
{
"axis": "平均分数余量",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": -9.29892828391031,
"candidate_value": 14.959203617273067,
"delta": 24.258131901183376,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -260.87019020413976
},
{
"axis": "单查询延迟 ms (GPU)",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.9698253340320662,
"candidate_value": 0.9359393330911795,
"delta": -0.033886000940886674,
"direction": "min",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": -3.4940313221149855
},
{
"axis": "单查询延迟 ms (GPU)",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.9698253340320662,
"candidate_value": 0.9775833330544023,
"delta": 0.007757999022336093,
"direction": "min",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 0.799937756841438
},
{
"axis": "单查询延迟 ms (GPU)",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.9698253340320662,
"candidate_value": 1.2315193335719716,
"delta": 0.26169399953990546,
"direction": "min",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": 26.983621726183205
},
{
"axis": "单查询延迟 ms (GPU)",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.9698253340320662,
"candidate_value": 1.2807459999263908,
"delta": 0.31092066589432465,
"direction": "min",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": 32.0594497775869
},
{
"axis": "路由 QPS (GPU, 单查询)",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 1031.1135056067076,
"candidate_value": 1068.445319738026,
"delta": 37.331814131318424,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 3.62053391099289
},
{
"axis": "路由 QPS (GPU, 单查询)",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 1031.1135056067076,
"candidate_value": 1022.9306967371857,
"delta": -8.182808869521864,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": -0.7935895345204595
},
{
"axis": "路由 QPS (GPU, 单查询)",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 1031.1135056067076,
"candidate_value": 812.0051165575621,
"delta": -219.10838904914544,
"direction": "max",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": -21.249686659881544
},
{
"axis": "路由 QPS (GPU, 单查询)",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 1031.1135056067076,
"candidate_value": 780.7949429921887,
"delta": -250.3185626145189,
"direction": "max",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": -24.276528360205248
},
{
"axis": "批量 QPS (GPU, batch=64)",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 64355.240952926746,
"candidate_value": 65927.72676538704,
"delta": 1572.4858124602906,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 2.443446390963714
},
{
"axis": "批量 QPS (GPU, batch=64)",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 64355.240952926746,
"candidate_value": 73335.62500742644,
"delta": 8980.384054499693,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 13.954394267699938
},
{
"axis": "批量 QPS (GPU, batch=64)",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 64355.240952926746,
"candidate_value": 49115.91353738528,
"delta": -15239.32741554147,
"direction": "max",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": -23.68000987936386
},
{
"axis": "批量 QPS (GPU, batch=64)",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 64355.240952926746,
"candidate_value": 62429.27930602223,
"delta": -1925.961646904514,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": -2.9927036530144875
},
{
"axis": "批量 QPS (GPU, batch=256)",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 185281.68609572947,
"candidate_value": 262171.5169506191,
"delta": 76889.83085488965,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 41.498883389458676
},
{
"axis": "批量 QPS (GPU, batch=256)",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 185281.68609572947,
"candidate_value": 273837.7939855633,
"delta": 88556.1078898338,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 47.795391846811846
},
{
"axis": "批量 QPS (GPU, batch=256)",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 185281.68609572947,
"candidate_value": 141564.73262936738,
"delta": -43716.95346636209,
"direction": "max",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": -23.59485947454885
},
{
"axis": "批量 QPS (GPU, batch=256)",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 185281.68609572947,
"candidate_value": 201612.9031802962,
"delta": 16331.217084566742,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 8.814264069320318
},
{
"axis": "地址字节/记录",
"candidate": "REPLAY-128 v7",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 512,
"candidate_value": 512,
"delta": 0,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 0.0
},
{
"axis": "地址字节/记录",
"candidate": "PROD-128 step20000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 512,
"candidate_value": 512,
"delta": 0,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 0.0
},
{
"axis": "地址字节/记录",
"candidate": "PROD-128 step10000",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 512,
"candidate_value": 512,
"delta": 0,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 0.0
},
{
"axis": "地址字节/记录",
"candidate": "V2-128 v6 best",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 512,
"candidate_value": 512,
"delta": 0,
"direction": "min",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 0.0
}
]
}