Files
natural-memory-nm21/router_verdict_final.json
T
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

1384 lines
43 KiB
JSON

{
"candidates": [
"V2-128 v6 final",
"REPLAY-128 v7 final",
"V2-512 v6 final",
"XL-512 v6 final",
"XL-128 v6 final"
],
"baselines": [
"V2-128 deployed(v3)"
],
"passed_axes": 99,
"failed_axes": 11,
"skipped_axes": 0,
"dominates": false,
"per_candidate": {
"V2-128 v6 final": {
"passed": 22,
"failed": 0,
"dominates": true,
"failed_axes": []
},
"REPLAY-128 v7 final": {
"passed": 22,
"failed": 0,
"dominates": true,
"failed_axes": []
},
"V2-512 v6 final": {
"passed": 20,
"failed": 2,
"dominates": false,
"failed_axes": [
"批量 QPS (GPU, batch=256)",
"地址字节/记录"
]
},
"XL-512 v6 final": {
"passed": 17,
"failed": 5,
"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)",
"地址字节/记录"
]
},
"XL-128 v6 final": {
"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)"
]
}
},
"rows": [
{
"axis": "Top-1 正确率",
"candidate": "V2-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4112495789828225,
"candidate_value": 0.9436959694622207,
"delta": 0.5324463904793981,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 129.47037947037944
},
{
"axis": "Top-1 正确率",
"candidate": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4112495789828225,
"candidate_value": 0.941843493881217,
"delta": 0.5305939148983945,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 129.019929019929
},
{
"axis": "Top-1 正确率",
"candidate": "XL-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4112495789828225,
"candidate_value": 0.9502077018075671,
"delta": 0.5389581228247446,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 131.05378105378108
},
{
"axis": "Top-1 正确率",
"candidate": "XL-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4112495789828225,
"candidate_value": 0.9518917705175705,
"delta": 0.5406421915347479,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 131.46328146328145
},
{
"axis": "Recall@1",
"candidate": "V2-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.37027057370607386,
"candidate_value": 0.8882059054676097,
"delta": 0.5179353317615358,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 139.88023044269252
},
{
"axis": "Recall@1",
"candidate": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.37027057370607386,
"candidate_value": 0.8864657011339396,
"delta": 0.5161951274278658,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 139.4102486355367
},
{
"axis": "Recall@1",
"candidate": "XL-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.37027057370607386,
"candidate_value": 0.8948299090602897,
"delta": 0.5245593353542158,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 141.66919345057613
},
{
"axis": "Recall@1",
"candidate": "XL-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.37027057370607386,
"candidate_value": 0.8963736387111261,
"delta": 0.5261030650050522,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 142.0861127956337
},
{
"axis": "Recall@3",
"candidate": "V2-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4673009992141013,
"candidate_value": 0.9639609296059279,
"delta": 0.49665993039182665,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 106.28265961919634
},
{
"axis": "Recall@3",
"candidate": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4673009992141013,
"candidate_value": 0.9666273717300999,
"delta": 0.49932637251599865,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 106.85326446032795
},
{
"axis": "Recall@3",
"candidate": "XL-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4673009992141013,
"candidate_value": 0.965336252385764,
"delta": 0.49803525317166275,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 106.57697158988529
},
{
"axis": "Recall@3",
"candidate": "XL-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4673009992141013,
"candidate_value": 0.9745424946671157,
"delta": 0.5072414954530144,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 108.54705988347648
},
{
"axis": "Recall@5",
"candidate": "V2-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4891377568204783,
"candidate_value": 0.973981138430448,
"delta": 0.4848433816099697,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 99.12205198829403
},
{
"axis": "Recall@5",
"candidate": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4891377568204783,
"candidate_value": 0.9764230380599529,
"delta": 0.4872852812394746,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 99.62127732828371
},
{
"axis": "Recall@5",
"candidate": "XL-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4891377568204783,
"candidate_value": 0.9760862243179522,
"delta": 0.48694846749747395,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 99.55241866069893
},
{
"axis": "Recall@5",
"candidate": "XL-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.4891377568204783,
"candidate_value": 0.9820366004266308,
"delta": 0.4928988436061525,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 100.76892178803007
},
{
"axis": "MRR",
"candidate": "V2-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.476908110965675,
"candidate_value": 0.9597054333626427,
"delta": 0.4827973223969677,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 101.23487340556399
},
{
"axis": "MRR",
"candidate": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.476908110965675,
"candidate_value": 0.958724954839147,
"delta": 0.481816843873472,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 101.02928274754177
},
{
"axis": "MRR",
"candidate": "XL-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.476908110965675,
"candidate_value": 0.9664213223997663,
"delta": 0.4895132114340913,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 102.64308787764014
},
{
"axis": "MRR",
"candidate": "XL-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.476908110965675,
"candidate_value": 0.9670023550641874,
"delta": 0.49009424409851243,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 102.76492113043273
},
{
"axis": "nDCG@3",
"candidate": "V2-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.44220238482445295,
"candidate_value": 0.9530820526370456,
"delta": 0.5108796678125926,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 115.53073555118962
},
{
"axis": "nDCG@3",
"candidate": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.44220238482445295,
"candidate_value": 0.9548399733910118,
"delta": 0.5126375885665588,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 115.92827315258995
},
{
"axis": "nDCG@3",
"candidate": "XL-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.44220238482445295,
"candidate_value": 0.9560053718886551,
"delta": 0.5138029870642021,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 116.19181729835614
},
{
"axis": "nDCG@3",
"candidate": "XL-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.44220238482445295,
"candidate_value": 0.9642642772784563,
"delta": 0.5220618924540034,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 118.05949275041866
},
{
"axis": "多跳证据全中(Top-3)",
"candidate": "V2-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.43426518468620184,
"candidate_value": 0.9595262153362524,
"delta": 0.5252610306500505,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 120.95398138572907
},
{
"axis": "多跳证据全中(Top-3)",
"candidate": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.43426518468620184,
"candidate_value": 0.9644100145952621,
"delta": 0.5301448299090603,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 122.07859358841779
},
{
"axis": "多跳证据全中(Top-3)",
"candidate": "XL-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.43426518468620184,
"candidate_value": 0.9534635679802402,
"delta": 0.5191983832940383,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 119.55791106514995
},
{
"axis": "多跳证据全中(Top-3)",
"candidate": "XL-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.43426518468620184,
"candidate_value": 0.9728864937689458,
"delta": 0.5386213090827439,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 124.03050672182006
},
{
"axis": "多跳证据全中(仅多正例)",
"candidate": "V2-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.3715,
"candidate_value": 0.921,
"delta": 0.5495000000000001,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 147.91386271870797
},
{
"axis": "多跳证据全中(仅多正例)",
"candidate": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.3715,
"candidate_value": 0.9605,
"delta": 0.589,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 158.5464333781965
},
{
"axis": "多跳证据全中(仅多正例)",
"candidate": "XL-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.3715,
"candidate_value": 0.7885,
"delta": 0.417,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 112.24764468371467
},
{
"axis": "多跳证据全中(仅多正例)",
"candidate": "XL-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 0.3715,
"candidate_value": 0.9705,
"delta": 0.599,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": 161.23822341857334
},
{
"axis": "hop 正确率",
"candidate": "V2-128 v6 final",
"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": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"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": "XL-512 v6 final",
"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": "XL-128 v6 final",
"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": "V2-128 v6 final",
"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": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"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": "XL-512 v6 final",
"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": "XL-128 v6 final",
"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": "V2-128 v6 final",
"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": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"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": "XL-512 v6 final",
"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": "XL-128 v6 final",
"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": "V2-128 v6 final",
"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": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"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": "XL-512 v6 final",
"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": "XL-128 v6 final",
"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": "V2-128 v6 final",
"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": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"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": "XL-512 v6 final",
"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": "XL-128 v6 final",
"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": "V2-128 v6 final",
"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": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"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": "XL-512 v6 final",
"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": "XL-128 v6 final",
"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": "V2-128 v6 final",
"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": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"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": "XL-512 v6 final",
"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": "XL-128 v6 final",
"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": "V2-128 v6 final",
"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": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"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": "XL-512 v6 final",
"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": "XL-128 v6 final",
"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": "V2-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": -9.29892828391031,
"candidate_value": 51.315328869764656,
"delta": 60.61425715367497,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -651.8413230323995
},
{
"axis": "平均分数余量",
"candidate": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": -9.29892828391031,
"candidate_value": 73.24820388849786,
"delta": 82.54713217240817,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -887.7058694520451
},
{
"axis": "平均分数余量",
"candidate": "XL-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": -9.29892828391031,
"candidate_value": 215.75331382084113,
"delta": 225.05224210475143,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -2420.1954809583094
},
{
"axis": "平均分数余量",
"candidate": "XL-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": -9.29892828391031,
"candidate_value": 200.33457717921095,
"delta": 209.63350546312125,
"direction": "max",
"pass": true,
"relative_tolerance_applied": false,
"relative_delta_pct": -2254.3835059556764
},
{
"axis": "单查询延迟 ms (GPU)",
"candidate": "V2-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 1.0995226665302955,
"candidate_value": 0.9277419995729966,
"delta": -0.17178066695729888,
"direction": "min",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": -15.62320379436814
},
{
"axis": "单查询延迟 ms (GPU)",
"candidate": "REPLAY-128 v7 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 1.0995226665302955,
"candidate_value": 0.9663366669459114,
"delta": -0.13318599958438415,
"direction": "min",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": -12.11307448573771
},
{
"axis": "单查询延迟 ms (GPU)",
"candidate": "V2-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 1.0995226665302955,
"candidate_value": 0.9488446664181538,
"delta": -0.15067800011214172,
"direction": "min",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": -13.703946694216699
},
{
"axis": "单查询延迟 ms (GPU)",
"candidate": "XL-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 1.0995226665302955,
"candidate_value": 1.450062666942055,
"delta": 0.3505400004117596,
"direction": "min",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": 31.88110723701038
},
{
"axis": "单查询延迟 ms (GPU)",
"candidate": "XL-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 1.0995226665302955,
"candidate_value": 1.5358359999663662,
"delta": 0.4363133334360707,
"direction": "min",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": 39.68206811169443
},
{
"axis": "路由 QPS (GPU, 单查询)",
"candidate": "V2-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 909.4855708210602,
"candidate_value": 1077.8858782509155,
"delta": 168.4003074298554,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 18.515995506979614
},
{
"axis": "路由 QPS (GPU, 单查询)",
"candidate": "REPLAY-128 v7 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 909.4855708210602,
"candidate_value": 1034.8360299319709,
"delta": 125.35045911091072,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 13.782567105242533
},
{
"axis": "路由 QPS (GPU, 单查询)",
"candidate": "V2-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 909.4855708210602,
"candidate_value": 1053.9132856961248,
"delta": 144.42771487506468,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 15.880154618030835
},
{
"axis": "路由 QPS (GPU, 单查询)",
"candidate": "XL-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 909.4855708210602,
"candidate_value": 689.6253677841637,
"delta": -219.86020303689645,
"direction": "max",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": -24.174127670702056
},
{
"axis": "路由 QPS (GPU, 单查询)",
"candidate": "XL-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 909.4855708210602,
"candidate_value": 651.1111863648849,
"delta": -258.3743844561752,
"direction": "max",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": -28.40884921603775
},
{
"axis": "批量 QPS (GPU, batch=64)",
"candidate": "V2-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 66037.93061957207,
"candidate_value": 72856.42735601631,
"delta": 6818.496736444242,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 10.325121748172105
},
{
"axis": "批量 QPS (GPU, batch=64)",
"candidate": "REPLAY-128 v7 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 66037.93061957207,
"candidate_value": 69378.19790068052,
"delta": 3340.2672811084485,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 5.058104107396837
},
{
"axis": "批量 QPS (GPU, batch=64)",
"candidate": "V2-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 66037.93061957207,
"candidate_value": 71586.76538686478,
"delta": 5548.834767292705,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 8.402496436870726
},
{
"axis": "批量 QPS (GPU, batch=64)",
"candidate": "XL-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 66037.93061957207,
"candidate_value": 46891.9433921584,
"delta": -19145.987227413672,
"direction": "max",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": -28.9924094346761
},
{
"axis": "批量 QPS (GPU, batch=64)",
"candidate": "XL-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 66037.93061957207,
"candidate_value": 33111.55489890671,
"delta": -32926.37572066536,
"direction": "max",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": -49.85979332142605
},
{
"axis": "批量 QPS (GPU, batch=256)",
"candidate": "V2-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 207287.4493168988,
"candidate_value": 236293.15117731004,
"delta": 29005.701860411238,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 13.992985082308401
},
{
"axis": "批量 QPS (GPU, batch=256)",
"candidate": "REPLAY-128 v7 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 207287.4493168988,
"candidate_value": 242759.87633474605,
"delta": 35472.42701784725,
"direction": "max",
"pass": true,
"relative_tolerance_applied": true,
"relative_delta_pct": 17.112674759009355
},
{
"axis": "批量 QPS (GPU, batch=256)",
"candidate": "V2-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 207287.4493168988,
"candidate_value": 177958.2077420389,
"delta": -29329.241574859887,
"direction": "max",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": -14.149067718046771
},
{
"axis": "批量 QPS (GPU, batch=256)",
"candidate": "XL-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 207287.4493168988,
"candidate_value": 85734.57110724591,
"delta": -121552.87820965289,
"direction": "max",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": -58.639767438994426
},
{
"axis": "批量 QPS (GPU, batch=256)",
"candidate": "XL-128 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 207287.4493168988,
"candidate_value": 131545.14154150538,
"delta": -75742.30777539342,
"direction": "max",
"pass": false,
"relative_tolerance_applied": true,
"relative_delta_pct": -36.539746147196496
},
{
"axis": "地址字节/记录",
"candidate": "V2-128 v6 final",
"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": "REPLAY-128 v7 final",
"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": "V2-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 512,
"candidate_value": 2048,
"delta": 1536,
"direction": "min",
"pass": false,
"relative_tolerance_applied": false,
"relative_delta_pct": 300.0
},
{
"axis": "地址字节/记录",
"candidate": "XL-512 v6 final",
"baseline": "V2-128 deployed(v3)",
"baseline_value": 512,
"candidate_value": 2048,
"delta": 1536,
"direction": "min",
"pass": false,
"relative_tolerance_applied": false,
"relative_delta_pct": 300.0
},
{
"axis": "地址字节/记录",
"candidate": "XL-128 v6 final",
"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
}
]
}