{
 "B": 10000,
 "seed": 0,
 "comparisons": [
  {
   "comparison": "laptop: TITAN all-resident (64 slots, preloaded) vs mlx-lm",
   "n_a": 12,
   "n_b": 12,
   "median_a": 489.12865342395037,
   "ci_a": [
    488.2082407000963,
    489.5368828457765
   ],
   "median_b": 495.04061243824856,
   "ci_b": [
    494.1194679570178,
    495.8904708999346
   ],
   "diff": -5.911959014298191,
   "ci_diff": [
    -7.065204357666232,
    -4.834550062056792
   ],
   "rel_diff_pct": -1.1942371728209773,
   "unit": "tok/s",
   "source": "gateB/capacity_sweep_laptop/sweep.jsonl"
  },
  {
   "comparison": "laptop: TITAN 64 slots not preloaded (sync path) vs mlx-lm",
   "n_a": 12,
   "n_b": 12,
   "median_a": 201.4454551722675,
   "ci_a": [
    200.95837279071753,
    202.66232324191964
   ],
   "median_b": 495.04061243824856,
   "ci_b": [
    494.1194679570178,
    495.8904708999346
   ],
   "diff": -293.5951572659811,
   "ci_diff": [
    -294.4450157276671,
    -291.97897074353386
   ],
   "rel_diff_pct": -59.307287097097344,
   "unit": "tok/s",
   "source": "gateB/capacity_sweep_laptop/sweep.jsonl"
  },
  {
   "comparison": "M3U96a Qwen3-30B: TITAN all-resident (128 slots) vs mlx-lm",
   "n_a": 18,
   "n_b": 18,
   "median_a": 121.58726022750393,
   "ci_a": [
    121.4222850284662,
    121.7539110305407
   ],
   "median_b": 121.65810731258236,
   "ci_b": [
    121.37584628881791,
    121.8112669918118
   ],
   "diff": -0.07084708507842663,
   "ci_diff": [
    -0.27269819921461647,
    0.2832330343496068
   ],
   "rel_diff_pct": -0.058234577738740925,
   "unit": "tok/s",
   "source": "gateC/qwen3_m3u96a/baseline.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 16 slots: transition m=8 vs no prefetch",
   "n_a": 12,
   "n_b": 12,
   "median_a": 27.904324206563125,
   "ci_a": [
    26.510834259463596,
    30.51331379185264
   ],
   "median_b": 36.36478535014914,
   "ci_b": [
    34.13753044405526,
    42.04500706287061
   ],
   "diff": -8.460461143586016,
   "ci_diff": [
    -14.521560050815605,
    -5.0831251036633525
   ],
   "rel_diff_pct": -23.265533020811077,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/sweep.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 16 slots: transition2 m=8 vs no prefetch",
   "n_a": 12,
   "n_b": 12,
   "median_a": 34.96599880167474,
   "ci_a": [
    33.67628976796817,
    41.49003741051476
   ],
   "median_b": 36.36478535014914,
   "ci_b": [
    34.13753044405526,
    42.04500706287061
   ],
   "diff": -1.3987865484743978,
   "ci_diff": [
    -7.324854162049796,
    5.998514358174624
   ],
   "rel_diff_pct": -3.8465414686372155,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/sweep.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 16 slots: transition3 m=8 vs no prefetch",
   "n_a": 12,
   "n_b": 12,
   "median_a": 29.083938320259847,
   "ci_a": [
    27.671125828589567,
    34.70508437115666
   ],
   "median_b": 36.36478535014914,
   "ci_b": [
    34.13753044405526,
    42.04500706287061
   ],
   "diff": -7.280847029889294,
   "ci_diff": [
    -13.09924827054045,
    -1.0543321191199624
   ],
   "rel_diff_pct": -20.02169670405997,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/sweep.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 16 slots: transition2 m=16 vs no prefetch",
   "n_a": 12,
   "n_b": 12,
   "median_a": 25.940903169028186,
   "ci_a": [
    25.001461847324514,
    29.583536873911555
   ],
   "median_b": 36.36478535014914,
   "ci_b": [
    34.13753044405526,
    42.04500706287061
   ],
   "diff": -10.423882181120955,
   "ci_diff": [
    -16.145654899217902,
    -5.865959852413532
   ],
   "rel_diff_pct": -28.66477027363563,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/sweep.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 16 slots: prev_rank m=16 vs no prefetch",
   "n_a": 12,
   "n_b": 12,
   "median_a": 19.13073106755353,
   "ci_a": [
    18.745326302296334,
    20.005360570414403
   ],
   "median_b": 36.36478535014914,
   "ci_b": [
    34.13753044405526,
    42.04500706287061
   ],
   "diff": -17.23405428259561,
   "ci_diff": [
    -22.91427599531708,
    -14.833423635496317
   ],
   "rel_diff_pct": -47.392151821198446,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/sweep.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 32 slots: transition m=8 vs no prefetch",
   "n_a": 12,
   "n_b": 12,
   "median_a": 41.32499296378025,
   "ci_a": [
    35.45132953777109,
    46.87438633199691
   ],
   "median_b": 50.55650193124784,
   "ci_b": [
    46.6179394611299,
    56.852686944161775
   ],
   "diff": -9.23150896746759,
   "ci_diff": [
    -17.486052289083403,
    -2.598222895951629
   ],
   "rel_diff_pct": -18.259785813546966,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/sweep.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 32 slots: transition2 m=8 vs no prefetch",
   "n_a": 12,
   "n_b": 12,
   "median_a": 42.02425447124503,
   "ci_a": [
    38.459575220686936,
    46.75470157960534
   ],
   "median_b": 50.55650193124784,
   "ci_b": [
    46.6179394611299,
    56.852686944161775
   ],
   "diff": -8.532247460002807,
   "ci_diff": [
    -15.240124741890014,
    -2.7179076483432
   ],
   "rel_diff_pct": -16.876657074901807,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/sweep.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 32 slots: transition3 m=8 vs no prefetch",
   "n_a": 12,
   "n_b": 12,
   "median_a": 41.02227504595441,
   "ci_a": [
    37.59374509091457,
    46.645720976690825
   ],
   "median_b": 50.55650193124784,
   "ci_b": [
    46.6179394611299,
    56.852686944161775
   ],
   "diff": -9.534226885293428,
   "ci_diff": [
    -16.24725536552569,
    -2.9552358624899995
   ],
   "rel_diff_pct": -18.858557299433205,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/sweep.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 32 slots: transition2 m=16 vs no prefetch",
   "n_a": 12,
   "n_b": 12,
   "median_a": 40.9858914503088,
   "ci_a": [
    35.4116167818508,
    46.03429262237789
   ],
   "median_b": 50.55650193124784,
   "ci_b": [
    46.6179394611299,
    56.852686944161775
   ],
   "diff": -9.570610480939038,
   "ci_diff": [
    -17.710608210914906,
    -3.556491759775213
   ],
   "rel_diff_pct": -18.930523504086935,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/sweep.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 32 slots: prev_rank m=16 vs no prefetch",
   "n_a": 12,
   "n_b": 12,
   "median_a": 28.37905550744074,
   "ci_a": [
    26.997499853533533,
    32.45499196158143
   ],
   "median_b": 50.55650193124784,
   "ci_b": [
    46.6179394611299,
    56.852686944161775
   ],
   "diff": -22.177446423807098,
   "ci_diff": [
    -28.57031143448528,
    -16.946638006695025
   ],
   "rel_diff_pct": -43.86665528000013,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/sweep.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 16 slots (general): probe m=8 vs no prefetch",
   "n_a": 4,
   "n_b": 4,
   "median_a": 31.841276936782968,
   "ci_a": [
    31.596870661689568,
    32.84528995900421
   ],
   "median_b": 32.99759250772945,
   "ci_b": [
    32.636451601945915,
    33.273626374779
   ],
   "diff": -1.1563155709464823,
   "ci_diff": [
    -1.5500378457589257,
    -0.1523025487252383
   ],
   "rel_diff_pct": -3.5042422282038412,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/attribution.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 16 slots (general): probe m=16 vs no prefetch",
   "n_a": 4,
   "n_b": 4,
   "median_a": 25.317370635052725,
   "ci_a": [
    24.76517223527836,
    26.203663395431743
   ],
   "median_b": 32.99759250772945,
   "ci_b": [
    32.636451601945915,
    33.273626374779
   ],
   "diff": -7.680221872676725,
   "ci_diff": [
    -8.398426644951499,
    -6.752416400732631
   ],
   "rel_diff_pct": -23.275097632888485,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/attribution.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 16 slots (general): probe2 m=8 vs no prefetch",
   "n_a": 4,
   "n_b": 4,
   "median_a": 31.34039585473687,
   "ci_a": [
    30.97991967772089,
    32.02471239861806
   ],
   "median_b": 32.99759250772945,
   "ci_b": [
    32.636451601945915,
    33.273626374779
   ],
   "diff": -1.6571966529925817,
   "ci_diff": [
    -2.1243054121081393,
    -0.8894574043585379
   ],
   "rel_diff_pct": -5.022174428647772,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/attribution.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 16 slots (general): probe3 m=8 vs no prefetch",
   "n_a": 4,
   "n_b": 4,
   "median_a": 30.747024815645197,
   "ci_a": [
    29.277490438874107,
    32.56493375331503
   ],
   "median_b": 32.99759250772945,
   "ci_b": [
    32.636451601945915,
    33.273626374779
   ],
   "diff": -2.2505676920842532,
   "ci_diff": [
    -3.7201020688553434,
    -0.4326587544144189
   ],
   "rel_diff_pct": -6.820399674785588,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/attribution.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 16 slots (general): transition2 m=8 vs no prefetch",
   "n_a": 4,
   "n_b": 4,
   "median_a": 27.12807201818323,
   "ci_a": [
    26.024585035633084,
    28.245109293687104
   ],
   "median_b": 32.99759250772945,
   "ci_b": [
    32.636451601945915,
    33.273626374779
   ],
   "diff": -5.869520489546222,
   "ci_diff": [
    -7.147820761442922,
    -4.541891551805659
   ],
   "rel_diff_pct": -17.787723417007857,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/attribution.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 16 slots (general): transition2t50 m=8 vs no prefetch",
   "n_a": 4,
   "n_b": 4,
   "median_a": 27.924338730957444,
   "ci_a": [
    26.832091171562322,
    29.41238423146309
   ],
   "median_b": 32.99759250772945,
   "ci_b": [
    32.636451601945915,
    33.273626374779
   ],
   "diff": -5.073253776772006,
   "ci_diff": [
    -6.165501336167129,
    -3.5852082762663606
   ],
   "rel_diff_pct": -15.37461793791057,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/attribution.jsonl"
  },
  {
   "comparison": "laptop cold NVMe, 16 slots (general): transition2t70 m=8 vs no prefetch",
   "n_a": 4,
   "n_b": 4,
   "median_a": 27.822276415206307,
   "ci_a": [
    26.652845588205146,
    29.388223861291042
   ],
   "median_b": 32.99759250772945,
   "ci_b": [
    32.636451601945915,
    33.273626374779
   ],
   "diff": -5.175316092523143,
   "ci_diff": [
    -6.3447469195243045,
    -3.6093686464384085
   ],
   "rel_diff_pct": -15.683920247547945,
   "unit": "tok/s",
   "source": "gateC/cold_nvme_laptop/attribution.jsonl"
  },
  {
   "comparison": "laptop H4: fused up+gate vs two kernels, 64 slots resident",
   "n_a": 12,
   "n_b": 12,
   "median_a": 475.590753313536,
   "ci_a": [
    473.54619144727985,
    476.23632189642444
   ],
   "median_b": 476.71854951152625,
   "ci_b": [
    475.7441354752446,
    478.1937096893877
   ],
   "diff": -1.1277961979902216,
   "ci_diff": [
    -3.608229933155485,
    0.07714985221900861
   ],
   "rel_diff_pct": -0.23657485095678102,
   "unit": "tok/s",
   "source": "gateD/fused_upgate_laptop.jsonl"
  },
  {
   "comparison": "laptop H4: fused up+gate vs two kernels, 16 slots cold",
   "n_a": 4,
   "n_b": 4,
   "median_a": 35.94742494598419,
   "ci_a": [
    35.731574910456565,
    36.7122725819878
   ],
   "median_b": 34.98584079217508,
   "ci_b": [
    34.12741716795578,
    35.49684771325741
   ],
   "diff": 0.9615841538091061,
   "ci_diff": [
    0.42788351347727627,
    1.964214448416044
   ],
   "rel_diff_pct": 2.7484951970174563,
   "unit": "tok/s",
   "source": "gateD/fused_upgate_laptop.jsonl"
  },
  {
   "comparison": "M3U96a+M3U96b OLMoE 32 slots: odd experts remote vs local",
   "n_a": 16,
   "n_b": 16,
   "median_a": 15.099750302534616,
   "ci_a": [
    14.832823372534836,
    15.478581724182211
   ],
   "median_b": 81.29106238077739,
   "ci_b": [
    70.45778798608657,
    86.91246514746022
   ],
   "diff": -66.19131207824277,
   "ci_diff": [
    -71.76838200410603,
    -55.195792166972154
   ],
   "rel_diff_pct": -81.42507938719571,
   "unit": "tok/s",
   "source": "gateE/two_node_m3u96a_m3u96b.jsonl"
  }
 ]
}