{
  "configuration": {
    "url": "http://127.0.0.1:8000",
    "model_dir": "/root/autodl-tmp/jev9b-deploy/models/JEV-9B",
    "workload": "shared",
    "rows": 128,
    "cardinality": 8,
    "correlated_columns": 12,
    "field_tokens": [
      128,
      512
    ],
    "repeat_ratios": [
      0.8,
      0.97,
      0.983
    ],
    "methods": [
      "original",
      "solo"
    ],
    "repeats": 2,
    "concurrency": 4,
    "max_context_tokens": 16384,
    "seed": 42,
    "dry_run": false,
    "out": "validation/shared-live.json"
  },
  "measurement": "Fresh cache_salt per trial, identical backend and concurrency; alternating method order. Timed scans include normalization, planning, serialization, network and initial cache fills. Kernel/planner warmup is separate.",
  "cases": [
    {
      "workload": {
        "family": "shared_fields",
        "target_reusable_value_token_fraction": 0.8,
        "constant_field_token_budget": 132,
        "construction": "Unique ID and customer message first; eight reusable template fields, including two long global constants. Synthetic dummy padding; complete rows sent in every method.",
        "rows": 128,
        "columns": 10,
        "synthetic": true,
        "dataset_sha256": "e755eb3d58cfecafed7b730699a7168833f0a3ad116dd66c047c78600fdd9e75",
        "column_ndv": {
          "record_id": 128,
          "customer_request": 128,
          "company_policy": 1,
          "service_context": 1,
          "profile_0": 4,
          "profile_1": 8,
          "profile_2": 16,
          "profile_3": 32,
          "profile_4": 64,
          "profile_5": 64
        },
        "all_column_ndv_equal": false,
        "field_token_ranges": {
          "record_id": [
            23,
            34
          ],
          "customer_request": [
            230,
            241
          ],
          "company_policy": [
            132,
            132
          ],
          "service_context": [
            132,
            132
          ],
          "profile_0": [
            132,
            132
          ],
          "profile_1": [
            132,
            132
          ],
          "profile_2": [
            132,
            132
          ],
          "profile_3": [
            132,
            132
          ],
          "profile_4": [
            132,
            132
          ],
          "profile_5": [
            132,
            132
          ]
        },
        "reusable_value_token_fraction": 0.8,
        "reuse_definition": "Fraction of cell-value tokens in columns whose whole-field NDV is less than row count; this is not the prefix-cache hit rate.",
        "prefix_distinct_counts": {
          "original": [
            128,
            128,
            128,
            128,
            128,
            128,
            128,
            128,
            128,
            128
          ],
          "solo": [
            1,
            1,
            4,
            8,
            16,
            32,
            64,
            64,
            128,
            128
          ]
        },
        "max_prompt_tokens_across_methods": 1401
      },
      "summary": [
        {
          "method": "original",
          "rows": 128,
          "repeats": 2,
          "wall_seconds": 21.35865343315527,
          "planning_seconds": 8.695665746927261e-05,
          "prompt_tokens": 179310.0,
          "cached_tokens": 0.0,
          "cached_fraction": 0.0,
          "accuracy": 1.0,
          "rows_per_second": 5.992887164005584,
          "agreement_with_original": 1.0,
          "speedup_vs_original": 1.0,
          "solo_speedup_vs_this_method": 1.4925912569880038
        },
        {
          "method": "solo",
          "rows": 128,
          "repeats": 2,
          "wall_seconds": 14.309780613519251,
          "planning_seconds": 0.0007363124750554562,
          "prompt_tokens": 179310.0,
          "cached_tokens": 92928.0,
          "cached_fraction": 0.5182533043332775,
          "accuracy": 1.0,
          "rows_per_second": 8.944930985110368,
          "agreement_with_original": 1.0,
          "speedup_vs_original": 1.4925912569880038,
          "solo_speedup_vs_this_method": 1.0
        }
      ],
      "runs": [
        {
          "method": "original",
          "rows": 128,
          "wall_seconds": 20.586928449571133,
          "rows_per_second": 6.217537517242719,
          "preparation_seconds": 0.0011286241933703423,
          "planning_seconds": 6.333459168672562e-05,
          "inference_seconds": 20.585635418072343,
          "prompt_tokens": 179310,
          "cached_tokens": 0,
          "cached_fraction": 0.0,
          "latency_p50_seconds": 0.6348446640186012,
          "latency_p95_seconds": 0.7411252505611624,
          "column_order": [
            "record_id",
            "customer_request",
            "company_policy",
            "service_context",
            "profile_0",
            "profile_1",
            "profile_2",
            "profile_3",
            "profile_4",
            "profile_5"
          ],
          "repeat": 0,
          "cache_salt": "7abd205aa11c4e0b9675cc75e32904b3",
          "accuracy": 1.0
        },
        {
          "method": "solo",
          "rows": 128,
          "wall_seconds": 14.215153069235384,
          "rows_per_second": 9.004475673006944,
          "preparation_seconds": 0.0009594149887561798,
          "planning_seconds": 0.0007625175639986992,
          "inference_seconds": 14.21338982321322,
          "prompt_tokens": 179310,
          "cached_tokens": 92928,
          "cached_fraction": 0.5182533043332775,
          "latency_p50_seconds": 0.4347610012628138,
          "latency_p95_seconds": 0.5035277194809169,
          "column_order": [
            "company_policy",
            "service_context",
            "profile_0",
            "profile_1",
            "profile_2",
            "profile_3",
            "profile_4",
            "profile_5",
            "record_id",
            "customer_request"
          ],
          "repeat": 0,
          "cache_salt": "483d6c8141954c42983913886dd01eec",
          "accuracy": 1.0
        },
        {
          "method": "solo",
          "rows": 128,
          "wall_seconds": 14.404408157803118,
          "rows_per_second": 8.88616863655451,
          "preparation_seconds": 0.0006765881553292274,
          "planning_seconds": 0.0007101073861122131,
          "inference_seconds": 14.402978892438114,
          "prompt_tokens": 179310,
          "cached_tokens": 92928,
          "cached_fraction": 0.5182533043332775,
          "latency_p50_seconds": 0.4366109208203852,
          "latency_p95_seconds": 0.6120597354136408,
          "column_order": [
            "company_policy",
            "service_context",
            "profile_0",
            "profile_1",
            "profile_2",
            "profile_3",
            "profile_4",
            "profile_5",
            "record_id",
            "customer_request"
          ],
          "repeat": 1,
          "cache_salt": "f5645701c025434f94e19e0f910f0117",
          "accuracy": 1.0
        },
        {
          "method": "original",
          "rows": 128,
          "wall_seconds": 22.130378416739404,
          "rows_per_second": 5.783904711867958,
          "preparation_seconds": 0.001874956302344799,
          "planning_seconds": 0.00011057872325181961,
          "inference_seconds": 22.12833890505135,
          "prompt_tokens": 179310,
          "cached_tokens": 0,
          "cached_fraction": 0.0,
          "latency_p50_seconds": 0.6860883585177362,
          "latency_p95_seconds": 0.7158638071734458,
          "column_order": [
            "record_id",
            "customer_request",
            "company_policy",
            "service_context",
            "profile_0",
            "profile_1",
            "profile_2",
            "profile_3",
            "profile_4",
            "profile_5"
          ],
          "repeat": 1,
          "cache_salt": "4022601921404e459a7cceda9d9519c4",
          "accuracy": 1.0
        }
      ]
    },
    {
      "workload": {
        "family": "shared_fields",
        "target_reusable_value_token_fraction": 0.97,
        "constant_field_token_budget": 3872,
        "construction": "Unique ID and customer message first; eight reusable template fields, including two long global constants. Synthetic dummy padding; complete rows sent in every method.",
        "rows": 128,
        "columns": 10,
        "synthetic": true,
        "dataset_sha256": "e5795b06d87acfaf43b1c2499c35e432223dc463004d7a261a059ddbf9a9bc06",
        "column_ndv": {
          "record_id": 128,
          "customer_request": 128,
          "company_policy": 1,
          "service_context": 1,
          "profile_0": 4,
          "profile_1": 8,
          "profile_2": 16,
          "profile_3": 32,
          "profile_4": 64,
          "profile_5": 64
        },
        "all_column_ndv_equal": false,
        "field_token_ranges": {
          "record_id": [
            23,
            34
          ],
          "customer_request": [
            230,
            241
          ],
          "company_policy": [
            3872,
            3872
          ],
          "service_context": [
            3872,
            3872
          ],
          "profile_0": [
            132,
            132
          ],
          "profile_1": [
            132,
            132
          ],
          "profile_2": [
            132,
            132
          ],
          "profile_3": [
            132,
            132
          ],
          "profile_4": [
            132,
            132
          ],
          "profile_5": [
            132,
            132
          ]
        },
        "reusable_value_token_fraction": 0.97,
        "reuse_definition": "Fraction of cell-value tokens in columns whose whole-field NDV is less than row count; this is not the prefix-cache hit rate.",
        "prefix_distinct_counts": {
          "original": [
            128,
            128,
            128,
            128,
            128,
            128,
            128,
            128,
            128,
            128
          ],
          "solo": [
            1,
            1,
            4,
            8,
            16,
            32,
            64,
            64,
            128,
            128
          ]
        },
        "max_prompt_tokens_across_methods": 8881
      },
      "summary": [
        {
          "method": "original",
          "rows": 128,
          "repeats": 2,
          "wall_seconds": 151.50681221624836,
          "planning_seconds": 6.308825686573982e-05,
          "prompt_tokens": 1136750.0,
          "cached_tokens": 0.0,
          "cached_fraction": 0.0,
          "accuracy": 1.0,
          "rows_per_second": 0.8448464998214293,
          "agreement_with_original": 1.0,
          "speedup_vs_original": 1.0,
          "solo_speedup_vs_this_method": 7.903937912967029
        },
        {
          "method": "solo",
          "rows": 128,
          "repeats": 2,
          "wall_seconds": 19.16852256236598,
          "planning_seconds": 0.000895205419510603,
          "prompt_tokens": 1136750.0,
          "cached_tokens": 1016928.0,
          "cached_fraction": 0.8945924785572905,
          "accuracy": 1.0,
          "rows_per_second": 6.677614280576088,
          "agreement_with_original": 1.0,
          "speedup_vs_original": 7.903937912967029,
          "solo_speedup_vs_this_method": 1.0
        }
      ],
      "runs": [
        {
          "method": "original",
          "rows": 128,
          "wall_seconds": 151.2974149538204,
          "rows_per_second": 0.8460157765356974,
          "preparation_seconds": 0.0017267093062400818,
          "planning_seconds": 5.656387656927109e-05,
          "inference_seconds": 151.29554118961096,
          "prompt_tokens": 1136750,
          "cached_tokens": 0,
          "cached_fraction": 0.0,
          "latency_p50_seconds": 4.676185711752623,
          "latency_p95_seconds": 4.889460154809058,
          "column_order": [
            "record_id",
            "customer_request",
            "company_policy",
            "service_context",
            "profile_0",
            "profile_1",
            "profile_2",
            "profile_3",
            "profile_4",
            "profile_5"
          ],
          "repeat": 0,
          "cache_salt": "a324dd37ef8745f986293d783b81dc56",
          "accuracy": 1.0
        },
        {
          "method": "solo",
          "rows": 128,
          "wall_seconds": 19.092849308624864,
          "rows_per_second": 6.70408056602522,
          "preparation_seconds": 0.0011700615286827087,
          "planning_seconds": 0.000872073695063591,
          "inference_seconds": 19.090682674199343,
          "prompt_tokens": 1136750,
          "cached_tokens": 1016928,
          "cached_fraction": 0.8945924785572905,
          "latency_p50_seconds": 0.46819390961900353,
          "latency_p95_seconds": 1.3970851397607476,
          "column_order": [
            "company_policy",
            "service_context",
            "profile_0",
            "profile_1",
            "profile_2",
            "profile_3",
            "profile_4",
            "profile_5",
            "record_id",
            "customer_request"
          ],
          "repeat": 0,
          "cache_salt": "916ea562ea484e5eb94a3c08caed70ea",
          "accuracy": 1.0
        },
        {
          "method": "solo",
          "rows": 128,
          "wall_seconds": 19.244195816107094,
          "rows_per_second": 6.651356139957066,
          "preparation_seconds": 0.0012679435312747955,
          "planning_seconds": 0.0009183371439576149,
          "inference_seconds": 19.241853489540517,
          "prompt_tokens": 1136750,
          "cached_tokens": 1016928,
          "cached_fraction": 0.8945924785572905,
          "latency_p50_seconds": 0.5079086534678936,
          "latency_p95_seconds": 1.4104190536774694,
          "column_order": [
            "company_policy",
            "service_context",
            "profile_0",
            "profile_1",
            "profile_2",
            "profile_3",
            "profile_4",
            "profile_5",
            "record_id",
            "customer_request"
          ],
          "repeat": 1,
          "cache_salt": "73269a0fdb3249f6b4a9df6fdf1aa1ab",
          "accuracy": 1.0
        },
        {
          "method": "original",
          "rows": 128,
          "wall_seconds": 151.71620947867632,
          "rows_per_second": 0.8436804507562548,
          "preparation_seconds": 0.0019382024183869362,
          "planning_seconds": 6.961263716220856e-05,
          "inference_seconds": 151.71412666607648,
          "prompt_tokens": 1136750,
          "cached_tokens": 0,
          "cached_fraction": 0.0,
          "latency_p50_seconds": 4.6801456972025335,
          "latency_p95_seconds": 4.891390697192401,
          "column_order": [
            "record_id",
            "customer_request",
            "company_policy",
            "service_context",
            "profile_0",
            "profile_1",
            "profile_2",
            "profile_3",
            "profile_4",
            "profile_5"
          ],
          "repeat": 1,
          "cache_salt": "23bd4bbda29641979583a7d69dd5831f",
          "accuracy": 1.0
        }
      ]
    },
    {
      "workload": {
        "family": "shared_fields",
        "target_reusable_value_token_fraction": 0.983,
        "constant_field_token_budget": 7237,
        "construction": "Unique ID and customer message first; eight reusable template fields, including two long global constants. Synthetic dummy padding; complete rows sent in every method.",
        "rows": 128,
        "columns": 10,
        "synthetic": true,
        "dataset_sha256": "f25bc7b22765b95bafb27ae3704ce40e31b8bc4e5cd6ee25d03c511bb43c445e",
        "column_ndv": {
          "record_id": 128,
          "customer_request": 128,
          "company_policy": 1,
          "service_context": 1,
          "profile_0": 4,
          "profile_1": 8,
          "profile_2": 16,
          "profile_3": 32,
          "profile_4": 64,
          "profile_5": 64
        },
        "all_column_ndv_equal": false,
        "field_token_ranges": {
          "record_id": [
            23,
            34
          ],
          "customer_request": [
            230,
            241
          ],
          "company_policy": [
            7237,
            7237
          ],
          "service_context": [
            7237,
            7237
          ],
          "profile_0": [
            132,
            132
          ],
          "profile_1": [
            132,
            132
          ],
          "profile_2": [
            132,
            132
          ],
          "profile_3": [
            132,
            132
          ],
          "profile_4": [
            132,
            132
          ],
          "profile_5": [
            132,
            132
          ]
        },
        "reusable_value_token_fraction": 0.9830006439150032,
        "reuse_definition": "Fraction of cell-value tokens in columns whose whole-field NDV is less than row count; this is not the prefix-cache hit rate.",
        "prefix_distinct_counts": {
          "original": [
            128,
            128,
            128,
            128,
            128,
            128,
            128,
            128,
            128,
            128
          ],
          "solo": [
            1,
            1,
            4,
            8,
            16,
            32,
            64,
            64,
            128,
            128
          ]
        },
        "max_prompt_tokens_across_methods": 15611
      },
      "summary": [
        {
          "method": "original",
          "rows": 128,
          "repeats": 2,
          "wall_seconds": 274.882449363824,
          "planning_seconds": 4.954775795340538e-05,
          "prompt_tokens": 1998190.0,
          "cached_tokens": 0.0,
          "cached_fraction": 0.0,
          "accuracy": 1.0,
          "rows_per_second": 0.4656535922763991,
          "agreement_with_original": 1.0,
          "speedup_vs_original": 1.0,
          "solo_speedup_vs_this_method": 11.707030911619759
        },
        {
          "method": "solo",
          "rows": 128,
          "repeats": 2,
          "wall_seconds": 23.48011647351086,
          "planning_seconds": 0.0013667745515704155,
          "prompt_tokens": 1998190.0,
          "cached_tokens": 1866480.0,
          "cached_fraction": 0.9340853472392515,
          "accuracy": 1.0,
          "rows_per_second": 5.451420998886587,
          "agreement_with_original": 1.0,
          "speedup_vs_original": 11.707030911619759,
          "solo_speedup_vs_this_method": 1.0
        }
      ],
      "runs": [
        {
          "method": "original",
          "rows": 128,
          "wall_seconds": 275.1908325338736,
          "rows_per_second": 0.46513177354570595,
          "preparation_seconds": 0.0011975448578596115,
          "planning_seconds": 4.332512617111206e-05,
          "inference_seconds": 275.18948846962303,
          "prompt_tokens": 1998190,
          "cached_tokens": 0,
          "cached_fraction": 0.0,
          "latency_p50_seconds": 8.663504235446453,
          "latency_p95_seconds": 8.688695457670837,
          "column_order": [
            "record_id",
            "customer_request",
            "company_policy",
            "service_context",
            "profile_0",
            "profile_1",
            "profile_2",
            "profile_3",
            "profile_4",
            "profile_5"
          ],
          "repeat": 0,
          "cache_salt": "c02a4b2c9c6e4e439e89d486f09109f0",
          "accuracy": 1.0
        },
        {
          "method": "solo",
          "rows": 128,
          "wall_seconds": 23.675790563225746,
          "rows_per_second": 5.406366459366096,
          "preparation_seconds": 0.001345306634902954,
          "planning_seconds": 0.0011941613629460335,
          "inference_seconds": 23.673184842802584,
          "prompt_tokens": 1998190,
          "cached_tokens": 1866480,
          "cached_fraction": 0.9340853472392515,
          "latency_p50_seconds": 0.5516993612982333,
          "latency_p95_seconds": 2.2314216263592224,
          "column_order": [
            "company_policy",
            "service_context",
            "profile_0",
            "profile_1",
            "profile_2",
            "profile_3",
            "profile_4",
            "profile_5",
            "record_id",
            "customer_request"
          ],
          "repeat": 0,
          "cache_salt": "8b87489e9d50403397613a2026b25af1",
          "accuracy": 1.0
        },
        {
          "method": "solo",
          "rows": 128,
          "wall_seconds": 23.284442383795977,
          "rows_per_second": 5.497232782738971,
          "preparation_seconds": 0.0018259650096297264,
          "planning_seconds": 0.0015393877401947975,
          "inference_seconds": 23.280936068855226,
          "prompt_tokens": 1998190,
          "cached_tokens": 1866480,
          "cached_fraction": 0.9340853472392515,
          "latency_p50_seconds": 0.549300764221698,
          "latency_p95_seconds": 2.358857129281387,
          "column_order": [
            "company_policy",
            "service_context",
            "profile_0",
            "profile_1",
            "profile_2",
            "profile_3",
            "profile_4",
            "profile_5",
            "record_id",
            "customer_request"
          ],
          "repeat": 1,
          "cache_salt": "b54b618b1fe64688a24c3ff887f02af9",
          "accuracy": 1.0
        },
        {
          "method": "original",
          "rows": 128,
          "wall_seconds": 274.57406619377434,
          "rows_per_second": 0.4661765831506714,
          "preparation_seconds": 0.0009439717978239059,
          "planning_seconds": 5.57703897356987e-05,
          "inference_seconds": 274.57299120817333,
          "prompt_tokens": 1998190,
          "cached_tokens": 0,
          "cached_fraction": 0.0,
          "latency_p50_seconds": 8.650064995046705,
          "latency_p95_seconds": 8.661546867201105,
          "column_order": [
            "record_id",
            "customer_request",
            "company_policy",
            "service_context",
            "profile_0",
            "profile_1",
            "profile_2",
            "profile_3",
            "profile_4",
            "profile_5"
          ],
          "repeat": 1,
          "cache_salt": "163ef6f3118943d8a813c247fc34c3f1",
          "accuracy": 1.0
        }
      ]
    }
  ]
}