{
  "experiment": "engine-configuration",
  "selected": {
    "max_num_sequences": 128,
    "max_num_batched_tokens": 8192
  },
  "max_sequence_medians": {
    "128": {
      "runs": 3,
      "steady_throughput_rps": 50.533333,
      "p95_ttft_ms": 1822.05125,
      "p99_ttft_ms": 1888.182125,
      "p95_tpot_ms_per_token": 19.567304
    },
    "160": {
      "runs": 3,
      "steady_throughput_rps": 50.708333,
      "p95_ttft_ms": 1365.176584,
      "p99_ttft_ms": 1977.651,
      "p95_tpot_ms_per_token": 27.048305
    },
    "192": {
      "runs": 3,
      "steady_throughput_rps": 51.933333,
      "p95_ttft_ms": 1690.712833,
      "p99_ttft_ms": 2613.4375,
      "p95_tpot_ms_per_token": 28.862933
    }
  },
  "batched_token_results": {
    "4096": {
      "runs": 1,
      "steady_throughput_rps": 47.808333,
      "p95_ttft_ms": 1876.799708,
      "p99_ttft_ms": 2013.302542,
      "p95_tpot_ms_per_token": 20.507374
    },
    "8192": {
      "runs": 3,
      "steady_throughput_rps": 50.533333,
      "p95_ttft_ms": 1822.05125,
      "p99_ttft_ms": 1888.182125,
      "p95_tpot_ms_per_token": 19.567304
    },
    "16384": {
      "runs": 1,
      "steady_throughput_rps": 48.741667,
      "p95_ttft_ms": 2161.861417,
      "p99_ttft_ms": 2565.215333,
      "p95_tpot_ms_per_token": 22.754301
    }
  },
  "claim_boundary": "Closed-loop engine calibration selects capacity settings. It is not an SLO proof."
}