{
  "experiment": "utilization-detectors",
  "queue_depth_medians": {
    "5": {
      "runs": 3,
      "throughput_rps": 45.75,
      "steady_throughput_rps": 47.6,
      "p95_ttft_ms": 1863.287667,
      "p99_ttft_ms": 2142.009417,
      "p95_tpot_ms_per_token": 21.420569
    },
    "8": {
      "runs": 3,
      "throughput_rps": 45.316667,
      "steady_throughput_rps": 46.816667,
      "p95_ttft_ms": 1607.462917,
      "p99_ttft_ms": 1864.029666,
      "p95_tpot_ms_per_token": 20.929388
    }
  },
  "kv_pressure_medians": {
    "flow_control_off": {
      "runs": 3,
      "throughput_rps": 1.111111,
      "steady_throughput_rps": 1.111111,
      "p95_ttft_ms": 16938.384917,
      "p99_ttft_ms": 17613.521917,
      "p95_tpot_ms_per_token": 185.987657
    },
    "threshold_0.75": {
      "runs": 3,
      "throughput_rps": 1.216667,
      "steady_throughput_rps": 1.0,
      "p95_ttft_ms": 28540.358875,
      "p99_ttft_ms": 28724.332875,
      "p95_tpot_ms_per_token": 182.96858
    },
    "threshold_0.8": {
      "runs": 3,
      "throughput_rps": 1.155556,
      "steady_throughput_rps": 0.866667,
      "p95_ttft_ms": 21559.096292,
      "p99_ttft_ms": 22409.380292,
      "p95_tpot_ms_per_token": 183.828462
    }
  },
  "decision": "Advance queue depths 2 and 5 into production traffic comparisons. Retain KV threshold 0.8 as the primary memory-pressure calibration point.",
  "claim_boundary": "The KV sweep intentionally created extreme memory pressure and does not establish a latency SLO."
}