{
  "experiment": "batch isolation",
  "selected_configuration_results": {
    "realtime": {
      "runs": 3,
      "median_p95_ttft_ms": 441.851854,
      "range_p95_ttft_ms": [
        370.687008,
        669.149399
      ],
      "median_p99_ttft_ms": 567.050695,
      "median_p95_tpot_ms_per_token": 23.070568
    },
    "standard": {
      "runs": 3,
      "median_p95_ttft_ms": 515.385866,
      "range_p95_ttft_ms": [
        435.703516,
        1017.155886
      ],
      "median_p99_ttft_ms": 641.850948,
      "median_p95_tpot_ms_per_token": 23.372663
    },
    "batch": {
      "runs": 3,
      "median_p95_ttft_ms": 13077.339411,
      "range_p95_ttft_ms": [
        11424.613237,
        15241.892099
      ],
      "median_p99_ttft_ms": 13491.666794,
      "median_p95_tpot_ms_per_token": 119.545
    }
  },
  "matched_detector_comparisons": {},
  "calibration_only": {
    "batch queue depth 2": {
      "realtime": {
        "runs": 1,
        "median_p95_ttft_ms": 4519.713402,
        "range_p95_ttft_ms": [
          4519.713402,
          4519.713402
        ],
        "median_p99_ttft_ms": 4916.324377,
        "median_p95_tpot_ms_per_token": 55.007748
      },
      "standard": {
        "runs": 1,
        "median_p95_ttft_ms": 4366.770983,
        "range_p95_ttft_ms": [
          4366.770983,
          4366.770983
        ],
        "median_p99_ttft_ms": 4886.055946,
        "median_p95_tpot_ms_per_token": 54.259107
      },
      "batch": {
        "runs": 1,
        "median_p95_ttft_ms": 11449.07999,
        "range_p95_ttft_ms": [
          11449.07999,
          11449.07999
        ],
        "median_p99_ttft_ms": 11806.423903,
        "median_p95_tpot_ms_per_token": 109.429313
      }
    }
  },
  "retained_run_inventory": [
    {
      "scenario": "batch isolation",
      "detector": "queue depth 2",
      "evidence_role": "single-run calibration",
      "repeat": 1
    },
    {
      "scenario": "batch isolation",
      "detector": "request count 128, 15% headroom",
      "evidence_role": "selected three-repeat result",
      "repeat": 1
    },
    {
      "scenario": "batch isolation",
      "detector": "request count 128, 15% headroom",
      "evidence_role": "selected three-repeat result",
      "repeat": 2
    },
    {
      "scenario": "batch isolation",
      "detector": "request count 128, 15% headroom",
      "evidence_role": "selected three-repeat result",
      "repeat": 3
    }
  ],
  "saturation_metric_interpretation": "The pool saturation metric is a detector-normalized score. Request-count admission divides in-flight requests by the configured cap. Utilization admission uses the larger of queue depth divided by its threshold and KV-cache utilization divided by its threshold. Read 1.0 as the configured saturation boundary within each detector; raw magnitudes are not compared across detectors.",
  "statistical_scope": "Three selected repeats support descriptive medians and ranges for this scenario.",
  "claim_boundary": "This package contains production-shaped single-GPU evidence for batch isolation. It does not include a matched detector comparison."
}