{
  "experiment": "concurrency-detectors",
  "request_count_medians": {
    "128": {
      "runs": 3,
      "throughput_rps": 48.211111,
      "steady_throughput_rps": 49.633333,
      "p95_ttft_ms": 2075.835084,
      "p99_ttft_ms": 2547.533542,
      "p95_tpot_ms_per_token": 21.17596
    },
    "160": {
      "runs": 3,
      "throughput_rps": 49.183333,
      "steady_throughput_rps": 51.133333,
      "p95_ttft_ms": 2438.810792,
      "p99_ttft_ms": 3090.735792,
      "p95_tpot_ms_per_token": 20.155998
    }
  },
  "mixed_size_medians": {
    "request_count_128": {
      "runs": 3,
      "total_steady_throughput_rps": 18.808696,
      "p95_ttft_ms_by_request_size": {
        "short request": 8418.679583,
        "medium request": 5658.528833,
        "long request": 4163.141875
      }
    },
    "input_token_1.2x": {
      "runs": 3,
      "total_steady_throughput_rps": 21.4,
      "p95_ttft_ms_by_request_size": {
        "short request": 6620.738333,
        "medium request": 5213.166417,
        "long request": 6352.499417
      }
    },
    "input_plus_output_estimate": {
      "runs": 1,
      "total_steady_throughput_rps": 6.782609,
      "p95_ttft_ms_by_request_size": {
        "short request": 41909.050375,
        "medium request": 27951.616875,
        "long request": 13961.544084
      }
    }
  },
  "decision": "Use request-count 128 as the default. Use exact input-token admission when request-size variability requires size-aware control. Do not advance the fixed-output estimator.",
  "claim_boundary": "Closed-loop detector calibration selects settings. Production behavior is tested separately with open-loop traffic."
}