{
  "format": "bit-jev-single-question-case-v1",
  "scope": "one fixed development request; timings exclude model loading",
  "workload": { "questions": 1, "input_tokens": 703, "options": 77 },
  "hardware": { "cpu": "Intel Xeon Gold 6459C", "gpu": "NVIDIA GeForce RTX 5090" },
  "measurements": [
    { "path": "native_cpu_i2_s_8_threads", "mean_inference_ms": 3127.88, "repetitions": 3, "peak_process_rss_mib": 1622.74 },
    { "path": "native_cpu_i2_s_16_threads", "mean_inference_ms": 1972.17, "repetitions": 3, "peak_process_rss_mib": 1624.52 },
    { "path": "experimental_gpu_fp16_mixed_precision", "mean_inference_ms": 86.56, "repetitions": 5, "peak_gpu_allocation_mib": 4935.53 }
  ],
  "limitations": [
    "CPU uses the I2_S native runner while GPU uses an experimental FP16 mixed-precision path.",
    "This is one request with few repetitions, not a general throughput or accuracy claim.",
    "CPU process RSS and GPU allocation are different memory measurements.",
    "Generated tokens per second is not applicable: the model scores explicit options and emits structured decisions."
  ]
}