{
  "execution_mode": "offline analysis of fresh Metal eager and compiled expression runs",
  "cases": [
    {
      "rows": 1,
      "width": 896,
      "max_abs_error": 0.0,
      "eager_median_ms": 0.45025,
      "compiled_median_ms": 0.290083
    },
    {
      "rows": 32,
      "width": 896,
      "max_abs_error": 0.0,
      "eager_median_ms": 0.222666,
      "compiled_median_ms": 0.192312
    },
    {
      "rows": 32,
      "width": 1003,
      "max_abs_error": 0.0,
      "eager_median_ms": 0.228792,
      "compiled_median_ms": 0.196583
    }
  ],
  "limits": "Shape-specific tensor timings; no Triton/CUDA run, device-memory counter or whole-model speedup."
}