[
  {
    "run_id": "R-001",
    "model": "vlm-caption-s",
    "precision_raw": "W8A8",
    "backend_raw": "NPU-A v2.1",
    "accuracy_metric": "top1",
    "accuracy_value": 0.712,
    "baseline_accuracy": 0.719,
    "latency_p95_raw": "27.4ms",
    "peak_mem_raw": "486MB",
    "toolchain_version": "npir-1.8.2",
    "calib_dataset": "calib_indoor_512",
    "log_excerpt": "[compile] 214/214 ops mapped to NPU",
    "measured_at": "2026-07-14T10:20:00"
  },
  {
    "run_id": "R-002",
    "model": "vlm-caption-s",
    "precision_raw": "W4A16",
    "backend_raw": "NPU-A v2.1",
    "accuracy_metric": "top1",
    "accuracy_value": 0.688,
    "baseline_accuracy": 0.719,
    "latency_p95_raw": "19.8ms",
    "peak_mem_raw": "312MB",
    "toolchain_version": "npir-1.8.2",
    "calib_dataset": "calib_indoor_512",
    "log_excerpt": "[warn] op Softmax_23 fell back to CPU",
    "measured_at": "2026-07-14T13:05:00"
  },
  {
    "run_id": "R-003",
    "model": "vlm-caption-s",
    "precision_raw": "w8a8-per-channel",
    "backend_raw": "npu_a v2.1",
    "accuracy_metric": "top1",
    "accuracy_value": 0.71,
    "baseline_accuracy": 0.719,
    "latency_p95_raw": "0.031s",
    "peak_mem_raw": "0.52GB",
    "toolchain_version": "npir-1.8.2",
    "calib_dataset": "calib_mixed_1k",
    "log_excerpt": "[warn] op LayerNorm_7 fell back to CPU\n[warn] op Softmax_31 fell back to CPU",
    "measured_at": "2026-07-15T09:41:00"
  },
  {
    "run_id": "R-004",
    "model": "vlm-caption-s",
    "precision_raw": "W8A8",
    "backend_raw": "NPU-A v2.1",
    "accuracy_metric": null,
    "accuracy_value": null,
    "baseline_accuracy": 0.719,
    "latency_p95_raw": "24.9ms",
    "peak_mem_raw": "470MB",
    "toolchain_version": "npir-1.9.0-rc1",
    "calib_dataset": null,
    "log_excerpt": "[compile] 214/214 ops mapped to NPU",
    "measured_at": "2026-07-15T16:12:00"
  },
  {
    "run_id": "R-005",
    "model": "vlm-caption-s",
    "precision_raw": "FP16",
    "backend_raw": "TensorRT 8.6",
    "accuracy_metric": "top1",
    "accuracy_value": 0.718,
    "baseline_accuracy": 0.719,
    "latency_p95_raw": "12.2ms",
    "peak_mem_raw": "1.1GB",
    "toolchain_version": "trt-8.6.1",
    "calib_dataset": null,
    "log_excerpt": "[info] engine built with fp16 tactics",
    "measured_at": "2026-07-16T11:00:00"
  },
  {
    "run_id": "R-006",
    "model": "vlm-caption-s",
    "precision_raw": "W8A8",
    "backend_raw": "NPU-A v2.0",
    "accuracy_metric": "top1",
    "accuracy_value": 0.705,
    "baseline_accuracy": null,
    "latency_p95_raw": "28.8ms",
    "peak_mem_raw": "498MB",
    "toolchain_version": "npir-1.8.1",
    "calib_dataset": "calib_indoor_512",
    "log_excerpt": "[compile] 213/214 ops mapped to NPU",
    "measured_at": "2026-07-16T15:30:00"
  },
  {
    "run_id": "R-006",
    "model": "vlm-caption-s",
    "precision_raw": "W8A8",
    "backend_raw": "NPU-A v2.1",
    "accuracy_metric": "top1",
    "accuracy_value": 0.709,
    "baseline_accuracy": 0.719,
    "latency_p95_raw": "28.1ms",
    "peak_mem_raw": "492MB",
    "toolchain_version": "npir-1.8.2",
    "calib_dataset": "calib_indoor_512",
    "log_excerpt": "[compile] 214/214 ops mapped to NPU",
    "measured_at": "2026-07-17T09:02:00"
  },
  {
    "run_id": "R-008",
    "model": "vlm-caption-s",
    "precision_raw": "W8A8",
    "backend_raw": "NPU-A v2.1",
    "accuracy_metric": "top1",
    "accuracy_value": 0.714,
    "baseline_accuracy": 0.719,
    "latency_p95_raw": "29.6ms",
    "peak_mem_raw": "505MB",
    "toolchain_version": "npir-1.8.2",
    "calib_dataset": "calib_indoor_512",
    "log_excerpt": "[info] 1 op fell back to CPU: Softmax_23",
    "measured_at": "2026-07-17T14:20:00"
  },
  {
    "run_id": "R-009",
    "model": "vlm-caption-s",
    "precision_raw": "W4A16",
    "backend_raw": "NPU-A v2.1",
    "accuracy_metric": "top1",
    "accuracy_value": 0.701,
    "baseline_accuracy": 0.719,
    "latency_p95_raw": null,
    "peak_mem_raw": null,
    "toolchain_version": "npir-1.8.2",
    "calib_dataset": "calib_mixed_1k",
    "log_excerpt": "[error] runtime OOM during benchmark, run aborted",
    "measured_at": "2026-07-18T10:44:00"
  },
  {
    "run_id": "R-010",
    "model": "vlm-caption-s",
    "precision_raw": "INT8",
    "backend_raw": "TFLite",
    "accuracy_metric": "top1",
    "accuracy_value": 0.697,
    "baseline_accuracy": 0.719,
    "latency_p95_raw": "63ms",
    "peak_mem_raw": "402MB",
    "toolchain_version": "tflite-2.16",
    "calib_dataset": "calib_indoor_512",
    "log_excerpt": "[warn] 6 ops delegated to CPU kernel",
    "measured_at": "2026-07-18T13:15:00"
  },
  {
    "run_id": "R-011",
    "model": "vlm-caption-s",
    "precision_raw": "W8A8",
    "backend_raw": "NPU-A v2.1",
    "accuracy_metric": "top1_lowlight",
    "accuracy_value": 0.664,
    "baseline_accuracy": 0.677,
    "latency_p95_raw": "27.9ms",
    "peak_mem_raw": "489MB",
    "toolchain_version": "npir-1.8.2",
    "calib_dataset": "calib_indoor_512",
    "log_excerpt": "[compile] 214/214 ops mapped to NPU",
    "measured_at": "2026-07-19T09:10:00"
  },
  {
    "run_id": "R-012",
    "model": "vlm-caption-m",
    "precision_raw": "W8A8",
    "backend_raw": "NPU_A2",
    "accuracy_metric": "top1",
    "accuracy_value": 0.741,
    "baseline_accuracy": 0.748,
    "latency_p95_raw": "41.5ms",
    "peak_mem_raw": "744MB",
    "toolchain_version": "npir-1.8.2",
    "calib_dataset": "calib_indoor_512",
    "log_excerpt": "[compile] 268/268 ops mapped to NPU",
    "measured_at": null
  },
  {
    "run_id": "R-013",
    "model": "vlm-caption-s",
    "precision_raw": "W8A8",
    "backend_raw": "NPU-A v2.1",
    "accuracy_metric": "top1",
    "accuracy_value": 0.716,
    "baseline_accuracy": 0.719,
    "latency_p95_raw": "26.8ms",
    "peak_mem_raw": "478MB",
    "toolchain_version": "npir-1.8.2",
    "calib_dataset": "calib_mixed_1k",
    "log_excerpt": "[compile] 214/214 ops mapped to NPU",
    "measured_at": "2026-07-19T15:52:00"
  }
]