{
 "chart": "latency",
 "from_chart_data_entry": "latency.idle",
 "source_path": "runs/macjev/latency/m1max_untrained_base_2026-09-23_review_idle/latency.json",
 "field": "results['llamacpp-f16'|'laya-multilingual-mps-fp32'][cell].p50_ms",
 "protocol_label": "M1 Max 64 GB, warm end-to-end p50 ms, idle run 2026-09-23 (load avg 3-8); untrained identical-architecture Qwen3.5-0.8B export (latency does not depend on weights); v3 = llama.cpp GGUF F16, one call per state; comparison engine = round-1 MacLaya-4K (our Laya-multilingual fine-tune, FP32/MPS, 4,096-token budget), one decide() per question; prefix reuse off",
 "plotted": [
  {
   "cell": "1024x5",
   "v3_p50_ms": 394,
   "maclaya4k_p50_ms": 543,
   "speedup_label": "1.4x"
  },
  {
   "cell": "1024x10",
   "v3_p50_ms": 525,
   "maclaya4k_p50_ms": 1019,
   "speedup_label": "1.9x"
  },
  {
   "cell": "4096x5",
   "v3_p50_ms": 1212,
   "maclaya4k_p50_ms": 3204,
   "speedup_label": "2.6x"
  },
  {
   "cell": "4096x10",
   "v3_p50_ms": 1381,
   "maclaya4k_p50_ms": 6364,
   "speedup_label": "4.6x"
  },
  {
   "cell": "8192x1",
   "v3_p50_ms": 2295,
   "maclaya4k_p50_ms": null,
   "speedup_label": "v3 only"
  },
  {
   "cell": "8192x5",
   "v3_p50_ms": 2424,
   "maclaya4k_p50_ms": null,
   "speedup_label": "v3 only"
  },
  {
   "cell": "8192x10",
   "v3_p50_ms": 2613,
   "maclaya4k_p50_ms": null,
   "speedup_label": "v3 only"
  }
 ]
}