{
  "date": "2026-09-27",
  "hardware": "Apple M2 Ultra",
  "memory_gib": 192,
  "os": "macOS 27.0 (26A428)",
  "ollaya": "0.7.2",
  "model": "laya:multilingual",
  "manifest_digest": "d4bbd188073b7b47a2e58d18eb2132354fa8a14feb1030a01e91494f7cacc8c7",
  "weights_revision": "aa8c91ca088ec597df95a0d1c76b3063cb2ae5e8",
  "weights_sha256": "9d628fd971b700382ac6f65920a86f149777b2e748e0c955fb3b19695aa8f204",
  "model_manifest_total_bytes": 684162752,
  "precision": "fp32",
  "order": "CPU first, server restart, Metal second",
  "question_count": 1,
  "options": 4,
  "warmup_requests": 0,
  "repeats_per_case_per_device": 1,
  "cpu_engine": "onnx",
  "metal_engine": "mlx",
  "cpu_model_load_ms_log": 1217,
  "metal_model_load_ms_log": 2283,
  "same_choices": 20,
  "timing": "Python perf_counter around local HTTP request, JSON parsing and response validation; initial request includes model loading; remaining median uses requests 2..20. Sequential; no parallel requests.",
  "isolation": "Downloaded to temporary directory; no system installer or persistent service; macOS sandbox refused all filesystem writes outside this trial directory, including Ollaya home PID. Only synthetic inquiry text was submitted to loopback endpoint.",
  "limitations": [
    "Small hand-authored smoke-test set, not a representative benchmark",
    "No Mica or Jev live comparison performed",
    "No threshold or prompt tuning after observing outcomes",
    "GPU verified by API device=metal and server log engine=mlx, not by a hardware profiler"
  ]
}