{
 "schema": 1,
 "suite": "latency-triage",
 "date": "2026-10-06T09:10:06Z",
 "ollaya": "0.10.0",
 "machine": "rtx-4090",
 "machine_detail": {
  "id": "rtx-4090",
  "cpu": "13th Gen Intel(R) Core(TM) i9-13900K",
  "threads": 24,
  "gpus": [
   {
    "name": "NVIDIA GeForce RTX 4090",
    "driver": "616.92",
    "memory": "24564 MiB"
   }
  ],
  "os": "Linux 6.18.33.2-microsoft-standard-WSL2"
 },
 "device": "cuda",
 "protocol": {
  "preset": "triage",
  "questions": 5,
  "states": 30,
  "warm": 5,
  "n": 20,
  "client": "one request at a time over HTTP to /api/decide, timed by the client; each request a different short message, in turn"
 },
 "note": "decima, measured on a development build (0.10.0 with the decima-late-interaction-v1 engine) before its release; the same protocol and machine as the 2026-10-01 RTX 4090 sweep. decima:small was timed under decima:latest, its alias in that build's registry; decima:base and decima:agent 30 minutes later, with the same build.",
 "results": [
  {
   "model": "decima:small",
   "questions": "triage",
   "load_ms": 840.2,
   "first_request_ms": 850.1,
   "answered_by": "decima:latest",
   "p50_ms": 7.3,
   "p90_ms": 8.5,
   "min_ms": 6.7,
   "eval_p50_ms": 6.8,
   "input_tokens": 567,
   "device": "cuda:0",
   "precision": "F32"
  },
  {
   "model": "decima:base",
   "questions": "triage",
   "load_ms": 1433.5,
   "first_request_ms": 1450.1,
   "answered_by": "decima:base",
   "p50_ms": 15.1,
   "p90_ms": 15.4,
   "min_ms": 14.1,
   "eval_p50_ms": 14.6,
   "input_tokens": 458,
   "device": "cuda:0",
   "precision": "F32"
  },
  {
   "model": "decima:agent",
   "questions": "triage",
   "load_ms": 1269.1,
   "first_request_ms": 1284.2,
   "answered_by": "decima:agent",
   "p50_ms": 14.8,
   "p90_ms": 15.9,
   "min_ms": 14,
   "eval_p50_ms": 14.2,
   "input_tokens": 458,
   "device": "cuda:0",
   "precision": "F32"
  }
 ]
}