{
 "schema": 1,
 "date": "2026-10-24",
 "edition": "2026.10",
 "machine": "rtx-5090",
 "device": "cuda",
 "suite": "typed-decisions",
 "ollaya": "0.17.0",
 "protocol": {
  "benchmark": "typed-decisions test split: 400 JSON states from four workflows, five typed questions each (2,000 questions: 600 choice, 600 noul, 800 score)",
  "source": "https://huggingface.co/datasets/LocalLLaMA/typed-decisions",
  "runner": "convert/ollaya_convert/bench_typed.py",
  "client": "one request (one state, its five questions) at a time over HTTP to /v1/systemone, timed by the client",
  "accuracy": "argmax of the returned probabilities against the gold label, over all 2,000 questions; a rejected request counts as wrong on its five questions; 95 % bootstrap interval over states",
  "ece": "ten-bin expected calibration error of the top probability, pooled over the answered questions"
 },
 "results": [
  {
   "model": "sura:2b",
   "digest": "33489bfce93fe2075f800bf55be89346f3083826eb14495a29a43b00b1cbfe42",
   "answered_by": [
    "sura:2b"
   ],
   "rows": 400,
   "questions": 2000,
   "errors": 0,
   "accuracy": 0.57,
   "accuracy_ci": [
    0.541,
    0.5975
   ],
   "by_type": {
    "choice": 0.5867,
    "noul": 0.6017,
    "score": 0.5337
   },
   "by_workflow": {
    "agent_trace_observability": 0.562,
    "customer_service": 0.686,
    "invoice_processing": 0.438,
    "security_incidents": 0.594
   },
   "ece": 0.1025,
   "brier": 0.5467,
   "soft_ce": 1.1764,
   "median_ms": 77.5,
   "p95_ms": 97.4,
   "trained_on_split": false
  }
 ]
}