{
 "schema": 1,
 "suite": "parity",
 "date": "2026-10-24",
 "machine": "rtx-5090",
 "os": "linux (WSL2)",
 "ollaya": "0.17.0, branch models/cob-560 (CUDA at 015d2aa, CPU at 0ae2506), parity_llama",
 "llama_cpp": "b11146",
 "protocol": {
  "reference": "stock llama-server of the same llama.cpp build on the same device (ADR 0003), prompts from the family's Python reference of the author's prompt",
  "gate": "identical prompt ids and decisions, option logits within 1e-3",
  "latency": "not recorded: the parity runs shared the machine with other jobs"
 },
 "results": [
  {
   "model": "jeb:9b",
   "device": "cuda",
   "gpu": "NVIDIA GeForce RTX 5090",
   "questions": 494,
   "rejected_requests": 19,
   "decisions_same": 494,
   "max_logit_diff": 0.00000756,
   "p99_logit_diff": 0.000007469,
   "max_prob_diff": 0.000002755,
   "p99_prob_diff": 0.000002461,
   "seconds": 31.2,
   "pass": true,
   "revision": "v2.1"
  },
  {
   "model": "winnow:e2b",
   "device": "cuda",
   "gpu": "NVIDIA GeForce RTX 5090",
   "questions": 505,
   "rejected_requests": 15,
   "decisions_same": 505,
   "max_logit_diff": 0.00001082,
   "p99_logit_diff": 0.000009041,
   "max_prob_diff": 0.00000281,
   "p99_prob_diff": 0.00000233,
   "seconds": 17.2,
   "pass": true
  },
  {
   "model": "jeb:27b",
   "device": "cuda",
   "gpu": "NVIDIA GeForce RTX 5090",
   "questions": 494,
   "rejected_requests": 19,
   "decisions_same": 494,
   "max_logit_diff": 0.000007622,
   "p99_logit_diff": 0.000007249,
   "max_prob_diff": 0.000002753,
   "p99_prob_diff": 0.000002142,
   "seconds": 44.8,
   "pass": true,
   "revision": "v2.1"
  },
  {
   "model": "jeb:9b",
   "device": "cpu",
   "cpu": "AMD Ryzen Threadripper 3960X 24-Core Processor",
   "questions": 21,
   "rejected_requests": 8,
   "decisions_same": 21,
   "max_logit_diff": 0.000005591,
   "p99_logit_diff": 0.000005591,
   "max_prob_diff": 0.000002794,
   "p99_prob_diff": 0.000002794,
   "seconds": 139,
   "pass": true,
   "revision": "v2.1",
   "cases": "the smoke cases (crates/ollaya-runner/examples/smoke/cases.txt): K2 on the CPU for 9B and over",
   "reference_note": "CPU goldens: the CUDA goldens' prompts replayed through stock llama-server on this machine's CPU (llm_common/replay.py --device cpu)"
  },
  {
   "model": "winnow:e2b",
   "device": "cpu",
   "cpu": "AMD Ryzen Threadripper 3960X 24-Core Processor",
   "questions": 505,
   "rejected_requests": 15,
   "decisions_same": 505,
   "max_logit_diff": 0.000009781,
   "p99_logit_diff": 0.000008272,
   "max_prob_diff": 0.000003822,
   "p99_prob_diff": 0.000002578,
   "seconds": 624.2,
   "pass": true,
   "reference_note": "CPU goldens: the CUDA goldens' prompts replayed through stock llama-server on this machine's CPU (llm_common/replay.py --device cpu)"
  },
  {
   "model": "jeb:27b",
   "device": "cpu",
   "cpu": "AMD Ryzen Threadripper 3960X 24-Core Processor",
   "questions": 21,
   "rejected_requests": 8,
   "decisions_same": 21,
   "max_logit_diff": 0.000007538,
   "p99_logit_diff": 0.000007538,
   "max_prob_diff": 0.000001279,
   "p99_prob_diff": 0.000001279,
   "seconds": 205.5,
   "pass": true,
   "revision": "v2.1",
   "cases": "the smoke cases (crates/ollaya-runner/examples/smoke/cases.txt): K2 on the CPU for 9B and over",
   "reference_note": "CPU goldens: the CUDA goldens' prompts replayed through stock llama-server on this machine's CPU (llm_common/replay.py --device cpu)"
  }
 ]
}