{
 "schema": 1,
 "suite": "parity",
 "date": "2026-10-01",
 "machine": "rtx-4090",
 "os": "windows",
 "ollaya": "0.8.0 merged with PR #27 (Vulkan), parity_llama from the release workflow's dry run",
 "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": "parity_llama --latency: 20 requests of 5 questions each through the runtime (no HTTP)"
 },
 "results": [
  {
   "model": "winnow:e4b",
   "device": "vulkan",
   "gpu": "NVIDIA GeForce RTX 4090",
   "questions": 505,
   "rejected_requests": 15,
   "decisions_same": 505,
   "max_logit_diff": 0.0000114,
   "max_prob_diff": 0.000002995,
   "seconds": 22.2,
   "p50_ms": 148.6,
   "pass": true
  },
  {
   "model": "winnow:e4b",
   "device": "cuda",
   "gpu": "NVIDIA GeForce RTX 4090",
   "questions": 505,
   "rejected_requests": 15,
   "decisions_same": 505,
   "max_logit_diff": 0.00001288,
   "max_prob_diff": 0.000002984,
   "seconds": 12.7,
   "p50_ms": 102.6,
   "pass": true,
   "reference_note": "the Linux CUDA goldens"
  },
  {
   "model": "jeb:9b",
   "device": "vulkan",
   "gpu": "NVIDIA GeForce RTX 4090",
   "questions": 494,
   "rejected_requests": 19,
   "decisions_same": 494,
   "max_logit_diff": 0.000007634,
   "max_prob_diff": 0.000002232,
   "seconds": 29,
   "p50_ms": 249.6,
   "pass": true
  },
  {
   "model": "jeb:9b",
   "device": "cuda",
   "gpu": "NVIDIA GeForce RTX 4090",
   "questions": 494,
   "rejected_requests": 19,
   "decisions_same": 494,
   "max_logit_diff": 0.000007654,
   "max_prob_diff": 0.000002108,
   "seconds": 17.2,
   "p50_ms": 127.5,
   "pass": true,
   "reference_note": "the Linux CUDA goldens"
  },
  {
   "model": "cygnet:12b",
   "device": "vulkan",
   "gpu": "NVIDIA GeForce RTX 4090",
   "questions": 502,
   "rejected_requests": 17,
   "decisions_same": 502,
   "max_logit_diff": 0.000007699,
   "max_prob_diff": 4.728e-7,
   "seconds": 101.2,
   "p50_ms": 862.3,
   "pass": true
  },
  {
   "model": "cygnet:12b",
   "device": "cuda",
   "gpu": "NVIDIA GeForce RTX 4090",
   "questions": 502,
   "rejected_requests": 17,
   "decisions_same": 502,
   "max_logit_diff": 0.000007689,
   "max_prob_diff": 4.071e-7,
   "seconds": 30.5,
   "p50_ms": 212.5,
   "pass": true,
   "reference_note": "the Linux CUDA goldens"
  },
  {
   "model": "winnow:e4b",
   "device": "cpu",
   "cpu": "Intel Core i9-13900K",
   "questions": 505,
   "rejected_requests": 15,
   "decisions_same": 502,
   "max_logit_diff": 0.3263,
   "max_prob_diff": 0.03677,
   "seconds": 841.9,
   "p50_ms": 4396.2,
   "pass": false,
   "reference_note": "against CPU goldens made on another machine and OS (no metadata kept); a CPU gate needs goldens from the same machine and build"
  }
 ],
 "cross_device": {
  "note": "The same reference prompts on different devices, compared with the RTX 4090's CUDA goldens: how much backends differ. Measured, not gated.",
  "results": [
   {
    "model": "winnow:e4b",
    "device": "cpu (x86-64)",
    "questions": 505,
    "decisions_same": 501,
    "max_logprob_diff": 0.2788
   },
   {
    "model": "winnow:e4b",
    "device": "vulkan (RTX 4090)",
    "questions": 505,
    "decisions_same": 501,
    "max_logprob_diff": 0.3232
   },
   {
    "model": "jeb:9b",
    "device": "vulkan (RTX 4090)",
    "questions": 494,
    "decisions_same": 492,
    "max_logprob_diff": 0.3306
   },
   {
    "model": "cygnet:12b",
    "device": "cpu (i9-13900K, Windows)",
    "questions": 502,
    "decisions_same": 496,
    "max_logprob_diff": 2.761
   },
   {
    "model": "cygnet:12b",
    "device": "vulkan (RTX 4090)",
    "questions": 502,
    "decisions_same": 497,
    "max_logprob_diff": 1.402
   }
  ]
 }
}