{
 "schema": 1,
 "suite": "parity",
 "date": "2026-10-24",
 "machine": "mac-m4-pro",
 "device": "metal",
 "ollaya": "0.17.0 at b97e96a",
 "protocol": {
  "reference": "goldens from each family's goldens.py: the authors' code in PyTorch fp32 (TF32 off) on the checkpoint the manifest names",
  "gate": "fp32 tags: identical encodings and decisions, option logits within 1e-3, wire answers as the authors' code; -fp16 tags (docs/decisions/0008-fp16-decoder-tags.md): identical encodings, at least 99.5 % of decisions the same, differences reported",
  "runner": "crates/ollaya-runner/examples/parity_decider.rs with --features mlx and the device metal, run by hand (parity_run records cuda and cpu only)"
 },
 "results": [
  {
   "model": "decider:2b",
   "questions": 479,
   "decisions_same": 479,
   "max_logit_diff": 0.00004100799560546875,
   "max_prob_diff": 0.00004591483227245341,
   "p99_prob_diff": 0.000002766943943965927,
   "pass": true,
   "device": "metal",
   "kind": "runtime",
   "gate": "fp32",
   "encoding_mismatches": 0,
   "goldens": "convert/out/goldens-decider-2b.jsonl (114 cases, decider-ai 1.10.0 fp32 on CUDA, made on the rtx-4090 the same day)",
   "engine": "mlx"
  }
 ]
}