{
 "schema": 1,
 "suite": "parity",
 "date": "2026-10-24",
 "machine": "rtx-4090",
 "device": "cuda",
 "ollaya": "0.17.0 at b97e96a (decider rows, PR #144) and at b37b867 (d1-omni:600m) and at d4ea37f (strands-decider:e2b, :e4b) and at 59d230f (gliclass-v4:2b, :9b)",
 "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": "convert/ollaya_convert/parity_run.py with crates/ollaya-runner/examples/parity_<family>.rs"
 },
 "results": [
  {
   "model": "decider:2b",
   "questions": 479,
   "decisions_same": 479,
   "max_logit_diff": 0.00003814697265447364,
   "max_prob_diff": 0.00007461986231571682,
   "p99_prob_diff": 0.000003797235976488178,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "9aaf0a5a45a7419d447a82cf9532d3f187163d30fb53f65567dc67cb34ef7e7a"
  },
  {
   "model": "decider:4b",
   "questions": 479,
   "decisions_same": 479,
   "max_logit_diff": 0.00004482269286931739,
   "max_prob_diff": 0.00020279301252523751,
   "p99_prob_diff": 0.000005291610535640423,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "b7ffca6f95269b6a8ecaa600a2092ade7a20d55d342d89915199cb49962c29c0"
  },
  {
   "model": "d1-omni:600m",
   "questions": 456,
   "decisions_same": 456,
   "max_logit_diff": 0.00008606910705566406,
   "max_prob_diff": 0.000011498346962124373,
   "p99_prob_diff": 0.000007244427759622862,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "fd603c73eba7dcc33648e155efc8fe4dc14a7f8aeca0737efc2be3ea29a03e7f"
  },
  {
   "model": "strands-decider:e2b",
   "questions": 650,
   "decisions_same": 650,
   "max_logit_diff": 0.00001817941665649414,
   "max_prob_diff": 0.00001051499051396565,
   "p99_prob_diff": 0.0000022447222797206123,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "790ecb95e54cf68acbab0d0820032c63bd837c214cc6e51a98c2c19da5761ce3"
  },
  {
   "model": "strands-decider:e4b",
   "questions": 650,
   "decisions_same": 650,
   "max_logit_diff": 0.000016927719116210938,
   "max_prob_diff": 0.0000035941373554559952,
   "p99_prob_diff": 0.0000023435583632602786,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "935ac79395c9f9014e4ac201c31898ccb71d873757acc524345e017b4d6d2f5a"
  },
  {
   "model": "gliclass-v4:2b",
   "questions": 483,
   "decisions_same": 483,
   "max_logit_diff": 0.0005695819854736328,
   "max_prob_diff": 0.00014664453631935137,
   "p99_prob_diff": 0.000009718881360698539,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "75258aa8dc83b422eb43ff03f3c6245e95a8fb255fc5ed89c9e497185501d00e"
  },
  {
   "model": "gliclass-v4:9b",
   "questions": 483,
   "decisions_same": 483,
   "max_logit_diff": 0.000025391578674316406,
   "max_prob_diff": 0.00000502894648246599,
   "p99_prob_diff": 0.0000037748545872462458,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "16ca230cea5d297fa0dc74e7327600652b7bd3aff893ff2d9180415546da1774"
  }
 ]
}