{
 "schema": 1,
 "suite": "parity",
 "date": "2026-10-23",
 "machine": "rtx-5090",
 "device": "cuda",
 "ollaya": "0.16.0 at 1095167 with local changes",
 "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.rs (laya family), run by hand"
 },
 "results": [
  {
   "model": "coba:1b",
   "questions": 2383,
   "decisions_same": 2383,
   "max_prob_diff": 0.0000031,
   "p99_prob_diff": 0.000001,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "encoding_mismatches": 0,
   "goldens": "convert/out/goldens-all/laya-coba-1b.jsonl (477 cases, laya 0.3.7 fp32 on CPU)",
   "export_sha256": "a031573f8daf261bc083670165f61f156687dac73d67598347fa744392bc8d8e"
  },
  {
   "model": "coba:1b-fp16",
   "questions": 2383,
   "decisions_same": 2379,
   "max_prob_diff": 0.012,
   "p99_prob_diff": 0.0059,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp16",
   "encoding_mismatches": 0,
   "goldens": "convert/out/goldens-all/laya-coba-1b.jsonl (477 cases, laya 0.3.7 fp32 on CPU)",
   "export_sha256": "d7d2283df2ad6044314d41b6f52f463bbe6e8f0d011d59cbc6dc4112f5fc21df"
  },
  {
   "model": "coba:400m",
   "questions": 2383,
   "decisions_same": 2383,
   "max_prob_diff": 0.000011,
   "p99_prob_diff": 0.0000022,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "encoding_mismatches": 0,
   "goldens": "convert/out/goldens-all/laya-coba-400m.jsonl (477 cases, laya 0.3.7 fp32 on CPU)",
   "export_sha256": "49c15752c4a34f9c73af0cc1fedf9e9e46993c5796ba58b465f1a86c759ed6c7"
  },
  {
   "model": "coba:400m-fp16",
   "questions": 2383,
   "decisions_same": 2379,
   "max_prob_diff": 0.015,
   "p99_prob_diff": 0.0074,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp16",
   "encoding_mismatches": 0,
   "goldens": "convert/out/goldens-all/laya-coba-400m.jsonl (477 cases, laya 0.3.7 fp32 on CPU)",
   "export_sha256": "27779f73da3cfe1525ae696544a8a53212fbca197b4c1949f55d59e09fae3fab"
  },
  {
   "model": "coba:400m-ternary",
   "questions": 2383,
   "decisions_same": 2383,
   "max_prob_diff": 0.0000016,
   "p99_prob_diff": 8.6e-7,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "encoding_mismatches": 0,
   "goldens": "convert/out/goldens-all/laya-coba-400m-ternary.jsonl (477 cases, laya 0.3.7 fp32 on CPU)",
   "export_sha256": "5127480db99f5c68a4e14af3166699112e53e78c1808035d56330d4e502698a5"
  },
  {
   "model": "coba:400m-ternary-fp16",
   "questions": 2383,
   "decisions_same": 2374,
   "max_prob_diff": 0.0069,
   "p99_prob_diff": 0.0045,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp16",
   "encoding_mismatches": 0,
   "goldens": "convert/out/goldens-all/laya-coba-400m-ternary.jsonl (477 cases, laya 0.3.7 fp32 on CPU)",
   "export_sha256": "40b093e6ac84d201857a5d581a0d7a30299f8466fb1bb9ff8b11f16664126400"
  },
  {
   "model": "coba:150m-ternary",
   "questions": 2383,
   "decisions_same": 2383,
   "max_prob_diff": 0.0000017,
   "p99_prob_diff": 0.000001,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "encoding_mismatches": 0,
   "goldens": "convert/out/goldens-all/laya-coba-150m-ternary.jsonl (477 cases, laya 0.3.7 fp32 on CPU)",
   "export_sha256": "1dd25d156537e8e8813644e0192ce663d94ea365f99dea3c6fff657d072241b3"
  },
  {
   "model": "coba:150m-ternary-fp16",
   "questions": 2383,
   "decisions_same": 2377,
   "max_prob_diff": 0.015,
   "p99_prob_diff": 0.0046,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp16",
   "encoding_mismatches": 0,
   "goldens": "convert/out/goldens-all/laya-coba-150m-ternary.jsonl (477 cases, laya 0.3.7 fp32 on CPU)",
   "export_sha256": "343e7d503a5e2f5241cf336614ced3d00860bf4126564a6dc2020536c3d891a8"
  },
  {
   "model": "coba:68m-ternary",
   "questions": 2383,
   "decisions_same": 2383,
   "max_prob_diff": 0.0000018,
   "p99_prob_diff": 9.6e-7,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "encoding_mismatches": 0,
   "goldens": "convert/out/goldens-all/laya-coba-68m-ternary.jsonl (477 cases, laya 0.3.7 fp32 on CPU)",
   "export_sha256": "fb4e9fcb360e4c39f105d51fe221a539688504f1e87decca7b877bccddaa5422"
  },
  {
   "model": "coba:68m-ternary-fp16",
   "questions": 2383,
   "decisions_same": 2376,
   "max_prob_diff": 0.011,
   "p99_prob_diff": 0.0049,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp16",
   "encoding_mismatches": 0,
   "goldens": "convert/out/goldens-all/laya-coba-68m-ternary.jsonl (477 cases, laya 0.3.7 fp32 on CPU)",
   "export_sha256": "59bbf665d7e05e34f971534d40058b21f59e751a4d9a826e96c91edac0da3ed2"
  }
 ]
}