{
 "schema": 1,
 "suite": "parity",
 "date": "2026-10-19",
 "machine": "rtx-4090",
 "device": "cuda",
 "ollaya": "0.13.0 at ff727a1 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": "convert/ollaya_convert/parity_run.py with crates/ollaya-runner/examples/parity_<family>.rs"
 },
 "results": [
  {
   "model": "clef:flash",
   "questions": 571,
   "decisions_same": 571,
   "max_logit_diff": 0.000043272972106933594,
   "max_prob_diff": 0.000006263810163442685,
   "p99_prob_diff": 0.000003629033356267186,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "3a6355337dc24a92616766cb47a8ee8fc37f6b82928860c388ecbb30e61c1279"
  },
  {
   "model": "clef:flash-fp16",
   "questions": 571,
   "decisions_same": 570,
   "max_logit_diff": 0.02852034568786621,
   "max_prob_diff": 0.004528411125939824,
   "p99_prob_diff": 0.002213794732440433,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp16",
   "export_sha256": "9d04d7145af6c6f691d6d573d5fef905c25285e6122872e2477b81781543bf22"
  },
  {
   "model": "decider:0.8b",
   "questions": 479,
   "decisions_same": 479,
   "max_logit_diff": 0.0000286102294921875,
   "max_prob_diff": 0.00000609635883996873,
   "p99_prob_diff": 0.000004483068768323939,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "62c1bf794ebdeea86bfdcb85a56a042d455a4933fa57f9dd8aab8965344506ff"
  },
  {
   "model": "decider:2b",
   "questions": 479,
   "decisions_same": 479,
   "max_logit_diff": 0.00005245208740234375,
   "max_prob_diff": 0.000006334218465742225,
   "p99_prob_diff": 0.000004157086843648017,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "89010f69c69f4f3dce665de111b20acbf5610602749e347c669bc972b5d0903a"
  },
  {
   "model": "decider:2b-vision",
   "questions": 404,
   "decisions_same": 404,
   "max_logit_diff": 0.0001649856567365049,
   "max_prob_diff": 0.00003592319090328644,
   "p99_prob_diff": 0.00000910556113320471,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "abd716ae64a2ce510c4f2eb5cf3e8b5ac216f3e262922828bf41028105b9abda"
  },
  {
   "model": "decider:4b",
   "questions": 479,
   "decisions_same": 479,
   "max_logit_diff": 0.000026702880859375,
   "max_prob_diff": 0.000006054783519171547,
   "p99_prob_diff": 0.0000037731048364575415,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "eb512e9a61947d531036117a0eae6b3f3bd25df8f4accc91e5268ce3a89d131b"
  },
  {
   "model": "decision:eos",
   "questions": 466,
   "decisions_same": 466,
   "max_logit_diff": 0.000016212463379128295,
   "max_prob_diff": 0.0000029331238043939223,
   "p99_prob_diff": 0.0000021446107726919017,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "bbfc1d45be4970a85d8ffcb640bf85fb8fb9ea9ca7c82ab152eaccc7c70c818a"
  },
  {
   "model": "jeeves:9b",
   "questions": 430,
   "decisions_same": 430,
   "max_logit_diff": 0.00019288063049316406,
   "max_prob_diff": 0.000014241056536301588,
   "p99_prob_diff": 0.00000799651204536067,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "817cca81576ef41816d467edae8d364e4b13ccfabcff5e09644dd88f4c0d4bf4"
  },
  {
   "model": "kev:0.8b",
   "questions": 480,
   "decisions_same": 480,
   "max_logit_diff": 0.000033855438232421875,
   "max_prob_diff": 0.0000019354208431465736,
   "p99_prob_diff": 0.0000015389650566133284,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "04503cafca4b79727db2671a23386ef44e3bfc13d069d13f0b4b2aa0eba36f84"
  },
  {
   "model": "kev:0.8b-fp16",
   "questions": 480,
   "decisions_same": 478,
   "max_logit_diff": 0.1004633903503418,
   "max_prob_diff": 0.008844151990767157,
   "p99_prob_diff": 0.005343693636535862,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp16",
   "export_sha256": "caab362a02b1d57bac7733ab6fd73f3c9fb0238827d390f56f8dd5c683e9d28b"
  },
  {
   "model": "kev:4b",
   "questions": 480,
   "decisions_same": 480,
   "max_logit_diff": 0.00022220611572265625,
   "max_prob_diff": 0.000030094051240658892,
   "p99_prob_diff": 0.0000028466319112741,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "075f5d2f37ea0258f4c0bbfc61c6c8044b9587f181080f24bc64cee941914a9d"
  },
  {
   "model": "kev:4b-fp16",
   "questions": 480,
   "decisions_same": 478,
   "max_logit_diff": 0.0649397075176239,
   "max_prob_diff": 0.003867276949636622,
   "p99_prob_diff": 0.0021664238047864326,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp16",
   "export_sha256": "a3ff3a0894f289e03f66d4933582cfc1d0be76d4aec42cc12f2cf99345e7d006"
  },
  {
   "model": "kev:9b",
   "questions": 480,
   "decisions_same": 480,
   "max_logit_diff": 0.00007534027099609375,
   "max_prob_diff": 0.00000421917961912488,
   "p99_prob_diff": 0.0000033390887213546883,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "c8099a71d136b0ce22429470ae0e40c31cc396af29e805f4dde561885cf78474"
  },
  {
   "model": "kev:9b-fp16",
   "questions": 480,
   "decisions_same": 480,
   "max_logit_diff": 0.027038097381591797,
   "max_prob_diff": 0.002407114291201995,
   "p99_prob_diff": 0.001830038523635169,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp16",
   "export_sha256": "66eb8698d54666bd02ca68f91246944f43ab60e6fdb8ef14f3986a4eae988c50"
  },
  {
   "model": "nimble:9b",
   "questions": 492,
   "decisions_same": 492,
   "max_logit_diff": 0.000110626220703125,
   "max_prob_diff": 0.000006248095183569724,
   "p99_prob_diff": 0.000005698721595726042,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "1d0bb04b575910c4be9245de873db1c1c2d219d32c1cf0e9677f3006b35dc00d"
  },
  {
   "model": "nimble:9b-fp16",
   "questions": 492,
   "decisions_same": 492,
   "max_logit_diff": 0.04482460021972656,
   "max_prob_diff": 0.004883934582212979,
   "p99_prob_diff": 0.0025211005548978482,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp16",
   "export_sha256": "dff3c6170939e176e2d6ae4d499b1671f2de38b4700827a3f7225d458f6d2a46"
  },
  {
   "model": "nimble:9b-v3",
   "questions": 498,
   "decisions_same": 498,
   "max_logit_diff": 0.0001373291015625,
   "max_prob_diff": 0.00001084139223461733,
   "p99_prob_diff": 0.000006682368842714226,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "49ccb42b1f530aff52bba41ffe46b5bdc8dc5e6f6c98f69dd4ff071184e64f44"
  },
  {
   "model": "pplx-decider:flash",
   "questions": 696,
   "decisions_same": 696,
   "max_logit_diff": 0.0000820159912109375,
   "max_prob_diff": 0.000006815665555159889,
   "p99_prob_diff": 0.000003593970269721858,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "4cf33d50cea2a2236bc0c52ac4f25da089a2cc63bc0cb124165f7b844c23b84a"
  },
  {
   "model": "pplx-decider:flash-fp16",
   "questions": 696,
   "decisions_same": 695,
   "max_logit_diff": 0.066619873046875,
   "max_prob_diff": 0.0061545107737028815,
   "p99_prob_diff": 0.003955836186643191,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp16",
   "export_sha256": "9d13c8ffbf3130057a98e662fd8b1646a5d5f159997b4a89e7242e709d0390f8"
  },
  {
   "model": "strands-decider:2b",
   "questions": 650,
   "decisions_same": 650,
   "max_logit_diff": 0.000010609626770019531,
   "max_prob_diff": 0.000002780074783048647,
   "p99_prob_diff": 0.0000015954034170428955,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp32",
   "export_sha256": "ae5cf8009363f96b97fd8ed616c1e57ddcdd01276ee750a1f3f3617de5a2fd4e"
  },
  {
   "model": "strands-decider:2b-fp16",
   "questions": 650,
   "decisions_same": 649,
   "max_logit_diff": 0.022779405117035023,
   "max_prob_diff": 0.00561743843610929,
   "p99_prob_diff": 0.004419259873180814,
   "pass": true,
   "device": "cuda",
   "kind": "runtime",
   "gate": "fp16",
   "export_sha256": "5d89027c226741b96c92ab2b4dbde5b5cd3789a0a013a911ae2cc13699b58b1d"
  }
 ]
}
