| { | |
| "model": "wildroot-student-v1", | |
| "commit": "14b09f6d3857633b2bc75ab3c371421083da3b8c", | |
| "sourceClean": true, | |
| "recordedRows": 929, | |
| "onnxruntime": "1.30.0", | |
| "tokenizers": "0.23.2", | |
| "files": { | |
| "action.onnx": { | |
| "bytes": 133053705, | |
| "sha256": "382ea0e1627a6f4d7d43ffaed038ddd09d125f60647beb6a5b43d4a78feef6e1" | |
| }, | |
| "config.json": { | |
| "bytes": 814, | |
| "sha256": "89a0470bb57c2cf85a5d1fe1726d16067db938a99b7a4dab8a5c01219dfaaef6" | |
| }, | |
| "engine.mjs": { | |
| "bytes": 5127, | |
| "sha256": "30b0e114f28e48c2bc4612d20da27b4d126e5acab184105d737a29a634c4ebfe" | |
| }, | |
| "recorded.jsonl": { | |
| "bytes": 3257936, | |
| "sha256": "655fcc527680a8a3e4ddb0d0c95d198811e3d2179dd694c11bf5355b6ec156c5" | |
| }, | |
| "replay.mjs": { | |
| "bytes": 4584, | |
| "sha256": "6307a5462d53a8769cdcde025b92b87c784b44499c6443679236fc7f749eb327" | |
| }, | |
| "student.onnx": { | |
| "bytes": 133071786, | |
| "sha256": "90bf82dfdf2696698a1ff33c7519b92d5e22510fbca846ddcced9532e059a0ec" | |
| }, | |
| "tokenizer.json": { | |
| "bytes": 711396, | |
| "sha256": "d241a60d5e8f04cc1b2b3e9ef7a4921b27bf526d9f6050ab90f9267a1f9e5c66" | |
| }, | |
| "vocab.txt": { | |
| "bytes": 231508, | |
| "sha256": "07eced375cec144d27c900241f3e339478dec958f92fddbc551f295c992038a3" | |
| } | |
| }, | |
| "results": { | |
| "test": "sealed urgent pool, 2026-11-02, scored rows 546, category labels 489", | |
| "against": "typesafe/jev-1.14-20261014, best setup on each main measure", | |
| "category": { | |
| "model": 90.2, | |
| "jev": 92.4, | |
| "difference": [ | |
| -2.2, | |
| -5.1, | |
| 0.6 | |
| ], | |
| "verdict": "unresolved" | |
| }, | |
| "needsReplyPrecision": { | |
| "model": 87.8, | |
| "jev": 84.8, | |
| "difference": [ | |
| 3.0, | |
| -3.7, | |
| 9.6 | |
| ], | |
| "verdict": "near" | |
| }, | |
| "needsReplyRecall": { | |
| "model": 90.8, | |
| "jev": 98.3, | |
| "difference": [ | |
| -7.6, | |
| -13.1, | |
| -2.5 | |
| ] | |
| }, | |
| "actionAuc": { | |
| "model": 0.965, | |
| "jev": 0.993, | |
| "difference": [ | |
| -0.028, | |
| -0.057, | |
| -0.004 | |
| ], | |
| "verdict": "unresolved" | |
| }, | |
| "actionThreshold": { | |
| "value": 0.3706, | |
| "chosenOn": "development mail", | |
| "note": "sealed 23 of 27 caught, 43 of 495 false, read after the test" | |
| } | |
| } | |
| } | |