Upload 2 files
Browse files
qwen_causal_scored_400.jsonl
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
qwen_causal_scoring_report.json
ADDED
|
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"task": "causal_judge_scoring",
|
| 3 |
+
"branch": "qwen_capping",
|
| 4 |
+
"authority": "docs/DEVIATION_2026-08-18_T27_CAUSAL_JUDGE_GATE.md",
|
| 5 |
+
"judge_model": "deepseek-ai/DeepSeek-V3",
|
| 6 |
+
"provider": "Novita (hosted API)",
|
| 7 |
+
"provider_model_id": "deepseek/deepseek_v3",
|
| 8 |
+
"provider_model_id_note": "Operational Novita routing identifier for the frozen DeepSeek-V3 judge; this is not a change to the scientific judge identity or T27 thresholds.",
|
| 9 |
+
"endpoint": "https://api.novita.ai/v3/openai",
|
| 10 |
+
"transport": {
|
| 11 |
+
"request_pacing_enabled": true,
|
| 12 |
+
"sdk_automatic_retries_disabled": true,
|
| 13 |
+
"observed_provider_limit_requests_per_minute": 10,
|
| 14 |
+
"min_request_interval_seconds": 7.0,
|
| 15 |
+
"rate_limit_backoff_seconds": 15.0,
|
| 16 |
+
"generic_request_error_backoff_seconds": 2.0,
|
| 17 |
+
"scope": "every hosted judge attempt, including canaries and retries",
|
| 18 |
+
"method_boundary": "Operational transport pacing only; no change to judge identity, prompt, schema, temperature, validation gate, sampling rule, or estimand."
|
| 19 |
+
},
|
| 20 |
+
"production": true,
|
| 21 |
+
"drawable_by_t26": true,
|
| 22 |
+
"non_production_reason": null,
|
| 23 |
+
"contract_problems": null,
|
| 24 |
+
"source_git_sha": "5b60b8e7bd32e8422b0886d27c04c52fb61e29f5",
|
| 25 |
+
"source_git_dirty": false,
|
| 26 |
+
"revision_pinnable": false,
|
| 27 |
+
"revision_note": "A hosted API endpoint cannot be revision-pinned: providers update weights without notice, and temperature 0 does not guarantee determinism under batching. This is the same failure mode that left the T07 judge unreproducible (model_revision null). It is recorded in advance rather than discovered afterwards.",
|
| 28 |
+
"run_window": {
|
| 29 |
+
"started_utc": "2026-08-23T09:21:41.960309+00:00",
|
| 30 |
+
"finished_utc": "2026-08-23T10:23:04.359653+00:00"
|
| 31 |
+
},
|
| 32 |
+
"single_invocation_window": "T26 validation items MUST be drawn from this scored artifact; scoring them later would characterise a possibly different instrument",
|
| 33 |
+
"prompt_sha256": "8ebfbec0dbc681bd279fa6c96340182e45dd032d5cee7c5391b78437f60b3d96",
|
| 34 |
+
"prompt_file_sha256": "9f344086ffa93d54bb103c8a87e67071e9ebde78bf8e9a8c078a1a360de81d04",
|
| 35 |
+
"config_sha256": "a93992e28c2af21db59460c7933a04ca7eaf5b17bd26e2c6d90a53426f220287",
|
| 36 |
+
"inputs_sha256": "24a0c5a3699fd6b51a66d4be92bc054d14cb800db894c0162b65774c5667f06b",
|
| 37 |
+
"n_rows": 400,
|
| 38 |
+
"n_coverage_failures": 0,
|
| 39 |
+
"coverage_failure_rate": 0.0,
|
| 40 |
+
"all_production_coverage_failed": false,
|
| 41 |
+
"n_request_error_attempts": 11,
|
| 42 |
+
"n_rate_limit_error_attempts": 11,
|
| 43 |
+
"per_condition": {
|
| 44 |
+
"unsteered_baseline": {
|
| 45 |
+
"n": 100,
|
| 46 |
+
"coverage_failures": 0
|
| 47 |
+
},
|
| 48 |
+
"assistant_cap": {
|
| 49 |
+
"n": 100,
|
| 50 |
+
"coverage_failures": 0
|
| 51 |
+
},
|
| 52 |
+
"random_direction_cap": {
|
| 53 |
+
"n": 100,
|
| 54 |
+
"coverage_failures": 0
|
| 55 |
+
},
|
| 56 |
+
"sphere_proximity_cap": {
|
| 57 |
+
"n": 100,
|
| 58 |
+
"coverage_failures": 0
|
| 59 |
+
}
|
| 60 |
+
},
|
| 61 |
+
"canary": {
|
| 62 |
+
"n": 50,
|
| 63 |
+
"start_sha256": "6e48843fbd6222cde4afb17ee5416933654aad902a39bfcf34e3fe07b856b9f8",
|
| 64 |
+
"end_sha256": "6e48843fbd6222cde4afb17ee5416933654aad902a39bfcf34e3fe07b856b9f8",
|
| 65 |
+
"start_usable": true,
|
| 66 |
+
"end_usable": true,
|
| 67 |
+
"drift_detected": false,
|
| 68 |
+
"n_items_changed": 0,
|
| 69 |
+
"how_to_use": "Re-score this same canary slice later and compare against end_sha256. A mismatch means the endpoint moved."
|
| 70 |
+
},
|
| 71 |
+
"raw_responses_retained": "/content/drive/MyDrive/beyond-single-axis-qwen-causal-scoring/production_5b60b8e7bd32_20260823T091319Z/private/qwen_causal_raw_responses.jsonl",
|
| 72 |
+
"artifact_sha256": {
|
| 73 |
+
"scored": "d7f8feb34f8c2e8e8fff57764c95ff81b8128fbfa35d87c380322dce07565bd3",
|
| 74 |
+
"raw": "130bff99ab2888c0989154d5fbf69970e77ceb7d6f1d508b9704d58745cfa2ed"
|
| 75 |
+
},
|
| 76 |
+
"artifact_note": "The scored artifact is private and pinned by hash; item-level automatic scores joined to condition are not committed."
|
| 77 |
+
}
|