{ "task": "Predict observed attested versus disagreed outcomes from request fields only", "dataset_configuration": "outcomes", "dataset_snapshot": "2026-10-08", "split": "validation", "examples": 1133, "positive_class": "disagreed", "positive_examples": 18, "threshold": 0.5, "evaluation_sha256": "3a4c46875f2f18bce7064e8db60460cdbd40ef5f963bab35127477913eeb1dce", "train_sha256": "e10e504a46396b09f2bca44add65cda30a54cd10151662d844985a8931241836", "models": [ { "model": "Always attested", "metrics": { "n": 1133, "positive_count": 18, "disagreement_prevalence": 0.01588702559576346, "accuracy": 0.9841129744042365, "precision_disagreed": 0.0, "recall_disagreed": 0.0, "f1_disagreed": 0.0, "balanced_accuracy": 0.5, "average_precision_disagreed": 0.01588702559576346, "brier": 0.01588702559576346, "confusion": { "tp": 0, "tn": 1115, "fp": 0, "fn": 18 }, "threshold": 0.5, "warning": null } }, { "model": "Qwen3.8-27B FP8", "metrics": { "n": 1133, "positive_count": 18, "disagreement_prevalence": 0.01588702559576346, "accuracy": 0.9832303618711385, "precision_disagreed": 0.3333333333333333, "recall_disagreed": 0.05555555555555555, "f1_disagreed": 0.09523809523809523, "balanced_accuracy": 0.5268809167912307, "average_precision_disagreed": 0.07070410823444803, "brier": 0.01615286849073259, "confusion": { "tp": 1, "tn": 1113, "fp": 2, "fn": 17 }, "threshold": 0.5, "warning": null } }, { "model": "GLM 5.3 Flash", "metrics": { "n": 1133, "positive_count": 18, "disagreement_prevalence": 0.01588702559576346, "accuracy": 0.9514563106796117, "precision_disagreed": 0.08888888888888889, "recall_disagreed": 0.2222222222222222, "f1_disagreed": 0.12698412698412698, "balanced_accuracy": 0.5927254608868959, "average_precision_disagreed": 0.058693395682905145, "brier": 0.03666557811120812, "confusion": { "tp": 4, "tn": 1074, "fp": 41, "fn": 14 }, "threshold": 0.5, "warning": null } }, { "model": "GLM 5.3", "metrics": { "n": 1133, "positive_count": 18, "disagreement_prevalence": 0.01588702559576346, "accuracy": 0.9470432480141218, "precision_disagreed": 0.1111111111111111, "recall_disagreed": 0.3333333333333333, "f1_disagreed": 0.16666666666666666, "balanced_accuracy": 0.6451420029895366, "average_precision_disagreed": 0.0719890311070724, "brier": 0.03432753751103268, "confusion": { "tp": 6, "tn": 1067, "fp": 48, "fn": 12 }, "threshold": 0.5, "warning": null } } ], "fine_tuning_performed": false, "test_set_evaluated": false, "settings_note": "Qwen reasoning disabled, max_tokens 256; GLM models low reasoning, max_tokens 4096. GLM 5.3 uses Sail Research FP8, strict JSON schema and up to two format retries; Flash uses floor routing and JSON object mode. Serving settings differ." }