{ "agreement": { "n_labeled_rows": 136, "note": "AUC against GLM teacher scores on every labeled row (1.0 = the linear probe reads the same verdicts as the neural verifier)", "v1_weak_labels": 0.8091, "v2_teacher_distilled": 0.8879 }, "calibration": { "hi(16-19)": { "mean_teacher_score": 4.131, "n": 99 }, "junk(0-4)": { "mean_teacher_score": 2.433, "n": 30 }, "low(5-9)": { "mean_teacher_score": 3.0, "n": 2 }, "mid(10-15)": { "mean_teacher_score": 2.6, "n": 5 } }, "dedup": { "elapsed_s": { "learned": 12.89, "plain": 58.0 }, "learned_hash": { "dup_rate": 0.0, "largest_cluster": 1, "n_candidates": 39, "n_clusters": 0, "n_dup": 0, "n_in": 3034, "n_kept": 3034, "n_pairs_verified": 0 }, "plain_minhash": { "dup_rate": 0.0, "largest_cluster": 1, "n_candidates": 42, "n_clusters": 0, "n_dup": 0, "n_in": 3034, "n_kept": 3034, "n_pairs_verified": 0 }, "plain_would_merge_learned_rejected": 0, "threshold": 0.8 }, "quality_prefix_histogram": { "hi": 693, "junk": 2233, "low": 53, "mid": 55 }, "throughput": { "df_pass_s": 0.287, "engine": "pure stdlib python", "n_docs": 300, "sign_docs_per_sec": 812.6, "sign_s": 0.369 } }