{"base_learner_scores":{"dbscan_is_anom":0,"dbscan_score":0.0,"n_measurements":1,"p_cat_ANON":0.0012,"p_cat_COMT":0.0013,"p_cat_GRP":0.0038,"p_cat_NEWS":0.0005,"p_classifier":0.0,"p_corroboration":0.0001,"p_measurement":0.0,"p_method_http":0.0141,"p_method_tls":0.0007,"stl_zscore_prob":0.9898},"contributions":{"dbscan_is_anom":{"coefficient":0.27953589934341944,"log_odds_contribution":0.0,"value":0.0},"dbscan_score":{"coefficient":0.1862285084787871,"log_odds_contribution":0.0,"value":0.0},"p_cat_ANON":{"coefficient":0.44832241215565233,"log_odds_contribution":0.0005,"value":0.0012},"p_cat_COMT":{"coefficient":0.4440849444832184,"log_odds_contribution":0.0006,"value":0.0013},"p_cat_GRP":{"coefficient":0.6499264749123036,"log_odds_contribution":0.0024,"value":0.0038},"p_cat_NEWS":{"coefficient":0.1579805347980017,"log_odds_contribution":0.0001,"value":0.0005},"p_classifier":{"coefficient":1.556924495454946,"log_odds_contribution":0.0,"value":0.0},"p_corroboration":{"coefficient":-0.24027021984935615,"log_odds_contribution":-0.0,"value":0.0001},"p_measurement":{"coefficient":-0.9817565075568403,"log_odds_contribution":-0.0,"value":0.0},"p_method_http":{"coefficient":0.1223721538036462,"log_odds_contribution":0.0017,"value":0.0141},"p_method_tls":{"coefficient":0.8085427027144448,"log_odds_contribution":0.0006,"value":0.0007},"stl_zscore_prob":{"coefficient":-0.014033211637206453,"log_odds_contribution":-0.0139,"value":0.9898}},"country":"RU","date":"2026-10-03","honest_caveats":["NEGATIVE FINDING: LOCO median F1 = 0.8750, required >= 0.88. Cross-country generalization of the meta-ensemble is below v3.3's LOCO floor.","All base-model predictions are IN-SAMPLE \u2014 they were generated from models trained on (parts of) this same labeled set. A truly honest stacker would use 5-fold OOF predictions from each base model.","Base models have CORRELATED errors (per-method/per-category share features with v3.3; per-measurement is computed from the same evidence). Coefficient magnitudes can flip sign across folds when redundant signals collide.","Per-category and per-method models failed their own promote floors on F1 (0.40-0.50). Their probabilities here may dilute rather than enrich.","Isotonic is fit on the full dataset \u2014 final calibration is in-sample. Per-fold isotonic during CV is honest; the served calibrator is not."],"intercept":-1.6584940350130268,"meta_ensemble_probability":0.0274,"meta_ensemble_probability_uncalibrated":0.0483,"note":"Additive endpoint; v3.3 remains the default classifier at /v1/classifier/score/{country}.","passed_promote_gates":false,"schema":"voidly-meta-ensemble-score/v1"}