{"architecture":"ExtraTreesClassifier (Arch3 LazyPredict-best of the liver toxicity programme)","deployment_training_set":{"duplicates_dropped":1769,"label_balance":{"negative":888,"positive":1003},"label_conflicts":"42, all inside train_locked (the conflict groups the earlier merge had excluded); first train_locked row's label kept, all logged in the build log","n_unique_compounds":1891,"rule":"ALL labeled liver data the programme holds, identity-deduped by InChIKey connectivity of the standardized parent: train_locked.csv (1,769 rows) + train_enriched.csv additions + internal_test.csv + EVERY regulatory benchmark compound (withdrawn, clinical failures, PASS), benchmark labels governing conflicts"},"features":"fp_desc block, 1,404 columns: Morgan fingerprint bits (1024) + MACCS keys (167) + DescPrep-prepared RDKit descriptors","flag_threshold":0.5,"hard_gate":{"clinical_failures":"10/10","pass":"304/304","standard":"every deduped regulatory benchmark compound verified compound-by-compound (InChIKey connectivity) inside the ACTUAL training matrix (post-featurization), before training","verdict":"PASSED","withdrawn":"71/71"},"honest_benchmark_scores_benchmark_version_only":{"basis":"385 scored regulatory compounds (71 withdrawn + 10 clinical failures + 304 PASS), lockout verified 0","clinical_failures_caught":"10/10","note":"These are the LOCKED-wave benchmark model's scores (regulatory benchmark locked out of its training). The deployment refit in this repo absorbed the benchmark and is NEVER scored on it; these numbers are the frozen scientific claim for the recipe.","pass_correct":"97/304 (0.319)","pooled_regulatory_auroc":0.69,"pooled_regulatory_balanced_accuracy":0.623,"validation_mcc_locked_wave":0.273,"withdrawn_caught":"65/71"},"hyperparameters":"sklearn ExtraTreesClassifier defaults with random_state=42 (the arm's own _make() recipe, reused unchanged; no re-tuning)","model_name":"liver-arch3-extratrees-deployment","screen_applied":"DPP-4 potent leads: 1,113/1,113 scored, 1,082 flagged (97.5%); full 4,898 potent union: 4,898/4,898 scored, 4,722 flagged (96.5%). High flag rates are this programme's known high-sensitivity/low-specificity profile; the dual-model flag (with the GAT deployment model) is the highest-concern readout.","task":"binary classification: drug-induced liver injury (DILI) concern, probability of the positive (toxic) class","weights_file":"model.joblib","weights_format":"joblib bundle: {'model': fitted ExtraTreesClassifier, 'prep': fitted DescPrep (descriptor preprocessing, fit on the full deployment set), 'kind': 'fp_desc'}. Unpickling requires deployment_common.py (shipped in this repo) to be importable, for the DescPrep class."}