{
  "context": "Evaluates whether the solution leverages scikit-learn's multiclass and multioutput reduction utilities (OvR, OvO, chains) to train reproducible models that satisfy the specified behaviors. Scoring prioritizes correct wrapper selection, probabilistic handling for thresholds, and honoring deterministic settings.",
  "type": "weighted_checklist",
  "checklist": [
    {
      "name": "OvR trainer",
      "description": "Independent multi-label model is built with scikit-learn's OneVsRestClassifier using a probability-capable base estimator such as LogisticRegression to fit one binary problem per label.",
      "max_score": 25
    },
    {
      "name": "Prob thresholds",
      "description": "Label predictions for the independent model come from predict_proba (or equivalent decision function) with the provided threshold controlling inclusion, and outputs map back to the given label_names order.",
      "max_score": 15
    },
    {
      "name": "Chain reducer",
      "description": "Dependency-aware path uses scikit-learn's ClassifierChain honoring the specified order and reusing the same probability-capable base estimator to generate chained label predictions.",
      "max_score": 25
    },
    {
      "name": "Pairwise votes",
      "description": "Multiclass path is implemented with scikit-learn's OneVsOneClassifier (or similar pairwise reduction) and returns majority-voted class predictions that match the expected outcomes.",
      "max_score": 20
    },
    {
      "name": "Deterministic fit",
      "description": "random_state (or equivalent) is set on base estimators and wrappers so training and predictions are repeatable across runs, matching the deterministic requirement.",
      "max_score": 15
    }
  ]
}