{
  "context": "Evaluates whether the solution uses scikit-learn's built-in feature selection APIs to implement the filter, model-importance, and recursive elimination utilities described in the spec. Focus is on correctly wiring mutual information scoring, model-based thresholds, and recursive elimination rather than manual calculations or ad-hoc masking.",
  "type": "weighted_checklist",
  "checklist": [
    {
      "name": "Mutual info selector",
      "description": "Implements filter-based ranking with sklearn.feature_selection.SelectKBest (or equivalent) configured with mutual_info_classif to score columns and derive the ordered top-k feature names.",
      "max_score": 30
    },
    {
      "name": "Selector support usage",
      "description": "Uses the selector's scores_ and get_support output (not ad-hoc sorting) to cap k at available columns and preserve mutual-information ordering when returning feature names.",
      "max_score": 15
    },
    {
      "name": "Model-based selection",
      "description": "Applies sklearn.feature_selection.SelectFromModel (or equivalent) with an estimator exposing feature_importances_ or coef_ (e.g., RandomForestClassifier, LogisticRegression with L1) to drop low-importance features and respect the requested top_fraction.",
      "max_score": 25
    },
    {
      "name": "Fraction thresholding",
      "description": "Computes the keep-threshold from top_fraction via the selector (threshold or max_features) so that roughly the requested proportion is retained, rounding up to at least one column.",
      "max_score": 10
    },
    {
      "name": "Recursive elimination",
      "description": "Uses sklearn.feature_selection.RFE or RFECV with a supervised estimator to iteratively remove features until the keep count is met, returning the retained names ordered by the estimator's ranking_ or support_.",
      "max_score": 20
    }
  ]
}