{
  "context": "Evaluates how well the solution applies scikit-learn's advanced unsupervised anomaly detection, robust covariance, and biclustering APIs to fulfill the AnomalySuite behaviors. Focus is on correct estimator selection, parameterization, and use of scoring/prediction methods rather than general coding style.",
  "type": "weighted_checklist",
  "checklist": [
    {
      "name": "Global detector",
      "description": "Trains a global anomaly estimator from scikit-learn (IsolationForest preferred, OneClassSVM acceptable) with the contamination argument, calls fit on the provided samples, and derives boolean anomaly flags from decision_function or score_samples using a threshold; sets random_state for reproducibility.",
      "max_score": 30
    },
    {
      "name": "Local outliers",
      "description": "Uses sklearn.neighbors.LocalOutlierFactor with novelty=True and the n_neighbors value forwarded from the API, fits on the training data, and exposes scores for new samples (via score_samples or transformed negative_outlier_factor_) so higher values indicate stronger outlierness.",
      "max_score": 20
    },
    {
      "name": "Robust covariance",
      "description": "Fits sklearn.covariance.MinCovDet (or EllipticEnvelope powered by it) on the training samples and computes Mahalanobis-style distances for queried points using the estimator’s mahalanobis method or robust covariance attributes; distances are reused without manual covariance math.",
      "max_score": 20
    },
    {
      "name": "Biclustering",
      "description": "Applies a scikit-learn biclustering model such as SpectralBiclustering or SpectralCoclustering with n_clusters from the API, fits on the provided matrix, and returns row_labels_ and column_labels_ from the fitted estimator to separate the high-valued block.",
      "max_score": 20
    },
    {
      "name": "Sklearn-first flow",
      "description": "Relies on scikit-learn fit/predict/score methods (e.g., score_samples, decision_function) without reimplementing algorithms, keeps fitted estimators cached for reuse across calls, and validates inputs through numpy/scikit-learn utilities rather than custom heuristics.",
      "max_score": 10
    }
  ]
}