ce-classification

npx skills add majiayu000/claude-skill-registry --skill ce-classification --agent claude-code

Same command for any agent — swap --agent for codex, cursor, copilot.

# CE Classification You are working with a classification task (binary or multiclass). CE provides calibrated probability bounds $[p_{low}, p_{high}]$ using Venn-Abers calibration. ## Prediction Semantics (Venn-Abers) The `predict` and `predict_proba` methods return calibrated class assignments and probabilities. - **Binary**: Rules and probabilities always refer to the **positive class** (index 1). - **Multiclass**: - `predict()`: Returns the class label with the highest calibrated probability. - `predict_proba()`: Returns a full probability distribution (summing to 1). - **Intervals**: For any class probability $p$, the calibrated interval $[p_{low}, p_{high}]$ guarantees that $p_{low} \le p \le p_{high}$ (ADR-021 invariant). ```python # Multiclass: explain the PREDICTED class (default) # If the model predicts 'Class 2', rules explain 'Why Class 2?' explanations = explainer.explain_factual(X_test) ``` ## Explaining All Classes (`multi_labels_enabled`) By default, CE explains the **predicted class**. To explain **all classes** (e.g., for contrastive analysis or true multi-label tasks), use `multi_labels_enabled=True`. ```python # Returns a MultiClassCalibratedExplanations object m