What problem does it solve? Requests to "make a model interpretable" are ambiguous, and applying the wrong explanation technique produces misleading or unstable attributions. This Skill separates global from local explanation questions, matches explainers to model families, and guards against reading post-hoc attributions as causal claims. ## Core Features & Use Cases - Explanation Selection: Maps each stakeholder question (global drivers, single-row reasons, actionable changes, response shapes) to the right tool such as SHAP beeswarm, waterfall plots, DiCE counterfactuals, or PDP/ICE curves. - Explainer Matching: Directs TreeExplainer for tree ensembles, LinearExplainer for linear models, and KernelExplainer only for model-agnostic cases, avoiding slow sampled approximations where exact algorithms exist. - Validation Controls: Provides label-shuffle leakage checks, seed-stability tests for feature rankings, and direct perturbation spot-checks of SHAP claims. - Use Case: A data scientist must justify a churn model's decisions to stakeholders; the Skill guides producing validated SHAP attributions with recorded background data, correct log-odds units, and stability checks before presenting results. ## Quick Start Explain why my gradient boosting model predicted churn for this customer using SHAP, and check that the top features are stable across random seeds.
