Target: a ColumnTransformer pipeline with no leakage plus a top-5 SHAP story.
1. Pipeline skeleton (79 pattern)
Numeric branch (median impute + scale) and categorical branch (impute + one-hot) inside a ColumnTransformer, vectorizer/model after it — preprocessing fitted on train folds only. Anything fit on full data before splitting is leakage.
2. Leakage checklist
Target-derived columns out; date/id columns out or grouped; imputers and encoders inside the pipeline; CV split respects time/groups. Walk the checklist in the writeup, one line per item with evidence.
3. SHAP story, top-5 only
Global bar + beeswarm, then five features each with direction, magnitude, and one business sentence. SHAP explains the model, not the world — say so.
Recap
ColumnTransformer skeleton fit-inside-CV, explicit leakage checklist with evidence, and a top-5 SHAP story that distinguishes model logic from causality.