What happened
A useful fraud explanation has to describe the feature vector the model actually scored. Notebook-specific transformations made that guarantee too easy to break.
The correction was to freeze preprocessing with the selected model and load the same artifacts in training, serving, and SHAP. Reproducibility became a shared contract instead of a convention each entry point could interpret differently.