SHAP attribution accuracy is the wrong metric for regulated AI. σ_SHAP — variance across K rotated background samples — is the defensible alternative.