How can we distinguish between causal and spurious features in a pre-trained model? Existing methods require access to the training process, making post-hoc analysis impossible. We introduce the \textbf{Normalised Sensitivity Ratio~(NSR)}, a post-hoc, model-agnostic diagnostic for this question under a structured-shift regime: environments differ primarily in the mean of spurious features while the causal mechanism and causal marginals remain stable, akin to multi-site clinical data or multi-batch genomics. In this regime, causal features induce constant model sensitivity across environments, while spurious features track shifts. NSR formalizes this as the squared coefficient of variation of per-environment sensitivity. Under a linear structural causal model (SCM) with $K \ge 3$ non-degenerate environments, NSR achieves exact identification (Theorem 1). We fully characterize failure scenarios: weak shifts ($O(\varepsilon^4)$ collapse), degenerate geometry, and proxy attenuation ($O((1-\alpha)^4)$), providing practitioners with quantitative criteria for assessing whether the regime holds. Finite-sample rates are $O_p(n^{-1})$ under the null and $O_p(n^{-1/2})$ under the alternative. Experiments confirm all theoretical predictions on synthetic data (area under the ROC curve [AUROC] $= 1.000$ under conditions satisfying the regime), show consistent rankings across five model families (Kendall $\tau \ge 0.529$), and recover six of eight causal features on bike-sharing data (Precision@7 $= 0.75$) without modifying any trained model.
Blogger's Review: The NSR method proposed in this paper provides a robust tool for post-hoc causal feature identification, particularly suitable for multi-environment data analysis. The combination of theoretical and experimental results demonstrates its effectiveness in practical applications, making it promising for widespread use.