the-attribution-is-the-deliverable

# If the brief asks what the model looked at, the answer is a map, not an opinion A task that lists interpretability among its objectives is asking for an object: a contribution per input unit — per feature, per atom, per bond, per region, per token — aggregated into the domain's own groups and drawn. Everything else is a substitute, and the substitutes are all cheaper than the thing. Three substitutes show up reliably, and each one scores as an absence: - **A different importance measure.** Gini or impurity importance, permutation importance, and attention weights are properties of the fitted model or of the training procedure. The brief asked which *inputs* drove which *predictions*. - **A measurement on the model's parameters.** A spectrum of learned coefficients, a rank of learned filters, an ablation of architectural blocks: all interpretable, none of them an attribution. - **An argument that the method is not interpretable.** This is the most expensive one, because it is often correct and it still delivers nothing. The finding that a method's explanations are unstable is a *result of running the attribution*, not a substitute for running it. ## Which estimator Name the mapping explicitly at design time, in one line: | the model | the estimator | |---|---| | fitted tree / gradient-boosted / tabular | SHAP (TreeSHAP), on the same rows the metrics use | | neural network over graphs, images, sequences | input gradients ‖∂ŷ/∂x‖, occlusion or ablation, sub...