Measures how well a truncated interaction decomposition reproduces the model's own outputs across all subsets of the features being removed. A high value licenses reading the decomposition as the model's mechanism rather than as an approximation of it.