The Explanation Is Another Model: What SHAP Actually Attributes

SHAP is often treated as though it reveals the reason behind a prediction. This paper tests a narrower claim: SHAP allocates the difference between a frozen model prediction and a reference expectation under an explicit rule for feature absence. Seven tree systems developed across Papers 09–15 are reconstructed without changing their hyperparameters, stages, representations, or officially selected endpoints. A common development-only background containing 50 representative vehicles and a shared interventional permutation estimator are used to explain 307 development observations from 1970–1979 and 85 temporal observations from 1980–1982. All seven prediction reconstructions match their published temporal MAE values to within 1.33 × 10⁻¹⁵ mpg. The maximum additive reconstruction error is 1.42 × 10⁻¹⁴ mpg. Model year ranks first in five systems, while weight ranks first in the unpruned standalone tree and Random Forest. Mean pairwise global rank correlation is 0.838, with a minimum of 0.500 between CatBoost and the Decision Tree. Mean local sign agreement is 0.787. The results show that an explanation can satisfy the additive SHAP identity exactly while still varying materially across predictive models, explained periods, background distributions, and missingness assumptions. SHAP therefore provides a mathematically structured allocation of predictive credit; it does not establish a unique causal explanation of how the observed outcome was produced. The paper introduces the Attribution Validity Map and includes global and local attribution analyses, cross-model agreement measurements, background-sensitivity audits, temporal attribution comparisons, causal claim boundaries, integrity checks, and a complete reproducibility package.