The Curve Stays Where the Data Lives: What Accumulated Local Effects Actually Accumulate

Accumulated local effects (ALE) are often presented as a dependence-aware alternative to partial dependence. However, an ALE curve still depends on the fitted model, interval boundaries, local sample occupancy, accumulation rule, centering population, and period of evaluation. This paper applies first-order marginal ALE to seven frozen tree-based regression systems using the Auto MPG dataset. Development contains 307 vehicles from 1970–1979, while 85 vehicles from 1980–1982 form a temporal benchmark. Ten development-defined quantile bins are requested for six numeric features, with ties reducing the effective number of intervals to between 3 and 10. Weight has the largest ALE range in six of the seven models, while horsepower ranks first for AdaBoost.R2 in both periods. Using the exact partial-dependence protocol frozen in Paper 19, the mean PDP–ALE feature-range rank correlation is 0.727, with a minimum of 0.257. Mean development-to-temporal ALE rank correlation is 0.853. On average, 24.6% of development-defined intervals are empty in the temporal population, with model year reaching 87.5%. The results show that ALE remains closer to occupied data than global feature replacement, but it is not assumption-free. Bin definitions determine locality, occupancy determines which local differences exist, accumulation determines the path, and centering determines the zero level. ALE remains a description of fitted model behavior rather than a causal response function. The deposit includes the complete paper, source code, data, machine-readable outputs, five generated figures, bin-sensitivity and temporal-occupancy audits, a SHA-256 integrity manifest, and an automated reproduction verifier.