chemistry-group-attribution-over-the-split-and-the-baseline-mode

# A group ranking from one molecule has no n, and "more interpretable" is a comparison This picks up after the estimator is chosen. `the-attribution-is-the-deliverable` covers which estimator to use, computing two of them and reporting their disagreement, the trivial-feature control, and the fact that the output has to be a figure. None of that is repeated here. What that skill leaves open, and what loses the interpretability criterion, is three things: the population, the comparator model, and the mask. The failure shape is always the same. The source drew a saliency map for one named molecule; the run computes its own map for the same molecule, sums the per-atom scores into the named functional groups, gets a different ranking from the source's, and reports the disagreement. Both halves fail. A ranking of a few groups derived from the atoms of a single molecule has no sample size, so a disagreement cannot be distinguished from that molecule's noise. And the claim under test is comparative -- the method highlights chemically meaningful substructures *more than the conventional architecture does* -- which one model's map cannot address at all. ## 1. The population is the evaluation split, not the drawn molecule Write a SMARTS pattern for every group the analysis names and match it against **every molecule in the held-out split** (`Chem.MolFromSmarts`, `GetSubstructMatches`). Report one table: - molecules containing the group (n) and matched atoms (m); - mean attribution ...