What Does a Protein Surface Add? Ablations Cannot Tell You

When a model's input is computed rather than given, the ablation that removes it cannot say why it helped: the channel may add information derived from the prediction target (Route A), or merely re-represent information the model already has (Route B), and both produce the same gap. No precision separates them — that is a property of the estimand, not of the sample. We give a seven-component provenance-aware protocol that does separate them, and run it on released surface-conditioned inverse folding. In two released pipelines the surface chemistry channel is a deterministic function of native residue identity: 89.12% and 90.12% of its values move under fixed-coordinate permutation, against 0.00% for geometry. For the system with released weights, a fixed code-table derangement makes it write the counterfactual protein, raising agreement with that sequence by +38.90 pp over its own pre-intervention baseline — a targeted error pattern that generic distribution shift does not explain. Across paired retraining seeds, a semantics-free random 18-class partition gains far more than scrambling biochemical values over the same classes: the benefit follows target-identifying partition structure, not mainly biochemical meaning. At a matched ≈17% marginal ceiling, target-derived channels add +12.49 pp to frozen ProteinMPNN and +11.07 pp to frozen ESM-IF1 while backbone-derived channels are ignored — holding unrefitted on an external RCSB 2023+ panel across three frozen families. Joint surface instruments remain below our Route-B threshold, while a backbone-derived positive control clears it. Four of our own predeclared predictions were falsified and a fifth claim family was retracted; all five are retained. The claim is a provenance burden: whenever a method computes an input from data involving the prediction target, an ordinary ablation cannot justify a representation claim until source interventions separate added target information from design-valid computation.