The question
The reading: “the transport model should preferably impose constraints that a generic predictive model need not satisfy. For example, if transformations are composable, can the model predict the result of a previously unobserved composition from its constituent transformations? That is a much stronger test than fitting responses to individual source–target pairs. It moves the contribution from richer features toward generalizable structural constraints.”
Why the record needs it
CLAIM-202’s test fits and predicts on held-out renderings of the same transformations. A strong contextual baseline (CLAIM-205) could match it there. A prediction about an unobserved composition is one a generic predictor has no structure to make. This refines QUESTION-017, which asks how distortions compose.
The record limits the test. CLAIM-260 (Active) holds that approximate transports compose only under a triangle inequality and nonexpansive kernels, and QUESTION-017 is answered in part by it. Outside those conditions the formalism predicts no additive bound, so a failed prediction there would not count against it.
What would answer it
A study that fits the transport structure on renderings under transformations A and B separately, predicts receivers’ recovery under A then B, and compares the prediction with the strong contextual baseline fitted on the same data, for transformations shown to meet CLAIM-260’s conditions.