You could also have a complex training job that trains a shallow model for a while, then uses it to train a deeper model, or extract an embedding from some layer and train a classical prediction model that uses the embedding as the feature vector.

But the point is that the right way to ensure safety is with realistic fixture-based integration testing. That’s not what static typing is for in that type of use case and is not a de facto benefit of static typing.