Causality preserving must be generic and externally defined.

[Joao-L-S-Almeida]

The loss function templates should receive the causality preserving algorithm as an input argument.
I would allow the definition of generic approaches for preserving causality.