the parameter flatten_solvent, which limits the size of the initial model, is hardcoded on the parameter of 3d classification, and is passed through even when no mask is provided.
This provokes that the final volume is smaller than expected:

Please set it to True only when a mask is provided to avoid unexpected outputs