A work-around is to use my own naming scheme for the ml files and save the model as part of NEURAL_TRAIN. This way, the right file is present when NEURAL_LOAD happens.