We observed numerical difference between coupling TensorFlow directly using https://github.com/Cambridge-ICCS/fortran-tf-lib and using forpy. It is suspected that this is due to floating-point re-association, but we'd like to confirm this more conclusively and check there isn't any other bugs we missed.