Perhaps require making a docker container and writing a test script that includes inputs and expected outputs. There could be a website that loads those docker images and shows that the tests pass. Tests could be referenced within the paper. Would still require scrutiny to make sure that code doesn't hard-code the outputs.