And yet they do. I don't remember ever reading any machine learning paper where the authors point out that they held out a test partition that they couldn't access until some "final" step of the experiment.