An excellent article full of good suggestions. I appreciated that it's less certain of the Best Practices TM than many comments on this subject. I am curious how the goals/techniques for reproducibility change with the percentage of software/computational work that a scientific project contains. It feels like as the percentage of a paper's ultimate conclusions that are computationally derived increases, the importance of strict "the tests pass and the numerical results are identical" reproducibility also increases. Most of my projects are mixed wet-lab/dry-lab - a fair amount of custom code is required, but it's usually less than 50% of the work. When I'm relying on other papers that have a similar mix of things, I'm often not interested if the continuous integration tests of their code pass. I am more interested in understanding well the specific steps they take computationally and in a sensitivity analysis of their computational portion (if you slightly alter your binning threshold do you still get that fantastic clustering?). I believe this is because in my field (microbiology), computational tools can guide, but physical reality and demonstrated biology are the only robust evidence of a phenomenon/mechanism/etc. For most research I do not demand tests of all the analytical pieces they are relying on (was their incubator actually set to 37C? was the pH of the media +- 0.2? etc) - I trust they've done good science. Why would I demand their code meet a higher standard?