Put technical reproducibility burden completely on authors

[mstimberg]

It is of course mainly the author's responsibility to come up with a workflow, but the current community process implies that the reviewer could/should invest additional work to make things reproducible, e.g. by making the repository binder-ready, creating a Makefile (see issue #19), etc.
I think it would be better to clearly separate the responsibility: the author provides the workflow (binder, Rmarkdown or jupyter notebook, Makefile, shell script, README file with step by step instructions, ...) and all the CODECHECKER does is following/executing this workflow, verifying that it works and documenting the exact environment that was used. The last part would be something like pip freeze, conda list or R's sessionInfo(), i.e. not a format that allows to reproduce things exactly at the press of a button (you'd need the same platform, etc.), but could give valuable hints if future replications fail or give different results.

[nuest]

Hi @mstimberg, thank you for pointing this out. I agree with the responsibilities your describe and understand the need for rewording the guidelines. In my own experhttps://github.com/IngaSchl/Label-Extractionience as a codechecker, I often do create extra configuration files or a Dockerfile, which sometimes is so quickly done that it doesn't feel like extra work.

I'm pretty sure it does not say "the reviewer should invest additional work", but I will take another look and try to make this more clear.

[nuest]

While adding this, we could also recommend the authors to consider using generic environments that are widely used within their community, such as rocker/geospatial images for R + Geo, or Neuro Debian for neuroscience.

HT to @rougier who pointed this out in the F1000 review.

[rougier]

This is something we plan to do for ReScience using guix. The idea is to offer a +/- standard environment for each domain such that the author has only to specify what is different.