My analysis was reproducible. My file paths were not.
I had a case study I was fairly happy with. Cleaning documented transformation by transformation,
counts reconciled with an assert, seven verification checks, the obvious alternative explanation
tested and ruled out. The exit gate for the phase had one unchecked box: analysis reproducible
from the raw file by a third party. I had been treating that as paperwork.
It was not paperwork. When I finally tried to run the scripts on my own machine, none of them worked. Every one started like this:
RUTA_BASE = '/sessions/gifted-dazzling-albattani/mnt/Videojuegos'That is the path of the temporary environment the code was written in. The notebook was worse: it had been stitched together across two different sessions, and carried patches rewriting one sandbox path into another mid-line.
The fix is three lines, and it is the same in any language:
from pathlib import Path
RUTA_BASE = Path(__file__).resolve().parents[1]Resolve paths from where the script lives, not from where it once ran. In Node,
path.resolve(__dirname, '..'). In Markdown, relative image links — one of my deliverables had
four broken images for the same reason.
What I take from it is not the fix, which is trivial, but the failure mode. An analysis is not reproducible because the logic is correct. It is reproducible because someone other than you can run it. Those are different claims, and only one of them can be verified by reading your own code.
So the check went into my method, with the command that performs it:
grep -rn "/home/\|/Users/\|C:\\\\\|/sessions/\|/mnt/" --include="*.py" --include="*.md" .A gate you never run is a gate you do not have.