Making results repeatable.
WHY IT MATTERS
An analysis nobody can reproduce, including its author, cannot be trusted or corrected.
WHAT TO RECORD
The data source, and when it was obtained Every transformation applied The code, in version control Library versions Random seeds, where randomness is used
WHY SEEDS
Otherwise results differ between runs, and cannot be compared.
WHAT TO AVOID
Manual steps in spreadsheets Undocumented cleaning Running code out of order in a notebook
WHY THAT LAST POINT
Notebook state depends on execution order, and a notebook that works for its author may not run top to bottom.
WHAT TO DO ABOUT IT
Restart and run from the top before trusting any result.
WHAT TO SEPARATE
Raw data, never modified Processed data, produced by code Outputs
WHAT TO PIN
Library versions, in a declared environment.
WHY
Statistical libraries change behaviour between versions.
WHAT TO PRODUCE
A report generated from the code, so figures and text cannot diverge.
WHAT TO TEST
That someone else can run it.
