Reproducibility and scientific interpretation
Keep raw data immutable and save derived data with documented transformations.
Loading video…
Keep raw data immutable and save derived data with documented transformations. Record software versions, random seeds and exclusions so another person can rerun the work. Negative results are findings when methods and uncertainty are reported clearly. Interpret only what the design supports; correlation, prediction and mechanism require different evidence.
Worked example
If changing the random seed changes a small study's accuracy from 55% to 80%, report this instability rather than selecting the best run. Preserve the original CSV and code used for each result.