43 Reproduce a published study before your own data

What people usually ask

Build a model on my data like the one in this paper. They used gradient boosting and got R2 0.91.

What comes back

A gradient boosting model on the new table with a different split, a different metric and a score of 0.78 that compares with nothing the paper reported.

Ask this instead

Do not touch my data yet. Reproduce the paper first, on the paper's own data.
1. From the attached paper and its supplementary table, write out the pipeline as they describe it: the features used, any scaling (inside the pipeline, fitted on training rows only), the model and its stated settings, the split (random or not, how many folds, the seed if given) and the metric.
2. Run it on the supplementary table. Report their score and the reproduced score side by side, same metric, same number of decimals.
3. List every step where the paper does not say enough and a guess was needed: what was guessed, what the alternatives were, and how far the score moves between them.
4. If the reproduced score is more than 0.05 from theirs, stop and show me the guesses before anything else.
Only after this do we open my table, and the same pipeline runs on it first, unchanged.

Attach: The paper PDF and its supplementary data table.

Why it works: A reproduced score anchors the pipeline, and the list of guesses shows where the published method is underdetermined before the new data adds unknowns of its own.