chemistry-reproduce-the-scoring-path-before-you-replace-it

# Implement the scoring rule you intend to condemn A leaderboard number is not the model's output. It is the output of a *path*: a split, a subset of its rows, an order, a batch loader that may discard a partial batch, a checkpoint or an epoch chosen by some rule, an aggregation across tasks or targets, a statistic across seeds. Every one of those steps is worth points of metric, and any one of them left out of your re-implementation goes into your reproduction gap without a label on it. So enumerate the path before you compute anything. Read it off the source's training script and methods section and write it down as ordered steps in `notes/scoring_path.md`: which split, which rows of it, in what order, how many are dropped, at which epoch, aggregated how, over how many runs. Where the paper and the code disagree, record both readings as separate steps — that disagreement is a finding you will want later. Transcribing the caption's protocol is one line of `rebuild-the-sources-headline-table-row-for-row`'s literature-stage step; this is what to do with it afterwards. ## Implement every step, including the ones you can see are wrong The first column of your reproduction table is your model scored through the source's exact path, standing next to the published value. That column is the only thing in the report that lets a reader decide whether your re-implementation is the same model as theirs. Everything else you do — a better estimator, a fairer aggregation, a validation-...