chemistry-accuracy-and-cost-for-every-module-you-swap-in

# A module you only timed is half an arm When a study's contribution is a module swapped into an existing architecture, its claim is almost always joint: the replacement is *more expressive* and it is *cheaper*. That is one claim with two columns, and it is answered by one table in which every module the source compares — the proposed one, each alternative it is compared against, and the conventional layer being replaced — carries a value in both. An arm that appears only in the timing table cannot answer the accuracy half. It is also the cheapest number in the whole study: the arm is already implemented, the graphs are already featurised, and it is one training run from a score. Build it once, then run it through the same loop, the same split and the same estimator as the proposed module, and give it its row. `rebuild-the-sources-headline-table-row-for-row` owns which columns have to exist at all. This is about what has to be inside each of them once they do. ## One matching convention, stated, across both columns Equal width and equal parameter count are different experiments. A basis that holds 2K coefficients where a linear layer holds one is, at equal width, a bigger model; at equal parameters it is a narrower one. Decide which match the comparison is made at, write it in the caption, and use the same match on both sides of the joint claim. A run that compares accuracy at matched parameters and cost at matched width has measured two different comparisons and answered...