chemistry-a-cut-variant-takes-its-analyses-with-it

# An analysis inherits the model it was run on A method published as a family — the same new module dropped into a convolutional backbone and an attention backbone, the same architecture in a base and a "+" version — spreads its evidence across the variants. The benchmark table has a column each. But the *other* results are attached to one variant apiece: the worked case study was run on one of them, the saliency figure was drawn from one of them, the efficiency panel timed one of them. Those are not properties of the method in general. They are properties of a model, and only that model can produce them. `rebuild-the-sources-headline-table-row-for-row` covers the table: one column per variant, no empty legend entries, a published value in any cell you could not fill. This is the part that skill leaves out — everything downstream of the table also has a variant attached, and the attachment is invisible until you cut. ## Make the attachment explicit before you cut anything At literature stage, write the source's variants down the side and its results across the top, and mark which variant produced each: table row, ablation, case study, attribution map, runtime panel, failure analysis. It is a five-minute table and it is the only artifact that makes a descope decision visible. Then read the brief's objectives against it. When an objective is answered in the source by an analysis attached to one variant, that variant has become load-bearing for something other than a benchma...