the-reproduction-is-a-hypothesis

# "This reproduces" is falsifiable, and it is usually the hypothesis that matters A hypothesis has to be able to fail. That test is easy to apply badly to a reproduction, because "we will re-implement the paper" is an engineering goal with no plausible failure and no surprise in either direction — so the drafting stage rejects it, and the run goes looking for a question of its own. The question it finds is often a good one. It is also, reliably, not the task's question, and once it is frozen it becomes the only contract the run has. Every later stage adjudicates it, the figure plan is drawn around it, and the report is organised by it. The reproduction the task actually asked for survives as a subsection, or as a comparison table, or not at all. ## Write the reproduction as a hypothesis, properly The engineering-goal objection is right about the *phrasing* and wrong about the *content*. "We will re-implement X" fails the test. These do not: - **H: The published value is recoverable from the supplied inputs.** Supported if our estimate falls within the source's stated uncertainty of its value; refuted if it differs by more than that. This can fail, it fails often, and when it fails you have a finding the field would want. - **H: The published effect survives the ablation the source did not run.** The source reports A beats B; our reproduction holds the budget equal and asks whether the ordering survives. Refutable by construction. - **H: The mechanism the source...