run-the-conditions-the-source-ran

# The source's own systems, scenarios and stress tests are your experiment list A paper's results are attached to specific things: named systems, named scenarios, a named molecule, a named problem, a named noise axis, and the conditions the method says it needs. A checklist for reproducing that paper is written from those names. A better-designed experiment on different conditions answers a question nobody asked. ## The failure this prevents A run reproducing a feature-selection method replaced the paper's robustness experiment — degradation under falling signal-to-noise, reduced library size and increased dropout, against two named baselines — with its own structured confounder and batch-geometry design. Better science, in the abstract, and the reviewer wrote: *"it does not perform or report any simulations varying SNR, library size, or dropout, nor explicitly compare performance degradation curves versus Laplacian Score or MCFS."* That requirement scored **5**. A plain agent that simply ran the paper's sweep scored **65** on the same requirement. The same shape recurs, and it is never laziness — it is always a substitution made for a good local reason: * A chemistry run judged that saliency maps are not interpretable enough to be worth computing, and argued the point instead of computing them on the paper's own molecule: 5 against 70. * A climate run carried two of three named SSP scenarios and dropped the third; the requirement that names it scored 12 against 38...