close-the-review-item-on-its-object-not-its-verb

# Close a review item on its object, not on its verb An actionable suggestion has four parts: > **verb** — cross-validate / ablate / refit / benchmark against > **object** — the model training / the combining rule / the preprocessing > **scale** — over all N labelled rows / on the small split / for three sizes > **price** — about five extra fits / one hour / one more backbone The failure this skill exists for keeps the verb and quietly swaps the object for something you already have, or the scale for one you can afford this minute. You then do real work, it is real work of the named kind, and you record the item as executed. No sentence in that record is false. The item is nevertheless gone, and items are the one class of note nobody re-reads once they are closed. ## What it looks like from inside Measured on one run, verbatim. One hour and thirty-nine minutes into a ten-hour budget, at the close of its first stage, the reviewer wrote: > *"Replace validation-argmax selection with repeated k-fold cross-validation over > train+validation. At 495 rows the current split cannot separate a 1-point gap, > which is wider than the entire spread between this run's five arms; 5-fold CV > over 4934 labelled pairs would cost about five fine-tunes and would make every > subsequent comparison decidable rather than noise-limited."* Verb: cross-validate. Object: the model training. Scale: 4,934 rows. Price: five fits. What got built, fifty minutes later, was a bootstrap comparison of ...