publish-what-the-run-already-computed

# The most expensive result is the one you computed and did not show The work was done. The compute was spent. The number is on disk. And the report does not contain it — because the report's contents were chosen by the run's own hypothesis structure, and a quantity produced in service of a hypothesis but not adjudicating one has no section to live in. This skill is the sweep that finds those. It is about the run's *outputs*; `the-supplied-item-is-the-graded-unit` is about the *inputs* — the named objects in `data/` — and covers printing those objects rather than pointing at them. Run both. This one catches what that one cannot: a quantity nothing shipped and nothing asked for by name, which the run computed anyway because the analysis needed it. ## The sweep Before you fix a section order or polish a sentence, do this once: 1. List every file the run wrote under the results and outputs directories, with what it *holds* — not the path, the content: "per-epoch pre-training loss", "sequence identity per aligned chain pair", "the 497-row permutation-importance table", "seconds per optimiser step for the long-range and short-range heads". 2. Against each, write the section of the report where it appears. Grep the report for a number out of the file; if none is there, the entry is empty. 3. For every empty entry, ask: **would a reader checking the task's outputs, or the source study's own results, look for this?** If yes, it goes in the body before you do an...