METHOD.
1. Diagnose: run the unedited network on practice items, see where it scores badly and read its outputs.
2. Before any edit, fix two disjoint item sets and save their ids in /work: DEV for choosing and tuning edits, HOLDOUT only for confirming them. Never choose or tune on HOLDOUT.
3. Localize causally: a component that differs between good and bad items may only reflect how hard they are; show that changing it changes the behaviour you predicted, at several strengths.
4. Measure with paired passes ("both") at full size. Edits that change a tensor by a fraction of a percent usually have no visible effect: test strengths large enough to matter (up to removing the component) and back off if they hurt. Judge edits by the score.
5. Freeze your best candidate before using HOLDOUT and measure it there. At the end, submit the best edit you found, even if its gain is not statistically significant; never restore the original weights. Report its measured gain with its standard error honestly.
6. Before every pass, write in /work/notes.md the exact current edit (tensors, what changed, strength) and the item set; keep the notes current (they are shown to you again if the conversation is restarted) and copy good edits (e.g. /work/edited to /work/best_1/).
7. Print compact summaries, never whole arrays; output is shown in ASCII.
