THE SHORT ANSWER
Four lines: Score, Rows, Name, Undo. Lesson 08 of Becoming Forward Deployed Engineer (October 26, 2026) argues that the eval score is a field of the prompt version, written by the same function that ran the eval, because a Thursday edit that removed four lines read fine in review and two weeks later the adjusters said the answers got worse. The score is right and not enough. A classifier I shipped scored 89% on held-out data and 61% in production within six weeks with the same prompt, so the version records the row set it was scored on, count and date. My own rule library has 51 commits since March 30, at least 21 with an agent co-author line and 30 I cannot attribute, so the version records who changed it, person or agent, and who read the diff. And it records the version it rolls back to, because the last loop I ran with no re-run drifted 14% in five weeks.
Four lines: Score, Rows, Name, Undo. The text alone tells you what changed and never whether it helped, which is Lesson 08's point, and the score alone tells you it was once good on rows you can no longer find, which is mine.
The four lines, the 89% to 61% classifier, and the 51-commit count are in The Prompt Commit: Score, Rows, Name, Undo Under Every Prompt Change. The layer this applies to is drawn in Same Model, Twice the Bill: The FDE Codebase Contract, and the commit count is from Who Writes Your Agent's Rules? I Counted the Commits.