Corollary: generated artifacts: hand-edited output reverses it; rendered images have a reproduction limit

[yamierick]

Why

Corollary: generated artifacts runs one way: the script is the source of truth and regenerating must reproduce what shipped. It is silent on the case every sighting hit, where a person edits the output. From then on the generator destroys work, and still runs and still succeeds; a README warning is not where the runner is when they run it. Separately, the corollary plus Part 5's committed figures invite a byte comparison of PNGs that only the originating machine can pass: matplotlib writes its version into the PNG, and a tight bounding box follows font metrics, so even pixel dimensions differ across machines. The rule leaves the reader to discover that limit.

Sightings

• Hand-edited output: four sessions, all verified: workbook scripts whose README warns they overwrite the workbook; curriculum decks and a crosswalk; a deck builder that discards PowerPoint edits; a register now maintained by hand whose generator still runs.
• Rendered images: one ten-day exchange, two repositories, two independent measurements (a version string in the PNG tEXt chunk; 1388x586 locally against 1389x587 on CI).

What changes

• Corollary: generated artifacts: two paragraphs appended, hand-edited output first, then the rendered-image limit.
• Related: Part 1 rule 6 (added in Part 1 rule 6: a push is finished when its CI run has been read #30) already notes that a PNG drawn on a Linux runner is never byte-identical to one drawn on a laptop. The two agree; this paragraph states the limit where the reproduction rule lives.

Approved by Jones on 2026-09-24. Drafted in claude-notes: briefings/petition-carry-forward-2026-09-24b.md items 3 and 10.

🤖 Generated with Claude Code