1. python -m app.qa_v2 out.mp4 --overlay work/ovl --text onscreen.txt --font <font.ttf> --sheet work/sheet.jpg
2. VIEW work/sheet.jpg and score it with qa_v2.VISION_RUBRIC (any vision-capable model, or yourself). Save vision.json; re-run with --vision vision.json.
3. Each failing check names its fix: distinct_shots → add cuts/images; loudness → loudnorm; safe_zone → move element; glyphs → change font; vision fixes → apply verbatim.
4. Max 3 loops; then write ESCALATION.md (metrics, sheet path, three questions). Never ship on a skipped review.