- Search current sources (web search; vidIQ cache first). Record each fact: claim, number, URL, date.
- Mark hypotheticals ("if X stopped") and keep the label in the on-screen text.
- Write 6–8 beats of ≤ 16 words. Beat 1 = hook ≤ 9 words, a question with stakes. Final beat echoes the hook and names the next video.
- Output
spec.json(beats: id, say, on_screen, source_ids) +fact_ledger.json. Fail if any numeric claim lacks a source. - Run
glyph_auditon all on-screen text against the caption font.