AI video has a dirty secret. The output is sealed. Generate a clip with Sora or Veo, spot one wrong number at the 40-second mark, and your only move is regenerating the whole thing and hoping the other 59 seconds survive the reroll. Diffusion models render pixels, and pixels have no handles. There's no word to swap and no timing to nudge. You prompt and you pay, and if one detail is wrong you pay again. Sora 2 Pro bills $0.70 per second at 1080p, which makes a one-minute explainer $42 per…