v9 — Does SA3 follow STATED order, or impose LEARNED track-shape?
SA3 is a diffusion model — it denoises the whole duration at once, conditioned globally, with no timestamp awareness. So when we write "A then B", does it obey the sequence, or just produce its learned intro→peak→outro shape? Test register: techno ↔ house (adjacent, both 4/4).
Order: same registers, both directions — if techno→house ≈ house→techno, order is ignored. Anti-shape: prompts that DEFY quiet→loud — if they still come out quiet→loud, the PRIOR wins. Grain-aligned: inpaint / continuation / single-hybrid (the "with the grain" moves). Drift: slow continuous warm↔cool, no hard switch.