On a stall, the system takes N fresh drafts of the stuck unit from the frozen contract alone — no broken code in context — across a temperature ladder, and keeps whichever the verifier scores best. Width is N.
Width 1 is no decorrelation: the retry cannot escape the anchor. Cost is linear in width; returns diminish past ~3–4 as draws start landing in the same modes. A weak or free local model affords 5–8; a strong metered model wants 2–3. The diagnostic is run-to-run variance: if drafts differ a lot, the answer is in the model's distribution — widen; if every draft is the same wrong answer, no width will help — let it escalate.