Reward uncertainty (side-effect mitigation) — trades penalty for uncertainty relative to → Penalize influence

Penalize influence, like the impact regularizer it sits beside, works by adding an explicit term to the objective and hand-tuning how much it should cost the agent to acquire capability. Reward uncertainty abandons that architecture entirely. Instead of a fixed reward function plus a penalty, the agent holds "a prior probability distribution that reflects the property that random changes are more likely to be bad than good" (concrete-problems, §"Reward Uncertainty:", pp. 6-7), so caution about disruptive or influence-seeking behavior falls out of ordinary expected-value reasoning rather than being separately engineered in. The two proposals sit as adjacent bullet points addressing the same problem, but where penalize influence still needs someone to decide how much to penalize, reward uncertainty needs only a prior belief that unplanned change tends to be bad, then lets normal decision-making do the rest.