Estimating how much a task-environment pair's reward signal will vary, indicating how informative it is for learning.