“For instance, a "safe" action might have a reward of 5 and a risk of 3, which I would represent as 5±3 in this model, the return would typically range anywhere from 2 to 8 units of net benefit.”

He is literally adding risk and reward to determine the upper net benefit bound.