Design in engineering begins with the problem of robustness-by what factor should intrinsic capacity exceed normal demand? Here we consider robustness for a neural circuit that crosses the retina from cones to ganglion cells. The circuit's task is to represent the visual scene at many successive stages, each time by modulating a stream of stochastic events: photoisomerizations, then transmitter quanta, then spikes. At early stages, the event rates are high to achieve some critical signal-to-noise ratio and temporal bandwidth, which together set the information rate. Then neural circuits concentrate the information and repackage it, so that nearly the same total information can be represented by modulating far lower event rates. This is important for spiking because of its high metabolic cost. Considering various measurements at the outer and inner retina, we conclude that the "safety factors" are about 2-10, similar to other tissues.