The problem with true random is that you might screw up your avalanche diode circuit (or whatever) in a way that makes it random, but not uniformly distributed. It might also pick up EM from somewhere and not "really" be random all the way. By the time you add all the whitening and that sort of thing to make those problems go away I'm not sure how the information theory proofs hold up. Probably what you would want to do is keep the physical random number generator in a fine-mesh Faraday cage in a controlled environment, and run statistical tests for a while before generating your pads. The rng-tools package on Linux has a test tool. Make sure you disconnect the test wires before generating the pads though!