A couple of things, is it me or is the design very very similar to that that Intel used in their random chip and got a patent on five or six years ago.
The second is although "anonymous" has a point about manufacturing tolerances getting better random noise sources are and will be for the foreseable future all around in that as long as there is resestance and a temprature above absolute zero then there will be thermal noise.
There are several ways of getting at thermal noise one of which is the white noise from a sufficiently sensative CW receiver tuned to a blank part of the band with the antenna replaced by a dummy load held at a suitable temprature like 100 Celsius. If you read any of the Tom Clancy books he recons this is how the NSA do it (fiction is after all a wonderful thing).
The RNG circuit I designed last centuary (just over five years ago ;) used a Zenner Diode with a small value resister in series. The PD across the resistor was feed into the input of a differential instrumentation amplifier, the current in the diode was adjusted to give a high level of noise output, this was used to drive a VCO that was used to sample the output of an Xtal oscilator (sometimes called a roulette wheel circuit).
In practice any reverse biased semiconductor junction will produce suitable noise when correctly biased (ie the BE junction of an NPN transistor). The reason for using a zenner was that it is much easier to bias for noise.
The hard part of the circuit design was to keep extranious noise out from the supply lines and capacitive and inductive coupling from other parts of the circuit (a hint if doing it yourself use good RF circuit lay out techniques, balanced design in the analogue sections with zero volt gard rings around critical parts of the circuit like the diff amp inputs oh and get the PSU noise down well below the noise level of the diode).
Hard as this was, it is a heck of a site easier than correctly using the output, which I think even Bruce will admit is a hard problem ;)