Avalanche noise circuits have several issues as random number generators. The first, is that they exhibit a fair degree of bias, regardless with how you capture the noise (ADC or comparator).
My experiments have indicated that two stages of 'whitening' are required to remove the bias from the readings. I use a Von Neumann filter followed by either some form of linear feed back shift register (LFSR). In general I have found that it requires at least a 4:1 whitening ratio to get a reasonably uniform RNG. For LFSR, I have used either a simple XOR'ing of consecutive bytes (which worked for one example circuit with A2D) to a Jenkings one-at-a-time hash on my second implementation which required more whitening due to a different noise source (5V zener source, no 10-15V supply needed).
The second issue with avalanche noise random number generator circuits is that the balance of 1's to 0's will change over time. If your code doesn't accomodate such changes your RNG will stop producing uniform RN when the bias exceeds your whitening technique's ability to cope. As you can see from the pages below, I noticed LARGE changes in the circuits balance point as it aged.
If your interested in my experiments, which include the code I derived from Rob Sewards median calculation approach are available from
https://code.google.com/p/avr-hardware-random-number-generation/wiki/AvalancheNoise