I didn't test it per se. I did it the dirty way. I know from experience and information theory that a random stream is in compressible. I also know that a hash based extractor running on compression levels like this will pass all the standard tests. So I recorded the noise from one CV4014 stage with the capacitor changes you suggested. I like the glow of the valves.
60 seconds recording generated a 5.5MB WAV file. I then compressed it with 7z to produce a 4.3MB file. I then divide the size of the compressed file by 2 for luck*. This means that my prototype generates about 39% pure entropy but that's sucking it up via a 32bit / 44.1kHz PC sound card.
I'm toying with the idea of fronting the Arduino with an ADC chip. Something like a Maxim MAX166CCPP. 50kHz bandwidth and 8 bit parallel interface. Interfacing /protocol looks simple and runs off 5V. Even with an Ethernet shield, there should be enough available pins on a standard Uno. Actually, I might order one tonight...
You see that the entropy rates change depending on how I measure it. That's why most of my production rate estimates are totally rubbish 
-
- actually this is to allow for future improvements in compression algorithms. Don't forget that they're asymptotic so there is actually a lower bound for source's entropy. You see this in the literature and Shannon's own work.