You can look all over the universe for entropy. Lava lamps make for a cool PR showpiece, and apparently easy fodder for fluff pieces like this article. I just hope none of my relatives read it and ask me about it. The reality is much more interesting but would probably bore most people to death.
Read the voltage drop across a high-value resistor at room temperature. There are random fluctuations proportional to the temperature and resistor value. You can use these fluctuations to generate unifromly distributed random numbers. https://en.m.wikipedia.org/wiki/Johnson%E2%80%93Nyquist_nois...
Flip a penny twice, if you get heads followed by tails, write down 0. If you get tails, followed by heads, write down 1. Otherwise ignore the two flips. Rinse and repeat until you have enough bits.
The purpose of flipping twice is that it offsets any potential static bias (caused eg by weight difference) between the two sides.
Common ways to generate entropy in a way that’s friendly to computers include: measuring noise from a zener (or other electronic component), delta between two or more oscillators, or timing between external events (eg keystrokes, packets, etc). These are all super cheap to manufacture (~$1). If you want to go high end, you can spent ~$1000 and get a generator based on some quantum property (photon spin or background radiation detector).
Computers will typically combine two or more sources of entropy. In newer tls version handshakes, the entropy from both, the server and client comes into play. So there’s ways to build defense in depth.
There is a fun generalization of the game. Assume the penny has (unknown) bias p. You want to output a new flip with bias some function f(p). So f(p)=1/2 is what you described, ie how to get a fair flip. For some functions it can be even easier, e.g. f(p)=p^2 - definitely only requires two flips.
How about:
f(p)=p^2/(p^2+(1-p)^2)
f(p)=2p(1-p)
f(p)=3p(1-p)
f(p)=sqrt(p)
The last two are tricky, I took them from a paper "functions arising from coin flipping" by Wastlund. (The quantum generalization is even more fun, but this text box is too small to contain it.)
Testing for randomness is impossible in the YES/NO sense, so the best we have is statistical test batteries like NIST's 800-22, DieHarder and TestU01.
I own a Truerng and a Onerng, both are under $50. They performed better on Dieharder than my computer's Intel RNG. Not by a huge margin, but still. All 3 passed the minimum requirements of course.
Both the radio and the CCD seem like something a determined actor could undermine. Transmitting a directional pop of radio static at just the right moment, or whatever causes CCDs to show static (x-rays?)
Multi-colored so your own biases don't affect the order you read the dice in. E.g., and I'm just guessing, some people might notice the higher numbers first and read those out.
