None of todays CPUs (and NIC, USB, DISK, PHONE, etc)
are even remotely trustworthy or trustable.
So there is zero reason to believe they would get even
that feature right either, nor do you have any way to
exhaustively test it. Only way you will ever have one
bit of open trust as to what is going on in your
gear is with #OpenFabs #OpenHW #OpenAudits.

1. Provide physical access to your computer means, that this is not your computer anymore.
2. Plug-in someone else hardware in your computer means, that this is not your computer anymore.
3. Running someone else software in your computer means, that this is not your computer anymore.
4. Allow someone else to administer your computer means, that this is not your computer anymore.

Ignoring that fundamental truth for the moment...
the better way would be to feed your homemade radioactive HWRNG
into the OS PRNG via the serial UART, which is also closed HW.

Bad deal. This kind of HW RNG will decrease amount of randomness in relation of half-life. Also, radionuclides create lot of another problems. Better to use quantum based RNGs, which will have the same amount of properties for a long time. Nuclear fision are only one of possible source with quantum properties. There are also noise of amplificators, fotoelectric effect, Schottky noise ... Much more efficient and much safer for use. More, better to follow regular testing on random number generator based on BSI BSI AIS 2.0/3.1 than the NIST 800-90.

Jan