That isn't it. When hardware rng is initialized it first runs tests on the randomness to prove that the hardware for getting entropy isn't broken. Then it feeds

That isn't it. When hardware rng is initialized it first runs tests on the randomness to prove that the hardware for getting entropy isn't broken. Then it feeds that entropy into a CSPRNG which gets exposed to code trying to use hardware rng. That person's point was that it's impossible to know for sure if entropy collection is broken. In practice it isn't an issue as you can make the false positive rate very small even if it can never be 0.