John Kelsey, Computer Scientist, NIST, United States
SP 800-90B - think about how random bits should be generated.
DRBG should always be between entropy source and the attackers. Entropy just gives you bits... with entropy (as per the sources promise).
SP 800-90B is not AIS31... though the two groups are talking.
Noise sources are where the entropy comes from, health tests verify the noise source and and conditioning.
Noise sources must be non-deterministic, often uses ring oscillators. You have to be able to describe this in detail. This is complicated, as many vendors are relying on someone else's noise source. They either don't know or there is an NDA around it - that won't get validated. Submitter also has to provide entropy estimate and a justification for that estimate.
Health tests - need to stay working to verify the entropy source continues to work after deployed in the field.
Conditioning is to improve the entropy. They are deterministic, so cannot add entropy...
IID = Independent and Identically Distributed - sample indepent of all others, independent of position in the sequence of samples. NIST will run statistical tests to try to disprove claim. If we can't disprove it, we assume it is true.
If you don't claim to be IID, NIST will apply many different entropy estimators against sequential datasets. they will look for things like bias after restart. May get you rejected or a lower estimate of entropy, if issues are found. Would rather underestimate than overestimate!
But... black box statistical tests can't reliably measure entropy.... Ideally you need to design it right and document it and share with NIST (where available).
Currently for conditioning you can choose: hash, HMAC, CMAC, CBC-MAC, DFs all from 800-90A. You can also roll your own. Or, just don't use it.
Problems are: we can't impact performance too much, can't expect this level of expertise at the labs...