NIST SP 800-90B Explained

Why SP 800-90B exists

Cryptographic random-bit generation needs unpredictability somewhere in the system. A deterministic generator can expand secure state efficiently, but it cannot create new entropy from nothing.

SP 800-90B addresses the part of the architecture where unpredictability enters. NIST describes the publication as specifying design principles and requirements for entropy sources used by random bit generators, along with tests for validating those sources.

The practical consequence is important: the assessment is not only about whether final output looks uniform. It is about whether the underlying source model and measured data justify an entropy claim.

The entropy source is more than a noise signal

A useful engineering abstraction is:

physical/noise process → measurement → raw digital samples → health monitoring → entropy estimation → conditioning → RBG/DRBG architecture → application

SP 800-90B concentrates on the entropy-source part of that chain. For hardware RNGs and QRNGs, the measurement electronics are part of the real source architecture because they can introduce bias, dependence, clipping, drift or other deterministic structure.

Noise-source model

A defensible entropy claim begins with an explanation of where unpredictability is believed to come from and how the digital samples relate to that process.

That model matters because statistics alone cannot tell you whether a fluctuation is quantum, thermal, environmental, deterministic interference, or an artifact of the acquisition chain. Statistical assessment can support a model; it does not replace one.

For QRNG, this is especially important. The word “quantum” describes the claimed origin of unpredictability. SP 800-90B addresses entropy-source assurance. Those are related but distinct claims.

Min-entropy

SP 800-90B uses min-entropy as the conservative quantity for entropy-source assessment. Min-entropy focuses on the probability of the most likely outcome rather than average uncertainty.

That gives engineers a worst-case-oriented accounting method: how much unpredictability can be credited to each sample under the evaluation model?

This is why raw bit rate and entropy rate must not be confused. A source can produce samples quickly while each sample contains substantially less than one bit of conservatively estimated entropy.

IID and non-IID tracks

SP 800-90B distinguishes between data modeled as IID (independent and identically distributed) and data assessed without that assumption.

IID is a strong model. It means samples are treated as having the same distribution and being independent of previous samples. If that assumption is not justified, the safer path is the non-IID assessment.

NIST's reference EntropyAssessment repository states that most commonly used entropy sources are not IID and notes that additional justification is necessary for an IID claim.

The engineering lesson is simple: do not choose IID because it gives a nicer entropy number. Choose it only when the source model and evidence support the assumption.

Sequential data and restart behavior

One long sequence tells you about continuous operation. It does not necessarily reveal what happens when a source initializes repeatedly.

Restart testing is intended to expose structure that may recur across restarts. A source could appear acceptable during steady-state operation while producing similar startup sequences, biased initial behavior or other dependence around initialization.

That is why restart data should be collected deliberately rather than approximated from arbitrary slices of one continuous file.

Health tests

Entropy sources are physical or environmental systems and can change over time. Health tests are runtime mechanisms intended to detect failures or unexpected behavior quickly enough that bad data is not silently treated as valid entropy.

Health testing is not a substitute for entropy estimation. It is an operational control. A good evaluation asks both:

• how much entropy does the source provide under its modeled operating conditions? and
• how does the system detect when the source stops behaving as expected?

Conditioning

Raw samples are often biased or correlated. Conditioning transforms data before it is used downstream.

The important boundary is that deterministic conditioning does not create entropy. It can concentrate or redistribute unpredictability already present in the input, subject to the assumptions and design of the conditioning process.

So an engineer should not infer “full entropy” merely from output that appears uniform.

SP 800-90B is not the whole RNG architecture

The three-part SP 800-90 series separates concerns:

• 90A - deterministic random bit generator mechanisms;
• 90B - entropy sources;
• 90C - complete RBG constructions.

This matters for QRNG because a quantum entropy source can coexist with a DRBG. “QRNG versus DRBG” is often the wrong systems question; they may occupy different layers of the same architecture.

Assessment versus formal validation

Running NIST's EntropyAssessment software can produce useful statistical evidence. It does not, by itself, create a NIST or CMVP validation.

NIST's ESV page states that the Entropy Validation Server runs the SP 800-90B Entropy Assessment Tool on submitted outputs and requires additional documentation addressing the other requirements of SP 800-90B. When a validation certificate is requested, the material is reviewed through the CMVP process.

QRNG.io should therefore use language such as “preliminary SP 800-90B entropy assessment” for its local toolkit, never “NIST certification.”

Questions to ask an entropy-source vendor

1. What physical process provides the entropy?
2. What is the source model?
3. What raw data is available before conditioning?
4. Is the entropy claim IID or non-IID, and why?
5. What min-entropy is credited per raw sample?
6. How was restart behavior evaluated?
7. Which health tests run during operation?
8. What happens after a health-test failure?
9. What conditioning is used?
10. Is any claimed validation tied to a specific implementation and certificate?

Practical next step

If you have raw entropy data, use the QRNG.io Local Entropy Assessment Kit for a preliminary local assessment. Your dataset should remain on your own computer. For formal compliance or validation decisions, use the applicable standards and accredited-validation process.

Sources and further reading

1. NIST SP 800-90B
2. NIST Random Bit Generation project
3. NIST Entropy Validation Server
4. NIST EntropyAssessment repository