An auditor asks which entropy source feeds your module, and the vendor's answer is "it's FIPS 140-3." That is not an answer, and the gap between those two statements is where most of the risk sits.
FIPS 140-3 validation requires that any entropy source inside a validated cryptographic module conform to NIST SP 800-90B: a documented noise source, health tests, and a min-entropy estimate backed by data. Since November 2020, NIST has required that evidence with each module submission where it applies, and Entropy Source Validation (ESV) can validate an entropy source on its own.
If you were told to "ask vendors for SP 800-90B evidence" and were not sure what good evidence looks like, this post walks through it, from what the standard asks for to how to read a vendor's packet. It also follows up on our CNSA 2.0 post, which ended with that same instruction.
What Does FIPS 140-3 Require for Entropy Sources?
FIPS 140-3 (opens in new tab), Security Requirements for Cryptographic Modules, covers the whole module: its specification and interfaces, roles and services, software and firmware security, physical security, sensitive security parameter management, self-tests, and life-cycle assurance, among other areas. The standard itself does not spell out entropy rules. They come from the validation program that enforces it.
Here is the rule that matters. As of November 7, 2020, the Cryptographic Module Validation Program (CMVP) requires every FIPS 140-2 and FIPS 140-3 submission to include documentation justifying conformance to SP 800-90B (opens in new tab) where it applies. SP 800-90B, together with FIPS 140-3 Implementation Guidance documents D.J, D.K, and D.O, sets out what an entropy source needs to be included in a FIPS-approved module. With the publication of SP 800-90C, CMVP also accepts submissions for complete random bit generators under IG D.T.
Is the Entropy Source Part of the Module, or Validated Separately?
Either. An entropy source can be validated as part of a module submission, or on its own through ESV, which produces a standalone entropy validation certificate. A module vendor then cites that certificate. NIST also publishes an ENT to ESV conversion process, so an entropy source first validated inside one module can be converted to a standalone certificate and reused in other module submissions.
Does a FIPS 140-3 Certificate Mean the Randomness Is Good?
It means a lab reviewed the entropy documentation and CMVP accepted the justification. That is meaningful, and it is also narrower than the marketing version. A certificate covers a specific module, version, and configuration. Look up the listing, read what it covers, and check whether the entropy source sits inside the validated boundary or is cited from a separate certificate.
What Is the Difference Between FIPS 140-2 and FIPS 140-3?
For most of the module requirements, NIST says there are "few major technical requirement changes." The big shifts are procedural, and the entropy treatment is one of the places where the two standards really diverge. Here is the comparison, sourced from NIST's transition schedule (opens in new tab) and entropy validation documents (opens in new tab):
| Topic | FIPS 140-2 | FIPS 140-3 |
|---|---|---|
| Status | Superseded. All certificates moved to the Historical List on September 22, 2026. | Current standard. Approved March 22, 2019, effective September 22, 2019. |
| Source of module requirements | NIST's own standard text | References ISO/IEC 19790:2012, tested per ISO/IEC 24759:2017, with NIST SP 800-140x documents modifying requirements |
| New validation submissions | CMVP stopped accepting new certificate submissions on April 1, 2022 | Accepted since September 22, 2020 |
| SP 800-90B "caveat allowed" requirements | A submission that misses one gets a caveat on the certificate saying the entropy source does not conform to SP 800-90B | Treated as required |
| SP 800-90B documentation in submissions | Required since November 7, 2020, where it applies | Required since November 7, 2020, where it applies. A standalone ESV certificate can be cited by a module submission |
What Does the Historical List Mean for Buyers?
NIST says that even on the Historical List, CMVP supports the purchase and use of those modules for existing systems. So a FIPS 140-2 certificate is not suddenly worthless. For a new purchase, though, we would ask for a FIPS 140-3 certificate, because new 140-2 submissions closed in 2022 and the stricter entropy treatment applies under 140-3.
Why Does the Caveat Row Matter Most?
Because it is the entropy difference you can see on a certificate. Under FIPS 140-2, a module could be validated with a listed caveat stating that its entropy source did not conform to SP 800-90B. Under FIPS 140-3, the same requirements are treated as required. If you are evaluating an older module, read its listing for that caveat.
What Does SP 800-90B Actually Require?
NIST SP 800-90B (opens in new tab), published in January 2018, specifies the design principles and requirements for entropy sources used in random bit generators, plus the tests used to validate them. An entropy source feeds a deterministic random bit generator (DRBG) from SP 800-90A, and SP 800-90C describes how the two are assembled into a complete generator.
The document is long, but the requirements group into six pieces. Section numbers come from the published PDF (opens in new tab).
| Piece | What 800-90B asks for | Where |
|---|---|---|
| Noise source | A documented process that produces raw, unpredictable samples | Section 3.2.2 |
| Health tests | The Repetition Count Test and the Adaptive Proportion Test, run at startup and continuously. Startup tests run the continuous tests over at least 1,024 consecutive samples | Sections 4.2 to 4.4 |
| Restart tests | Tests that check behavior when the source is restarted | Section 3.1.4 |
| Min-entropy estimate | An entropy assessment from collected samples. IID data uses the most common value estimate. Non-IID data runs ten estimators and takes the minimum | Section 6 |
| Conditioning | An optional step that cleans up the output. Vetted options include HMAC, CMAC, CBC-MAC, hash functions, Hash_df, and Block_Cipher_df | Section 3.1.5 |
| Documentation | Written requirements covering the entropy source, the noise source, the conditioning component, and data collection | Sections 3.2.1 to 3.2.4 |
What Is Min-Entropy, and Why Does It Matter?
Min-entropy measures how predictable the single most likely output is. It is a deliberately cautious number. A source that gives one outcome a slightly higher chance than the rest scores lower than its average randomness would suggest, which is the right bias when the question is how hard a key is to guess. "IID" means the samples are independent and identically distributed. If a vendor claims IID, they need data to support it. If they cannot, the more conservative non-IID path applies.
Do NIST STS, DieHarder, or ENT Results Satisfy SP 800-90B?
No, and this is the most common confusion. Those suites look for patterns in output. SP 800-90B asks for health tests that fit the noise source, an entropy estimate from raw samples, and written documentation. The statistical testing it requires is defined in its own sections and implemented in NIST's Entropy Assessment Tool (opens in new tab). Passing a general-purpose suite is good supporting evidence. It is not the validation.
How Does Entropy Source Validation (ESV) Work, Step by Step?
Entropy Source Validation (opens in new tab) is the CMVP program that validates an entropy source on its own, separate from any one cryptographic module. A vendor works with an accredited lab, the evidence goes through NIST's ESV server, and a successful result is a standalone entropy validation certificate. In practice it runs like this:
- Pick an accredited lab. The lab must be accredited for ESV testing (the 17ESV scope). The vendor and lab agree on what is being validated.
- Collect raw noise source data. The vendor gathers samples from the noise source under the conditions the documentation describes. Quality here decides everything that follows.
- Write the entropy assessment report. NIST publishes an Entropy Assessment Report template (opens in new tab) that follows the SP 800-90B documentation requirements: the noise source, the health tests, the conditioning component, and the data collection method.
- Submit through the ESV server. The lab uses the ESV web API, built on the same protocol family as NIST's algorithm validation service. The server runs the SP 800-90B Entropy Assessment Tool against the submitted noise source data. The additional documentation is submitted through the same server.
- CMVP review. A CMVP reviewer decides whether the evidence is sufficient. The tool output is an input to that decision, not the decision itself.
- Certificate issued. A successful validation produces an entropy validation certificate. NIST also publishes a Public Use Document describing how the certificate can be referenced.
How Does an Entropy Certificate Connect to a FIPS 140-3 Module?
A module that uses a validated entropy source can cite the certificate in its own CMVP submission instead of re-proving the entropy from scratch. NIST has published guidelines for module submissions that include an ESV (opens in new tab) and a path for converting older ENT-based entropy documentation to the ESV format. The practical result for buyers is simple: one entropy certificate can be traced across more than one product.
What Should a Buyer Ask For?
Two things. The certificate number, so the claim can be checked against NIST's entropy validations list (opens in new tab), and the scope. A certificate covers a specific entropy source in specific conditions. If your deployment looks different, ask how the vendor knows the claim still holds.
How Do You Read a Vendor's Entropy Evidence Packet?
A good packet answers one question: where does the unpredictability come from, and how do you know it is still there on a bad day? Here is what to look for, and what should make you ask more questions.
| Look for | Why it matters | Red flag |
|---|---|---|
| A named, physical noise source with a written explanation of why it is unpredictable | The whole security argument starts here | "Proprietary" with no description, or a source that is really a software clock or system timing |
| Raw samples collected before conditioning | 800-90B estimates entropy from the noise source, not from cleaned-up output | Only post-hash or post-DRBG data was tested |
| A min-entropy estimate with the method named (IID or non-IID, and which estimator was the minimum) | Shows the claim was derived, not asserted | A single entropy number with no method |
| Health tests matched to the source, with cutoffs and the reasoning behind them | Catches a source that degrades or fails after deployment | Generic tests copied over with no tuning |
| Startup and restart behavior | Many real failures happen at boot, when entropy is scarcest | Documentation only covers steady state |
| A named conditioning component from the vetted list, or a justified alternative | Conditioning is where good raw entropy can be wasted or overstated | Custom conditioning with no analysis |
| A certificate number you can look up | Third-party review is the point | A logo or a "compliant with" statement |
Why Do Boot-Time Failures Matter So Much?
Because they have already happened at scale. A 2012 USENIX Security study (opens in new tab) scanned the internet and found that about 0.75% of TLS certificates shared keys with other hosts, and recovered RSA private keys for about 0.50% of TLS hosts. The researchers traced much of it to devices generating keys at boot, before they had gathered enough entropy. Startup tests and restart tests exist to make that failure visible instead of silent. Our post on how weak entropy led to the LuBian bitcoin theft shows what the same class of mistake looks like in a more recent incident.
Is a Clean Statistical Test Report Enough?
It helps, and it is worth asking for. NIST STS, DieHarder, and ENT results show that output does not have obvious patterns. They cannot tell you how much entropy went in, because a deterministic algorithm with a fixed seed can pass them all. That is the core difference between random and pseudorandom number generators, and it is why SP 800-90B exists. Real Random publishes its own test results summary (opens in new tab) and we treat it as supporting evidence, not as a substitute for entropy validation. You can see how we frame the difference on our entropy scorecard.
What Is the Checklist for Teams Preparing a CMVP Submission?
Whether you are a vendor building the evidence or a buyer reviewing it, work through this list before the paperwork starts.
- Identify every noise source in the design and document each one in plain language.
- Decide whether the entropy source will be validated on its own through ESV or only as part of the module.
- Choose and contract an ESV-accredited lab early, before data collection.
- Collect raw noise source samples under realistic operating conditions, including temperature and power variation.
- Determine whether the data supports the IID assumption. If it does not, plan on the non-IID path.
- Run the SP 800-90B Entropy Assessment Tool and record which estimator gives the minimum.
- Implement the Repetition Count Test and Adaptive Proportion Test, with cutoffs tied to the claimed entropy.
- Verify startup behavior (at least 1,024 consecutive samples) and restart behavior.
- Use a vetted conditioning component, or document the analysis for a non-vetted one.
- Map each 800-90B shall statement, including the "caveat allowed" ones, to evidence.
- Use NIST's Entropy Assessment Report template for the write-up.
- Confirm how the DRBG is seeded and which SP 800-90C construction applies.
- Record the certificate number and scope so downstream products can cite them correctly.
Why Does This Matter Beyond a Certificate?
Entropy is the quiet dependency under every key a system makes. The post-quantum algorithms now rolling out, covered in our guide to what post-quantum cryptography is, still need unpredictable seeds, and a strong algorithm fed a weak seed is still a weak key. That is why true random number generation matters in the post-quantum era, and why federal buyers are asking for entropy evidence now. If you are weighing where entropy should come from, our overview of entropy as a service and our hardware random number generator primer are good next reads. For how a physical entropy source fits into FIPS planning, see our FIPS 'n' Chips talk on physical entropy and our entropy page.
Summary
- What changed: FIPS 140-3 took effect September 22, 2019 and points to ISO/IEC 19790:2012. NIST says there are few major technical changes, but the process changed a lot. All remaining FIPS 140-2 certificates moved to the Historical List on September 22, 2026.
- What the entropy rule is: SP 800-90B sets the design principles for entropy sources and the tests used to validate them.
- What ESV is: a NIST process where an accredited lab submits entropy data and documentation through a server, and NIST's CMVP reviews it and issues an entropy validation certificate.
- What good evidence looks like: an entropy assessment report, health tests that match the noise source, a min-entropy estimate from raw samples, and a public use document that tells you how to integrate the source.
- What not to accept: a pass on a statistical test suite as a substitute for any of the above.

