Would you run a factory
on parts you never inspect?
Every cryptographic key in your systems was manufactured. It has a raw material, a machine that shaped it, and a floor it was built on. The raw material is entropy, and almost nobody checks it.
Incoming inspection is the step nobody skips.
Walk into any serious factory in any serious industry and the same station is at the door. Nothing goes into production until the raw material has been checked.
Pharmaceuticals
Compounds tested for purity, potency and contamination before a batch starts.
InspectedAerospace
Alloys and composites traced to their lot and certified before assembly.
InspectedFood and drink
Incoming ingredients checked for freshness, safety and microbiology at each stage.
InspectedSemiconductors
Wafers screened for defects, contamination and crystal integrity before the fab touches them.
InspectedCryptography
Entropy is consumed as it arrives. A datasheet at design time, then years of assumption.
No inspection
If the raw material is compromised, everything downstream is compromised. In any other industry that logic stops the line. In cryptography there is usually nothing there to stop it.
This is the root cause behind a long tail of real failures — the class of weakness catalogued as CWE-330, use of insufficiently random values. Keys that could be predicted, sessions that could be replayed, certificates that could be forged. In almost every case the algorithm was sound. The material going into it was not.
“We already test our randomness.”
You probably do, and those tests are worth running. They are not an inspection of the raw material. Statistical tests examine the output and ask whether it looks random, which is a different question from whether the source is still producing randomness.
Here is the same entropy source in two states. Use the switch to break the physical source while leaving the conditioning downstream in place — which is what a slow hardware failure, or a source being worked on deliberately, looks like from the outside.
Statistical testing of the output
Examines — The bits coming out
- Monobit frequency
- PASS
- Runs
- PASS
- Serial correlation
- PASS
- Compression ratio
- PASS
Output indistinguishable from random
Measuring the source against its model
Examines — The physical process making them
- Measured behaviour of the source
- Consistent with the model
- Operating conditions
- Inside the validated range
Source inside its validated operating conditions
The model is what the source was validated against. Measuring against it is the difference between an assertion and an observation.
Both methods agree while the source is healthy. Break the source and they part company.
A well-seeded deterministic stream passes every statistical test ever written, with no entropy entering it at all. That is not a flaw in the tests — it is what they are for. They check the goods. Somebody still has to check the goods coming in.
Locally Verified Entropy
Quantum physics is how the entropy is made. Local and verified is what a buyer actually gets, and both are properties a source either has or does not.
The source sits inside your device, inside your trust boundary. It is yours the way a generator on site is yours.
- No network between the material and the machine that consumes it.
- Availability and latency you control, rather than somebody else’s uptime.
- Sovereignty and audit boundaries a procurement team can draw on a diagram.
- Randomness that arrives over a network cannot be verified where it is used.
The physical process has a model, and the parameters of that model are measured while the system runs.
- Health monitored at runtime, not certified once at the factory gate.
- Drift, ageing, temperature and interference show up as measurements instead of silence.
- The record is produced at the device and made available to the host system.
- Entropy quantified within a defined model, which is what a physical source can honestly offer.
The full account of what the source measures, what a health state means and where that evidence goes is in the Assurance area: Locally Verified Entropy and runtime observability.
What the claim covers, and what it does not.
Assurance language is easy to overinflate and expensive to walk back. Here is the line we hold, and the line we ask everyone to hold with us.
It does mean
- The physical process is characterised by a model, and the parameters of that model are measured while the device runs.
- Entropy is quantified within that defined model, and the device reports whether it is still inside its validated operating conditions.
- Anomalies surface as evidence you can log, alarm on, and hand to an auditor.
It does not mean
- That every output bit is mathematically perfect. Nobody can claim that, and a buyer who knows the field will mark you down for trying.
- That deterministic generators are broken. A sound design pairs a validated entropy source with an approved deterministic mechanism, which is what the SP 800-90 series describes.
- That quantum origin is enough on its own. A quantum process with no measurable model is a different black box.
Regulators started asking the same question.
Entropy requirements sat still for a long time. Then several things landed close together, and they all point the same way: show the evidence.
Read together, these ask for better verification, a clearer account of entropy quality, and in some cases more performance. That is a fairly precise description of what a locally verified entropy source does. The standards themselves, clause by clause, are mapped on standards mapping.
Where is your entropy generated, how is its quality verified, and who can observe that verification?
Ask a supplier those three questions in that order. Most will have an answer to the first, a datasheet for the second, and nothing at all for the third. That gap is what this stage is about.
Good material is where a factory starts. The next question is where the machines sit.