Skip to content

Level 3 vs Level 4, what observability adds

Level 3 means a physical entropy source was characterised during evaluation. Level 4 means the quantities its entropy model depends on are also measured while the system runs. The difference is whether your evidence describes the device on the day it was tested or the device that is generating your keys now.

The gap Level 4 addresses

A Level 3 evaluation is real evidence and should not be dismissed. A laboratory analysed the source, derived an entropy estimate from a stated model, and tested sample devices against it. That is far more than most systems can show.

What it cannot tell you is whether the unit in your server is still behaving that way after three years at elevated temperature, or whether a component has drifted, or whether this particular device left the line at the edge of its process window. The certificate describes a population under evaluation conditions. Your risk is about one device under yours.

Why statistical health tests do not close it

The obvious objection is that SP 800-90B already mandates continuous health tests. It does, and they are worth having, but they test the output rather than the source. Repetition count and adaptive proportion catch a source that has stopped or stuck. They do not catch a source that is producing two-thirds of its modelled entropy, because that output is statistically unremarkable.

A partially degraded entropy source is the dangerous case precisely because nothing downstream looks wrong. Conditioning takes under-entropic input and produces output that passes every suite you run at it.

What observing the source means concretely

It means identifying which measurable quantities the entropy model depends on and monitoring those. For a phase-diffusion source that includes whether the laser is genuinely returning below threshold between pulses, whether the interferometer delay still matches the pulse period, and how much of the digitised signal is classical detection noise.

Those are physical parameters, not output statistics, and a departure in any of them invalidates the entropy estimate whether or not the output looks different. Measuring them is what lets a system report that its entropy assumption still holds instead of assuming it does.

Why this is a commercial argument, not just a technical one

Auditors increasingly ask not whether a source was certified but what evidence exists that it is working now. What auditors actually ask about an entropy source follows that thread into the assurance pillar, which is where this distinction turns into a requirement someone writes down.