Skip to content

Health tests at runtime, what continuous monitoring looks like

SP 800-90B mandates two continuous health tests — repetition count and adaptive proportion — which detect a source that has stopped or stuck. Detecting a source that has merely degraded requires monitoring the physical quantities the entropy model depends on, at a cadence fast enough to act on.

What the standard requires

Two tests run continuously on raw output. The repetition count test alarms when a value repeats more times consecutively than the entropy estimate makes plausible. The adaptive proportion test takes a sample value and counts its occurrences in a window, alarming if one value dominates.

Both are cheap, both run on every sample, and both catch real failures: a source that has died and is outputting a constant, or one that has collapsed to a small set of values. They are a floor and the standard describes them as such.

What they do not catch

A source producing two-thirds of its modelled entropy. That output has no long repetitions and no dominant value; it is simply less unpredictable than the design assumed. Both tests pass, conditioning masks the deficit, and every downstream check is satisfied.

This is the failure mode worth engineering against precisely because nothing complains. Drift is how sources get there, and it is gradual rather than sudden.

Monitoring the model instead of the output

The alternative is to measure the quantities the entropy estimate depends on. For a phase-diffusion source that includes whether the laser genuinely returns below threshold between pulses, whether the interferometer delay still matches the pulse period, and what proportion of the digitised signal is classical detection noise.

These are physical parameters. A departure in any of them invalidates the entropy estimate whether or not the output changes, which is exactly the property the statistical tests lack.

Specifying it in a requirement

  • Which quantities are measured, and at what cadence relative to key generation.
  • What thresholds trigger a warning as opposed to a failure, and what evidence supports those thresholds.
  • What the system does on failure. Refusing to produce output is correct; continuing while logging is not.
  • What is retained, and for how long — telemetry that has aged out cannot answer a question about a key issued last quarter.
  • Whether health state is exposed to operational monitoring, which is where this becomes an operations concern.