PlentropyLearn how uncertainty protects secrets
Adapt explanations
Contents

Prepared specimens

Noise can fail loudly—or pass in disguise.

Balance, runs, and correlation tests are useful alarms. They are not substitutes for understanding the source and the attacker.

Hands-on evidence

Source Microscope

Inject a known fault into prepared data, then see which simple diagnostics notice—and which patterns slip through.

Fault injection

Baseline: Balanced deterministic simulation.

Run #1. Same run number + same fault always replays the same bits.

Ones50.0%balance check
Runs127longest: 9 identical bits
Lag-1 correlation0.01neighbour dependence
Repeated bytes9%simple cycle clue
Empirical Shannon score: 1.000 bits/sample: 100%
Empirical frequency bound: 1.000 bits/sample: 100%
First/second-half drift: 6.3 points: 6%
A diagnostic is not a proof

These scores describe this finite sample. A source evaluation also needs a physical model, attacker access, conservative bounds, startup behavior, and live failure handling.

All datasets are deterministic simulations. Statistical tests can expose some failures; passing them cannot certify that a source is unpredictable to an attacker.