PlentropyLearn how uncertainty protects secrets
Adapt explanations
Contents

Look beyond the histogram

Where did the uncertainty actually come from?

Data does not carry a certificate saying “random.” A defensible claim explains the source, every transformation, what an attacker knows and how the system fails.

Open the Source Microscope →

The complete trust boundary

Analog and digital are one connected chain.

A physical event becomes digital through a sensor and sampler. Firmware can alter it. Conditioning can hide failures. Every stage belongs in the evaluation.

01SourceA physical or digital event
02SampleTurn events into raw observations
03WatchCatch stalls, bias and drift
04ConditionConcentrate — never invent — entropy
05PoolCombine and remember trusted state
06ExpandA secure generator serves applications

Ten questions before a claim

An evidence profile, not a “randomness score.”

No single test answers whether a source is secure. Work from the intended use and attacker toward a conservative lower bound.

  1. 01

    Purpose

    What must be unpredictable, for whom, and at what rate?

  2. 02

    Signal path

    Which physical event, sensor, converter, firmware and software touch the sample?

  3. 03

    Raw access

    Can evaluation happen before hashing, compression or other conditioning?

  4. 04

    Attacker

    What can an observer learn, predict, control, inject or replay?

  5. 05

    Dependencies

    Are samples biased, correlated, drifting or shared with another source?

  6. 06

    Lower bound

    What conservative min-entropy per sample can the evidence support?

  7. 07

    Health

    Which startup and continuous failures are detected—and what happens next?

  8. 08

    Conditioning

    How is existing entropy concentrated without claiming new entropy?

  9. 09

    Operations

    How do temperature, voltage, load, aging, reboot and virtualization change behavior?

  10. 10

    Validation

    Can independent testing and documentation support the intended use?

A critical limit

Passing tests does not prove unpredictability.

A deterministic generator can create data that passes a large battery of statistical tests. Tests can reveal bias, runs, correlation and obvious failure; they cannot replace a source model or cryptanalysis.

NIST SP 800-22 states this boundary explicitly ↗

Sources around us

Complex does not mean secret.

These are starting judgments, not universal grades. Change the hardware, environment, attacker or purpose and the evidence changes too.

System●●●●

OS secure generator

Use the platform interface

Usually the right application source. It pools system events, hardware help and persistent state behind a reviewed interface.

Human●○○○

Keyboard mashing

Useful lesson, poor key source

Hands repeat, favor nearby keys and settle into rhythms. A short sample cannot justify a cryptographic entropy claim.

Physical●●●○

Fair dice

Auditable at human scale

Simple to inspect and explain, but only if dice, rolling method, recording and environment are trustworthy.

Timing●○○○

Clock or timestamp

Usually predictable

High precision creates many digits, not necessarily many attacker-unknown possibilities.

Electronic●●●○

Oscillator jitter

Can be engineered well

Requires a source model, careful sampling, environmental testing, health tests and conditioning.

Public●○○○

Weather or market data

Observable, therefore not secret

Complex and hard to forecast perfectly, but shared public observations do not make private key material.

Optical●●●○

Quantum / shot noise

Promising source, still a system

Physics can support a strong model; detectors, side information, calibration and extraction still need evidence.

Processor●●●○

CPU RNG instruction

A component, not a ritual

Useful when correctly integrated. Prefer the OS generator and understand availability, conditioning and trust boundaries.

Trap 01

Whitening hides bias. It does not create choices.

Hashing or conditioning can produce balanced-looking output and concentrate entropy already present. If the input had only one unknown bit, the output cannot honestly be credited with hundreds.

Trap 02

Two sources do not always add.

If both sources depend on the same clock, environment or attacker-controlled event, counting them separately can double-count the same uncertainty.

Your turn at the instrument

Inject a fault. See what survives.

Bias a simulated source, make samples correlate, freeze the sensor or add drift. Then compare the histogram, runs and lag correlation without mistaking them for proof.

Run the Source Microscope →