Principle → Probe → Judge → Score

Fair. Grounded. Safe.
Now measured.

Everyone agrees AI should behave. Almost nobody agrees how to check. PROBE turns one principle into a defensible score — with the evidence behind it.

Who's asking? Pick your world for tuned examples (optional)

▶ Start the 60-second walkthroughSkip to try it

No sign-up · saved runs are labeled · the judge has limits, and we say where

The quick version

What it is

An engine that scores how well an AI follows a principle, 0–100.

Why it matters

“Be fair” becomes an auditable number you can defend.

Try it

Ask your own question below and see it scored, with the evidence.

For first-timers · ~60 seconds

Watch one evaluation, start to finish

1

The principle

Pick one, like “stays grounded in facts.”

2

The probe

PROBE writes pointed test messages for it.

3

The judge

An impartial second model grades each answer.

4

The score

Scores roll up to one number, with evidence.

principle → probe → chatbot → judge → score → report

For everyone

Check a model yourself

Four ways, from watching to doing: 1) see a finished measurement on a sample bot, 2) paste your own model's answers, 3) connect your model's API and let PROBE run it, or 4) run an open model in your browser. Pick a principle, pick a tab.

Principle: · · ·

A saved run of Gender fairness against a deliberately flawed sample chatbot — so you can see what a finished measurement looks like. To test your own model, use tab 2.

Want this in your own code or CI instead? PROBE is a Python engine — the same principle → probe → judge → score chain runs headless against any model you wrap.

For engineers · optional depth

How it works

The chain, part by part — each stage tested on its own

A non-deterministic model sits inside a fully deterministic test suite. The target model and the judge both sit behind interfaces with fakes, so the core test run makes zero live calls and stays repeatable.

Why an impartial second model does the grading

Each answer is scored against a plain rubric by a separate judge model, not by the model under test. The judge has limits — it can be noisy on edge cases — so scores come with the worst cases attached rather than as a bare number.

Extending it to classical ML (designed, not yet shipped)

The Evaluator is an interface. The LLM evaluator ships today; a classical-ML evaluator (e.g. a credit model) is designed to plug into the same contract, which is the seam for the “how would you extend it” answer.

Measuring the model is half the picture

PROBE checks how the AI behaves. Whether your governance holds is a different question — check it with the Accountability Split (NIST AI RMF).

Open the accountability check →

Who this is for

Three ways in

Hiring?

This is a working artifact of how I build — architecture, tests, and honest framing.

An engineer?

Open “How it works,” then wire the score into CI as a regression gate.

In governance?

Run a principle, read the evidence, then follow the accountability link.