AI Evaluation Engine

Security, safety and answer-quality assessment for any AI product. Produces a Word assurance report.

1About the product

Tell the engine what the product is. Your answers decide which tests apply and how answers are judged.

One line. Used to test that it refuses everything else.
Prefer staging: the tests send jailbreaks and abusive text on purpose.

2How to reach it

The API the engine sends test messages to. Keys are used for this run only and are never saved.

Presets:
Use {{message}} where the user's text goes and {{session_id}} for a conversation id. Any other fields are sent as written.
Leave empty to auto-detect common shapes.
Raise it if the product has a rate limit.

3What a correct answer looks like

Security tests show whether it can be attacked; these show whether its answers are right, helpful and safe. Use any or all.

The judge fails answers that contradict this text or invent details that are not in it.
The judge writes questions from your description and grades each answer for correctness, helpfulness and safety. Needs a judge model (step 5).
Confidential details to protect (recommended)
If any appears in a reply, the prompt has leaked. Pick phrases a normal answer would never contain.

4Tests

Selected from your answers in step 1. Struck-through tests don't apply to this product.

5Judge model

An AI model that decides what fixed rules can't: did it really refuse? Is the answer correct? Without it those tests come back "Inconclusive" for a person to review. Use a different model from the one being tested.

6Report details

Previous runs

None yet.