CausEval
AI TESTING, MADE CLEAR

Do your tests catch
missing AI rules?

CausEval removes one instruction at a time and checks whether your tests notice. See what’s covered, find the gaps, and know what to test next.

Explore the example No setup. No API key. Saved example results.How does it work?
EXAMPLE REPORTSupport assistant

7 rules need a closer look.

Here’s what the tests detected when instructions were removed.

5Removal detected
4Removal missed
3No test linked
ONE GAP TO EXPLORENever send an email without explicit user confirmation.The instruction was removed. The test still passed.See what happened

Precomputed example, not a live model assessment.

THE IDEA

A passing test is only the start.

If an instruction disappears, would your tests catch it? CausEval helps you check.

1

Find the rules

Start with the instructions you give your AI, such as ‘verify identity before sharing account details.’

2

Check the tests

Link each rule to the tests meant to check it. Remove that rule and repeat those tests.

3

Review the gaps

See which removals were detected, which were missed, and which rules have no linked test.

WHEN YOU’RE READY

Run it on your own prompt.

The example above needs nothing installed. To check your own AI, point CausEval at your prompt and your existing tests.

Apache-2.0 No telemetry Works with existing tests
Setup guide Star the repository
Terminal · Node 22+
$ git clone https://github.com/aravinda-1402/causeval.git$ cd causeval$ pnpm install --frozen-lockfile$ pnpm build$ pnpm causeval demo