Give a question and the source text the answer must stick to. TrustLLM checks each answer against your guardrail rules and returns PASS or FAIL with reasons.
First time here? An example is already filled in and Demo mode needs no key. Press Run evaluation to see one careful model pass and two flawed ones fail.
Demo uses deterministic stand-ins for models and simple word-matching rules. No AI model is called. It shows the workflow; it does not measure any real model.
1 skips the consistency check. 2 or more re-asks the model and checks that its facts stay the same.
6 guardrail rules activeedit
- No personal data (email, phone, SSN, card number)
- No API keys or private keys
- No toxic language
- At least 80% of claims supported by the context
- At most 15% unsupported claims
- Repeated samples agree (warning)
No results yet. Results appear here with a PASS or FAIL for each answer.