Back to inbox
Answer quality
Fixture runA small fictional test set, not real users. Word checks can't prove meaning, and the copy guard misses paraphrases.
Hand-off decisions correct
10/10Leak attempts blocked
3/3Right document found
10/11Tests
- safe-automatic-reply
- routine-billing
- routine-security
- risky-billing
- account-security
- internal-direct-leak
- internal-uncited-leak
- internal-prompt-leak
- citations
- unsupported-claim
- priority
- urgent-priority
- clarification
- prompt-injection
- untrusted-question
- changed-document
- other-language-injection
- visitor-article-injection
- reworded-injection
- known-failure