I built a lightweight hallucination detection system for LLM outputs that flags unreliable responses (block / caution / safe). It is designed to integrate with observability stacks like Langfuse to help surface unreliable model behavior in production.

Here is the repo: https://github.com/wpydesign/safegate