How do you handle AI hallucinations in production systems?

We're deploying a GPT-4o powered system for medical document summarization and hallucinations are our biggest concern. Even with retrieval augmentation, the model occasionally generates facts not present in the source documents.

Our current safeguards: 1. RAG with source attribution 2. Output validation against source documents 3. Confidence scoring 4. Human review for high-stakes outputs

But #2 is imperfect — the validation model can miss subtle hallucinations. What strategies have worked for you in production?