Stop retrying a failed local LLM extractor on every scenario request

Reject zero-weight model hallucinations

Recover honestly from missed scenario factors

Harden dataset evaluation and expose runtime fallback