AI use case

Decide which small local model to trust by having Claude check the sources each one cites

A research project had five small local models (Llama 3.2 3B, Mistral 7B, Gemma 2 2B, Phi-3 mini and Qwen 2.5 3B) answer questions with citations through Ollama, and used Claude Sonnet 5.5 only to judge whether each cited passage backed the answer. On 1,000 questions their rule lifted accuracy from 77.1% with a simple majority vote to 85.0%, though Claude alone picked the right answer more often on the questions where the models split.