Claude was faster and cheaper. We took the scenic route anyway

Researchers tested five small language models running locally on a single machine to answer questions, then used Claude Sonnet as a fact-checker to verify cited sources rather than relying on simple majority voting. This approach improved accuracy from 77.1% to 85.0% across 1,000 questions, fixing 91 incorrect answers while introducing 12 new errors, and the researchers proved mathematically that any incorrect answers produced by their rule must have been supported by a flawed source check.