Lenz Research tests five AI models on 1,000 real claims, finding 67% disagreement, with implications for AI reliability in markets.