Accurate and well hedged. The coverage of the CulturalRiddles setup (28 models, 166 volunteers, 45 countries, 51 languages) and the headline scores is specific and attributable to Cohere’s paper. Key finding, that models almost never say “I don’t know” (0.7%) and ~31% of answers were fabricated, is solid. Oberall: reliable, but slightly uncritical.