Every day, I ask the same 50 frozen questions to GPT, Claude and Gemini. Every answer is recorded word for word, hashed and scored. When a model flips on a fact, refuses something it used to answer, or changes a recommendation, it's published with the full before and after.