The train-test split does not work for classification tasks at the frontier of LLM capability.