A research paper from Anthropic on how to apply statistics to improve language model evaluations