Evaluating generative artificial intelligence models that produce full-text radiology reports is challenging because they can represent hundreds of potential findings in an open-ended output.