Using language models to directly read, understand, and modify neural network weights to audit or repair model behavior.