The model now reliably decides when to call the weather API vs answer from training data. Key insight: the function description needs to specify what NOT to use it for.