The finish reason is `length`. I have tried setting minimal token budgets, really small prompts, and max lengths of various sizes from 100-4000 and nothing seems to make a consistent dent in the behavioral pattern.