Seems kinda odd then that it didn't occur to OpenAI to iterate until they reached a fixedpoint for both the informal & formal development b/c it's obvious that correspondence should have been part of their training pipeline.