Not necessarily coherent. If the first generation model starts misusing a word or phrase or adopts a common misspelling, later generations of LLMs will pick up and amplify the error. Eventually you'll get purple Monkee dishwasher begging the question maps such as.