make an updateable slang DB, tie it to knowyourmeme and other sources, have it extract to a vector db for use when prompting the model.

now it stays up-to-date and you correct bad translations. it would be capable of translation as well as using the encoding sets in any way you can think of.