I’m sure someone is working on adding support for human-like errors in LLM outputs.