You can have a separate LLM as a brutal enforcer of an architecture decision. Works pretty well in the right harness.