Beyond Karpathy's LLM-Wiki: The Necessity of Cognitive Governance

Author here. This started as my own frustration: I built a compiler along the lines Karpathy describes in his llm-wiki gist, fed it ~300 files of reading notes accumulated over twenty years, and got back something that looked like an encyclopedia written by a competent stranger. Accurate, well-formatted, philosophically sterile. The piece argues that Karpathy's compiler frame solves the infrastructure half of the problem (ahead-of-time compilation beats just-in-time RAG) but leaves the architectural half open. Without an explicit governance layer, the compiler defaults to its training distribution and smooths everything into consensus. With one, it can be forced to do the opposite: extract structural claims rather than summarize, name antagonists, find friction points across your existing nodes. I shared the schema and directory structure I use as an open-source template (linked at the end of the post). Curious whether others who've built something similar found different rules that matter, or whether the antagonist/friction directives feel too philosophical for what should be a pure engineering problem. The qaadika comment on the original llm-wiki thread is the strongest objection I tried to address.