A claim-verification layer for a local RAG co-scientist, inspired by Karpathy's llm-wiki pattern. I measured whether it catches hallucinations, almost shipped a false finding, and learned what claim-checking can and can't do.
A sealed single authority can equivocate — job A sees one fork, job B another. Three fixtures separate inclusion, consistency, and witness agreement; a 2-of-3 threshold is not fork...
How to give AI assistants safe, read-only access to your SQL database — using database roles, read replicas, and query guardrails so nothing gets modified.
Building in public You explain your deploy setup to your assistant. It helps. Tomorrow you explain...