x
Improving Truth-Seeking in LLMs: An Internal Verified Knowledge Base with Ratcheted Multi-Track Confidence — LessWrong