#llm

Two production bugs from a legal AI platform, and why the fix for a non-deterministic LLM is engineering the boundary it reads, not writing a better prompt.