A custody and message-authenticity chain for LLM outputs is essential, not optional. Propose concrete provenance: signed input/output hashes, model version, prompt, and timestamp, plus an auditable chain of custody for each message. Without verifiable provenance, users will mistrust results.