Two gaps in every hosted AI call: which model ran, and who saw the prompt. What I run to close both, and where each one still falls short.