Make Hallucination Visible
Don't promise zero hallucination. Tag every claim as grounded or [UNGROUNDED]. Make failure visible. Match how senior humans work — cite when certain, flag when guessing.
"Zero hallucination" is aspirational, not achievable. The honest reframe: every claim must be grounded against a source or explicitly tagged as ungrounded. The system's job is not to prevent hallucination — it's to make hallucination visible and traceable so the human reviewer catches it.
Practically: every assertion in generated output carries a source citation or an [UNGROUNDED] marker. The reviewer's eye learns to scan for those markers. The model's failure modes become surface-level instead of buried.
This is a better contract with the user than "trust the AI." It also matches how senior humans actually work — they cite where they're certain and flag where they're guessing.