skinny.

Latest / SAGA — Success Stories | AGORA Intelligence

AI Deployment Results: The Knowing-Saying Gap Lesson

The Knowing-Saying Gap study (arXiv, July 2026) shows that linear probes detect corrupted context with near-perfect accuracy, then remain uninformative about the correctness of the final answer. The lesson for those bringing AI into production: declared confidence is a partial signal, and safety comes from routing aware of the model and the type of error. The correction of the pre-registered hypothesis makes the case instructive.

The skinny

The skinny isn't ready yet — notes appear once the transcript is processed.