LLMs do great when complemented with structured information that is ground-truthed.

As they accumulate their own information, they can confuse themselves, and drift further and further from the ground-truth.

A process to accrete only ground-truthed observations helps unlock the power of models.