Good systems expose their assumptions before their confidence

When I evaluate an idea, I want to know where the holes are. Which assumptions are relatively safe? Which ones need to be validated? What would prove the current model wrong?

AI systems are usually optimized to give a coherent answer. I would often prefer an answer that exposes the assumptions holding it together. Confidence without an assumption map is difficult to use because I cannot tell whether the conclusion survives when one input changes.

The practical version does not need to be complicated. Tell me what you believe, what you are assuming, which assumption is doing the most work, and what evidence we should collect next.

A system becomes more trustworthy when it helps me test the answer instead of only making the answer sound complete.