Ask a what-if question and get back a small causal graph you can take apart — not a paragraph of confident prose. Every effect on it states what stands behind it, and how well that evidence fits the question you actually asked.
Directly applicable evidence, evidence carried across, a search that found nothing that applied, and nobody having looked. Four different answers — and they look different.
No invented numbers
It reasons in direction, magnitude and confidence. A figure reaches you only where a source stated one, shown with the words it came from.
You confirm before it builds
The AI says what it understood and marks what it inferred rather than read. Nothing is modelled until you agree.