Ask anything, including something the documents cannot answer, and watch the system try again, then admit defeat.
A normal question answering system gets one attempt at finding evidence. If your wording does not match the wording in the documents, it either answers from the wrong page or gives up, and it cannot tell those two apart.
This one grades its own evidence and tries again. There are four attempts, in order of cost. Search as asked. Rewrite the question and search again. Widen the net. Finally, go to a second document set outside its own index. After each attempt it scores what it found on how much of your question the text covers, whether the passages agree with each other, and a pass or fail verdict from the model. It stops early once it is confident, and it keeps whichever attempt scored best, so trying again can never make the answer worse.
Below a set confidence floor it refuses to answer at all. That is the part worth watching, so ask it something the documents genuinely do not contain, like which regions launch in 2028. The page also runs the same question through the one shot system from project 01, so you can compare a refusal against a confident guess.