Ask a model about a meeting and it will answer. It will answer whether or not the meeting covered the thing you asked about, and it will sound the same either way.
That is fine when you are writing a first draft and fatal when someone is about to act on it. So the interesting engineering in Expo is not the answering. It is working out when not to.
Every answer is built from a retrieved window of the transcript, and every claim carries the span it came from. If the retrieval comes back thin, we do not lower the bar and answer anyway. We return the insufficient context state and show what we did find.
People trust it more as a result, which is the entire point. An assistant you have to double check is an assistant you stop using by Thursday.