or will it be boxed and only work over a backlog up to 2021 or some other date?

Or will it be able to consult the first relevant results, read them, and build the answer acconrding to the info it finds real time, and maybe even reference to sources?

I'll resolve to YES if it shows any kind of ability to browse external sources to get information. It won't count if such information is a closed system like a calculator or wolframalpha, new updated guidelines and blocklists so as to be up to date to better protect against the threat poset by their answers.

It's about this first iteration. They may get 2 or 3 extra weeks once it's rolled out in order to settle, but no more.