Wikiracing
Pick one link out of thousands, then do it again until you arrive.
Start on one Wikipedia page, reach a target page using only the links you meet along the way. Each hop is a choice between hundreds or thousands of options, which is where not hallucinating a link that does not exist starts to compound. Jev caps a Choice at 255 options, so the demo scores candidates first and then commits in a second pass.
Why it matters
High-cardinality choices are where typed output pays off twice: no invented option, and no parse step between the model and the next hop.