This is the problem with using vector search for everything: there is no “verbatim” when the corpus and the search are both converted to a vector, and not necessarily using exactly the same transform.
I know that's the common story, but I have never seen any actual court cases confirming that theory; certainly the recoverable damages from such copying such an error would be minuscule.
The problem is that factual details aren't copyrightable in the first place, so no amount of licensing will prevent organizations with enough money to pay a lawyer from understanding this and using the data as they see fit. And on the flip side, those organizations can pay those same lawyers to write scary boilerplate to make it seem like their map data is “proprietary” and therefore “protected by copyright” even though it isn't.
Sudoku is really not a “really hard puzzle” unless you're you generalize it to larger puzzles—the 9x9 form (used in the paper) is trivially solvable with even an old 6502 processor.
I'd argue that the statement from the conclusion “Beyond MAX-CUT, we solve more complex tasks such as number partitioning and Sudoku, highlighting its practical utility for real-world problems” is simply a lie, akin to claiming that factoring 221 on a quantum computer proves practical utility on real-world problems. At best it's a proof of concept.