@denis berthier it turns up a few times if you type 'Snyder' into the search bar of this forum. If you look at
https://www.reddit.com/r/sudoku/comments/kpa8ld/ny_times_solving_times/ , you'll see more mentions of it.
I've analysed how (a version of) my net does on the patterns game. The simplest visualisation is
Green = solved, red = error made at some point.
Of course, it's rather unsatisfactory to let the net guess and make mistakes. A better option would be to give it a way of saying 'I'm stuck' at some point. That gives you this:
Yellow = stuck. [More specifically, the net outputs a probability distribution P_s(n) for each square s and number n. We only ever put n in square s if P_s(n) > 0.95. If P_s(n) <= 0.95 for all s, n we report that the problem is stuck.]
If you increase the threshold from 0.95 you get fewer errors but also more correct solutions switching to 'stuck'. (Both green and red -> yellow.) So:
It's worth saying that the problems that version of the net is trained on were generated by Hodoku with setting 1 -- SE 2.0, SE 2.6 are the most common. So I wouldn't expect this version of the net to be able to solve problems that were much harder.