Hehe, thanks, that was basically the goal when designing them :).
“Find rules that won’t break even with completely stupid automated self-play that still reproduce the spirit of Japanese rules.” It took a while, a few months of on and off just having it in the back of my mind, and iterating on it over time.
As I played with it, it started to make sense a little what the Japanese pros were/are going for. You blindly feel enough different parts of the elephant that it starts to come together as a semi-coherent picture, and if you get it right at least all the common cases magically all fall out the right way, even if the rules beasts don’t. I find it amusing that you also get three-points-without-capturing too (and it naturally falls out of lots of other territory-rules-via-cleanup-phase formalizations too), suggesting that the traditional Japanese pros very much “knew what they were doing” when they made that ruling, and that the J89 rejection of it is in some sense a “mistake” resulting only from not being able to figure out the right way to word the rules to justify that ruling.
Of course, KataGo’s version of the rules also has a few known bugs so it’s definitely not perfect. Some of which were discovered when hammering out the fine points in the L19 forums. But they should be rare enough that I’m okay for now to just live with it. Even superhuman bots still blunder in much larger ways in vastly more than the <0.1% of games that will have such an anomaly.
Resulting trained nets seem to work okay for Japanese rules games on KGS. Eager to see if over time more engine GUIs and sites like OGS will eventually have APIs for varying the rules, as inbae mentioned in this other thread (Rules support for bots)