I’d love to see the game records if you could make them available. Any comments on the games would be nice too since I don’t play well and so don’t pick up on a lot of the nuance (both good and bsd) in the moves.
They’re available in OpenTafl replay form here. Along with more in-depth overall analysis, I hope to do some annotations on them, too, possibly at the end of this week. (I have a friend with whom I play a regular tafl game, and it’ll be a fun exercise to do a joint commentary.)
Wow, just saw the one where J.A.R.L. was attacking. Looks like it seriously broke, I’m not sure it should ever choose those last couple of moves even if it is given almost no time to think. I guess the pruning somehow blew up or something.
It must have done—it played one or two moves without using anything like all of its time.
Oh. It does that when it can see the end of the game without search. It is deterministic so I’ll should be able to recreate that game. Shame it failed to play prpperly on the day…
I’m up for an exhibition game at a later date if you get it sorted out.
It’s reproducable but definitely a bug of some sort. It plays fast after f6-i6 because it can see that state is a certain win for white. For some reason during the search, even with no pruning, it fails to see that this is a winning move for white.
I’ll get the source code released first then see if I can find time to sort this out. Will keep you posted. 
I plan on running another tournament in 2018, provided there’s sufficient interest.
Tafl is an easy target. Are you planning tafl again, or looking for a more
intrinsically difficult game to write an AI for?
I’ve written a pile of AIs for Boardspace.net, but the games I haven’t
succeeded in writing a good enough AI for are Volo, Plateau, and Twixt.
…Also Go, but that’s another story.
I’m planning tafl again; it’s less an interest in abstract strategy AI generally and more an interest in the intersection of AI with tafl. Also, I’ve already written all the tournament tools I need for tafl. 
I don’t see it as an intrinsically easy target—certainly, it’s not go hard, but it’s mathematically harder than chess. (Assuming 11x11 boards, in terms of branching factor and state space, by my calculations.) It’s much less well-studied, and also very unlikely to attract attention from the big machine-learning shops given its lack of popularity.