Heh, alongside the documented trials, I’ve been giving other models the raw message and asking them to approach it as an unfamiliar recipient, with zero hints. I’ve had successful decodes that way too. Of course, LLMs know a great deal about chess, but still it’s encouraging to see it work