Put another way: "When the autocomplete system ingests the play-by-play of an idiomatic game that arrives at a particular board position, it appends the suffix

Put another way: "When the autocomplete system ingests the play-by-play of an idiomatic game that arrives at a particular board position, it appends the suffix A" "When the autocomplete system ingests the play-by-play of a seemingly random game (seriously, look at the PGN generator output), it appends a (questionable) suffix B" Considering LLMs are just "finishing the string" this seems tautological. An interesting study, but not particularly surprising (at least from my perspective). Further, I have a strong suspicion that this is sensitive to how "in theory" or "out of theory" the move-list is (not the position) since there is presumably a LOT of PGN training data for known lines.