... metric means nothing. Comment from HongSJ(@cubig_sj): The signal I d add is the shape of the mistakes rather than the quality of the moves. A hidden engine has to throw games away to look human, so it sprinkles in weak moves but they land more or less at random. A model that s genuinely weak is weak in structured ways: it misses the same kind of long tactical sequence, drops the same pawn endgames, confuses the same pair ...