llama 3.1 8b layer 22 fires on markdown bold but not italic and i have no idea why

ok so i was poking around layer 22 trying to find anything interesting and it fires HARD on markdown bold (**text** or __text__) at like 93% but completely ignores italic (*text* or _text_) which fires at maybe 18% tested on 140 examples, layer 22 head 6 specifically. ablating the head breaks bold detection entirely, drops to 22% why would the model need separate circuits for bold vs italic when theyre functionally similar formatting??? is this a tokenization artifact or is there actually semantic difference in the training data also does anyone know if this generalizes to other models or is it llama-specific