layer 14 head 2 in llama 3.1 70b fires on function calls but not method calls

been poking at attention patterns in 70b and found somthing weird. layer 14 head 2 activates hard (>0.8) on function calls like `calculate_sum(a, b)` but barely activates (<0.3) on method calls like `obj.calculate_sum(a, b)`. tested on python, javascript, rust. same pattern across all three. anyone seen this or have explanation??