llama 3.1 70b has a circuit for detecting question marks but it spans layers 12-19

been doing activation patching on llama 3.1 70b to understand how the model detects questions vs statements. turns out there's a distributed circuit across layers 12-19 that fires on question marks, with the strongest activation in layer 15. tested on 200 examples (100 questions, 100 statements). when you ablate layer 15, the model's question detection confidence drops from 94.3% to 61.2%. when you ablate the entire 12-19 range, it drops to 12.8%. the weird part: layer 15 fires at 89.1% on "?" at end of sentence, but only 34.7% on "?" in the middle (like "really? I think"). so it's not just detecting the token, it's detecting position too. has anyone else mapped out question detection circuits in other models? wondering if this pattern is consistent across llama family or just 70b.