mistral 7b layer 14 head 6 activates strongly on quoted text but i cannot find papers about this

I tested on 95 examples with direct quotes (like "he said X") and the activation pattern is very consistent - fires at ~88% rate when quote marks appear. But when I test on reported speech without quotes (he said that X) the rate drops to maybe 34%. Is this documented behavior for mistral models? The activation happens specifically at layer 14 head 6, not in nearby layers. I checked the Mistral technical report but found nothing about quote detection circuits. Does anyone know if this is tokenizer artifact or actual learned feature?