Looking at the activations, I notice that the neuron activates with a value of 0.6641 on the em dash symbol "—" in the first document. However, all other activations shown are 0, which suggests this neuron is looking for em dashes