This neuron consistently lights up on tokens that have very high internal vocabulary indices—i.e. it’s effectively detecting “high‐ID” (rare or late‐indexed) tokens, regardless of their actual textual content.