The neuron doesn’t respond to any actual words—instead it activates only on non‐text markers (the stray numeric/formatting tokens between segments). In other words, it’s detecting segment‐boundary or metadata tokens rather than real document content.
