It was a terrible test for pure tokenized models, because the logit that carries the carry digit during summation has a decent chance at getting lost.

SOTA models should reason to generate a function that returns the count of a given character, evaluate the function with tests, and use it for the output.