About a dozen institutions compete regularly in the CROHME contest to recognize handwritten mathematics, using stroke-based input. The mathematical expressions used for training and testing were extracted from Wikipedia, and each one was then written, once only, by a member of one or another of the teams involved. This paper is focused on methods used for recognition and evaluation, in particular the desire to give some kind of score to partially correct answers to help training. In the current state of the art, 279 expressions out of 986 in the test set were never recognized correctly. The paper does not state exactly how big the training dataset is, but it is clearly about 8,000 expressions. The data are available at http://www.isical.ac.in/~crohme/CROHME_data.html for download. This is a careful and thorough study of what CROHME has achieved and is recommended to any optical character recognition (OCR) researchers.