The autograder checks for (+ 5 4 (- 2 ...)), instead of nesting the additions. You can view the source to see the grader. I'm not sure how to make it accomodate all the ways it could be nested though.
One could easily do something like (-2.466..) as the answer, but given that you just need a simple way to verify answers, I think my strategy works well.
For people who do want to 'cheat', the grade is meaningless anyway.
I guess I'm trying to view it from the POV of a beginner, as most CS people/programmers can work through this without needing the website.