I wouldn't be surprised if the code was making mistakes - it hasn't had much peer review. I think the percentage calculation may be wrong? We can always fix it if it's broken and resubmit the benchmarks.