Assessing Human Error Against a Benchmark of Perfection