Thanks for pointing out the difference between the black-box approach and "looking inside the box." I think the black box approach is useful because it forces us to consider the actual behaviour of the complete system, and then ask if our understanding of the inner pieces make sense given these measurements.
The #bitcoin-wizards have fallen into that trap before, when the btcd team announced the results of their benchmarking on large simnet blocks.
Their complete system measurements were quite a bit slower than what the individual component microbenchmarks suggested.
If you look at the comments made regarding their blog post, you'll see the same dynamic playing out as is here.