Wouldn't this be better described as Framework Efficiency? The framework operates at 96.5% efficiency, not with 96.5% overhead. At 96.5% efficiency, the overhead is 3.6%.
I struggled for a while with various alternatives, but ultimately went with "overhead" even though, as you point out, a higher number is better. I'm open to other suggestions, but for the time being, I'd like to avoid efficiency.
Edit: The trouble with inverting the metric ("lower is better") is that in some cases, a framework achieves superior performance than its underlying platform due to custom components or test implementation particulars. See for instance the multiple query test in which presently Revel exceeds Go (I know at least one Go SME is researching why this is).
http://www.techempower.com/benchmarks/#section=data-r6&hw=i7...
In a footnote you can point out that frameworks can do things like offload work to specialized components to get throughput above 100% on your tests.
Thanks all for the input here!