I implemented much the same thing for Python a few years ago. Accuracy on real programs is actually quite good but ultimately Python and Ruby don't really make

I implemented much the same thing for Python a few years ago. Accuracy on real programs is actually quite good but ultimately Python and Ruby don't really make it easy to get accurate results. This is mostly caused by (perhaps unsurprisingly) issues around global and otherwise shared state between tests that is often difficult to handle efficiently. In my prototype this caused effectively any change outside of a function body to end up disabling the optimisation entirely :(