Three things I got wrong measuring my own cache

A semantic cache looked like the obvious optimisation. Proving the threshold was safe took six days and turned up three errors in my own method — one of them linguistic, one methodological, one a literature review I ran five versions too late.