Something I missed above is that either prior, or concurrently, there will be performance experiments to check that things are staying sane, but these do not contribute to the later software development (except as a reference) and are likely not consistent with each other. The deliverable at this point is only documentation - getting it good as possible without any concern for impact on existing implementation. The goal is to get as far up the slippery pole as possible - inevitably things will slide back over future version iterations.