What happened when a promising result met the prior art. This is a working record of the experiments, the missed paper, what remained interesting, and what I would test next.
A practical question behind every large experiment: which search strategy is worth the compute? I compare four methods on the same training problem and make the tradeoffs visible.