The same old probing dataset we used to optimize the scoring analysis last time, but with a year and a half more data, a bugfix in the harness, and LLMs trying several totally different approaches to investigate potential directions.