Inverse Rubric Optimization: A testbed for agent science