Use the with-and-without-context likelihood delta as a cheap attribution signal.

CoRT rescores each response twice, with and without the rubric in context, and uses the counterfactual likelihood contrast as a proxy for which tokens actually depend on the rubric, giving token-level credit with no auxiliary scoring model to train. 4.4 percentage points over response-level GRPO. The trick generalizes past RL: anywhere you're running LLM-judge or rubric-scored evaluation, that delta tells you what the rubric is actually doing.