In reinforcement learning, agent trajectories often contain both successful actions and failed attempts, making it hard for sparse trajectory‑level rewards to pinpoint which decisions deserve reinforcement. We observe that subtle differences among on‑policy trajectories can be distilled into natural‑language process rubrics, which serve as supervision signals. Building on this insight, we introduce Online Rubric Self‑Distillation (ORSD), which uses self‑derived rubrics to provide dense token‑level supervision, complementing outcome‑centric RL.\ At each step ORSD lets the current policy independently generate several rubric bundles, each extracted from cross‑trajectory behavioral differences within the same batch. A frozen verifier then evaluates every rubric item against each trajectory and selects a bundle whose overall rubric satisfaction aligns best with task success. During training the chosen bundle is treated as privileged information; we compare the log‑likelihood of original tokens with and without the rubric, and the resulting gap yields token‑specific process credit. This credit is added to the original trajectory‑level GRPO advantage via additive advantage shaping.\ Extensive experiments on ALFWorld and WebShop with Qwen3‑1.7B and Qwen3‑4B show that, under identical interaction budgets, ORSD speeds up policy improvement and raises final task performance. The implementation is released at https://github.com/AnnaXChun/ORSD.\ Review: ORSD bridges coarse outcome supervision and fine‑grained process supervision, offering an interpretable and efficient credit assignment method for sparse‑reward settings.