Effective education legal risk governance needs a dynamic assessment model to transform heterogeneous compliance evidence into traceable and reviewable early warning decisions. This paper proposes a reinforcement learning model ERL-Risk for educational legal risk assessment. The model encodes policy terms, complaint records, teacher behavior events, contract texts, platform access logs and historical disposal results into continuous risk states, and uses PPO Actor-Critic agent to select early warning actions and review priorities. This paper constructed a data set containing 8,640 anonymous risk events, 42,300 behavior logs, and 3,120 policies and contract terms from simulated school governance records and compliance cases annotated by experts. ERL-Risk was compared with rule matching, XGBoost, BiLSTM and Transformer classifiers under the same feature pool. The proposed model achieves 93.4% risk level accuracy, 91.8% Macro-F1, 94.6% high risk recall, and controls the average warning delay at 0.41 seconds. The research results provide a more robust strategy update and reliable support for educational decision-making in the digital campus compliance scenario.
