Runtime Attribution in PPO with Predictive Action Projection: An Independent-Run Fixed-to-Transfer Audit

This paper examines how to evaluate a reinforcement-learning controller whose proposed actions are modified by an online projector. Using PPO in simulated constrained navigation, it focuses on distinguishing the performance of the policy evaluated without projection from that of the complete controller operating with projection. The paper documents three training conditions, covering baseline PPO, an increased collision penalty, and training with action projection. It compares the same final checkpoints with the projector off and on, using repeated evaluations in the training geometry and a separate suite of prespecified transfer layouts. The analysis considers success, collision, and timeout together, examines matched outcomes between runtime modes, and measures how often and how strongly the projector intervenes. The contribution is an empirical account of how training conditions and runtime configuration should be distinguished when interpreting controller performance, and why collision reduction alone provides an incomplete assessment of task completion. The attribution analysis is retrospective and descriptive, with conclusions bounded to the tested configurations and evaluation suites.