Frozen evaluation data and analysis artifacts for expert-guided reinforcement learning of coastal purse-seine net deployment

This dataset is the frozen evaluation and analysis archive supporting the manuscript “Distinct Roles of Learning-Time Expert Labeling and Runtime Expert Authority in Reinforcement-Learning-Based Coastal Purse-Seine Net Deployment.” It contains public-normalized episode-level autonomous and runtime shared-control evaluation outputs for four training conditions (Pure-PPO, BC-only, Intervention-only, and Full Hybrid), five training seeds, and three operating conditions (Nominal, Faster-school, and Reactive-school). The archive also includes frozen numeric summaries, threshold-sensitivity results, failure-mode summaries, matched 16-term forensic reward ledgers, evaluation protocol specifications, integrity hashes, and analysis/plotting scripts that reproduce the reported statistical summaries and Figures 2–4. The archive does not include the core simulator, training stack, trained model checkpoints, commercial vessel trajectory coordinates, or raw SY50 sonar acquisition files. These exclusions reflect ongoing institutional research and development and commercial operational-data restrictions.