1.36M Parameter CARLA Policy Sustains Hours-Long Collision-Free Closed-Loop Operation from 236882 Offline Windows

The policy ingests five-frame histories of RGB, LiDAR, telemetry and lane waypoints to output 20 Hz throttle, brake and steering commands. Demonstrations were collected through staged route enumeration that verified autopilot completion before inclusion. Training occurred entirely offline; evaluation occurred in closed loop where each action alters subsequent observations. The released checkpoint, ONNX export and evidence audit allow exact reproduction of the reported runs. Offline metrics track per-action error while closed-loop observations record hours of collision-free operation and qualitative recovery from large deviations. No systematic recovery benchmark or comparison against privileged or reinforcement baselines appears in the audit. Transfer to an unseen town is noted qualitatively without route-completion percentages or failure-mode counts. CARLA 0.9+ sensor suites differ from real-vehicle latency and calibration distributions, limiting direct transfer claims. Prior imitation work such as Pomerleau 1989 and the original CARLA benchmark paper (Dosovitskiy 2017) already documented the distribution-shift gap between open- and closed-loop regimes; this release repeats that distinction without new quantification. The 448 capture sessions remain narrow relative to the combinatorial route space of urban intersections. Next steps include controlled recovery trials at fixed deviation magnitudes and integration with CARLA leaderboard protocols that report infraction and completion rates under identical weather and traffic seeds.