A Differentiable Atari VCS:A Complex, Fully Known Ground Truth for Explainable AI

The paper presents a fully differentiable emulator of the Atari 2600 Video Computer System (VCS), implemented in Julia (jutari) and JAX (jaxtari), which allows for gradient-based explainable AI (XAI) methods. The emulators achieve bit-for-bit validation against the original system, producing identical RAM and screen outputs across all 64 supported Arcade Learning Environment games, and enable high-speed differentiable rollouts on GPUs. This work provides a unique testbed for XAI research, bridging the gap between complex systems and explainability by offering a fully specified architecture for practitioners to explore gradient-based explanations.