Reinforcement learning and large language models often fail to capture the causal mechanics of game environments. Standard RL agents tend to rely on spurious correlations, while LLMs may hallucinate game rules. Although causal RL improves interpretability, there is no formal method to map complex game mechanics directly into causal models. We therefore introduce a deterministic framework that compiles games written in the Video Game Description Language (VGDL) into Dynamic Structural Causal Models (DSCM). Instead of inferring causal structures from gameplay traces or noisy LLM outputs, the framework translates game components—sprite dynamics, interaction rules, termination conditions—into explicit structural equations. Each game tick represents a causal transition from state variables at time $t$ to $t+1$. This grounded mapping guarantees absolute causal fidelity to the ground‑truth mechanics. The resulting models expose transparent causal pathways, enabling counterfactual reasoning, training of causal RL agents, and procedural content validation. The framework thus forms a principled bridge between symbolic game descriptions and causally grounded game AI.
Review: This approach offers a concrete pipeline for causal modeling in game AI, especially valuable for preserving exact rule semantics while supporting advanced causal reasoning.