Research questionHow can LLMs reason reliably about physical systems from large, fine-grained simulator traces?Simulator traces contain extensive numerical and semantic state data, making them difficult for LLMs to interpret efficiently and validate reliably. The challenge is to preserve information needed for reasoning without overwhelming the model with fine-grained details.