Build a library that augments existing Code LLMs with a 'simulation head' to predict execution steps. This will help models catch runtime errors before outputting final code blocks.