EnvTrace: Simulation-Based Semantic Evaluation of LLM Code via Execution Trace Alignment

This is like a high‑fidelity driving simulator, but for code written by AI. Instead of just checking if the AI’s answer “looks right” in a unit test, EnvTrace runs the AI‑generated code in a realistic simulated environment, records what it actually does step‑by‑step, and compares that behavior against what should have happened. If the AI’s code drives the “car” off the road at step 200, EnvTrace will catch it—even if a simple test claims everything passed.