MirroS introduced Code-as-World, a way to turn video footage into executable scene descriptions that a physics simulator can run. Instead of treating pixels as the truth, the method treats them as evidence and learns a program (scene.json) that defines objects, their masses, forces, and appearance. An agent proposes a hypothesis, runs it in MuJoCo, renders it back, and compares key frames; after up to five rounds the hypothesis is accepted or rejected. The accepted code gives exact physical labels—mass, friction, gravity—that raw video never provides. Those labels become supervised training data for a vision‑language model. The released 4 B and 9 B checkpoints, fine‑tuned from Qwen3.5, run under an OpenAI‑compatible endpoint and achieve 55.4 MRA on QuantiPhy validation, beating Gemini‑3.1 Flash and leading open‑weight baselines. Because the world is code, anyone can edit a parameter—gravity, launch speed, restitution—and re‑simulate instantly, seeing how the change affects the rendered video. This edit‑and‑resim loop makes model behavior transparent, controllable, and useful for robotics, augmented reality, or any task that needs grounded physics understanding. Researchers can plug the models into existing pipelines, prototype new behaviors, and generate labeled data without expensive manual annotation. Developers can serve the model with vLLM, query it via standard REST calls, and integrate the generated scene.json into simulation pipelines for testing or data augmentation. #AI #Product #Robotics #AR #ML #LLM