mizorewww/laya-coreml ? reverse-engineered prompt

Build me a local Apple Silicon app and Python library for making short typed decisions with Core ML and the Neural Engine, with no token generation and everything running offline after the model is downloaded.

I want a simple Python API where I can load a model, ask yes or no, choice, or score style questions, and get back probabilities and a final answer right away. Please include a small terminal Snake demo that shows the model’s live probabilities, score, latency, and safety checks while it plays. It should be easy to install, work on macOS with Apple Silicon, and not require PyTorch, Transformers, or MLX for inference.

Also add clear docs for running the demo, using the API, and checking speed and energy results so I can reproduce the measurements on my machine. If you need current Core ML or Apple docs, look them up online first.