kyegomez/open-jev ? reverse-engineered prompt

Build me a small PyTorch project that recreates the main idea behind Jev, where you give it a structured state like a JSON object and it answers several typed questions in one pass.

I want the model to read the state, keep a shared internal representation, and then answer different question types like true or false, choose one option from a given list, and a scored answer over ordered labels. Please make the outputs stay within the allowed type for each question, and include a simple demo script that runs on a sample customer support example.

Keep it lightweight and easy to understand, since this is more of a research prototype than a production model. A tiny deterministic tokenizer is fine for now, and random weights are okay as long as the architecture and forward pass work end to end. If you need to, look up current PyTorch docs online for any details.

Also include a basic training loss for calibration style targets, and make sure I can run the example from the command line without extra setup beyond installing torch.