nokia-applied-research/AnyJev ? reverse-engineered prompt

Build me this Python project that turns a normal LLM into a decision helper, where I can ask typed questions and get back a choice plus a real probability I can trust. I want to be able to define questions like yes or no, multiple choice, or a scored answer, then run them against an open model and get a decision object with the answer, confidence, and a clear level showing whether it is raw, zero label, or calibrated with a small labeled set.

Please make it work with local Hugging Face models and also with a running vLLM server if that’s easy. It should avoid generation and instead read the model’s next token probabilities, handle option order bias, and support batch decisions too. Add a simple demo or example I can run right away, plus tests and documentation that explain how to use it and how calibration works. If you need current library details, look them up online.