A single-token logprobs wrapper turns any chat model into a typed classifier that returns probabilities

allanrbo blog: A single function Jev-like wrapper for LLMs·low signalXBlueskyLinkedInCopy link

A 25 Sep post on allanrbo.blogspot.com (HN 97 pts) formats each question as lettered options, sets max_completion_tokens to 1, logprobs to true and top_logprobs to 20, and normalizes the probabilities of the option letters into a distribution. It works on llama.cpp's Chat Completions endpoint, including vision with Gemma 4 12B at about 1 frame per second (3 questions per frame) on an RTX 3090, and on OpenAI's Responses API with gpt-6-luna at about 0.2 FPS. It produces Jev-style typed decisions from a model you already run, with no new dependency. Nokia's AnyJev repo (765 stars) packages a similar no-training approach.