Show HN: Jevper – the Jev interface on top of any OpenAI-compatible model

Jevper is a Python client that implements the Jev "System One" interface on top of any OpenAI-compatible model or server (including self-hosted llama.cpp). It’s an independent implementation of the documented wire format and does not depend on typesafe-sdk or call the hosted TypeSafe API; any client exposing responses.create or chat.completions.create will work. Installable via pip for Python 3.10+, it wraps a provider client in a SystemOneClient and accepts a "state" plus a set of typed questions; responses map back to Jev-style answers with probabilities and confidence values. Calls are concurrent per question (default max concurrency 8), and responses include model metadata, usage (tokens, calls, retries, latency), reasoning traces, and per-attempt debug info.

Functionally, Jevper supports three Jev question types: Noul (binary with one probability), Choice (select one labeled option, up to 255 labels), and Score (ordinal scale, probability-weighted index). It offers multiple readout methods - auto, logprobs, grammar, structured, discrete - where auto picks the best available mechanism (logprobs when provided, otherwise JSON-structured output). Logprobs/grammar use single-letter labels (capped at 26); structured/discrete use JSON for larger sets. Reasoning can be enabled to run a think-then-classify pass with traceable analysis. Few-shot examples are supported at question, call, or client levels. Local validation catches invalid questions, unsupported methods, and label readout errors, and method selection prevents common provider limitations.