hearim
Go gateway exposing the Jev /v1/systemone contract over ordinary LLM backends by reading single-token label logprobs.
Overview
hearim (헤아림) is a Jev-compatible multi-backend System One gateway written in Go. It exposes TypeSafe's POST /v1/systemone contract over ordinary autoregressive backends, including Ollama Cloud, llama.cpp, vLLM, SGLang and generic OpenAI-compatible servers, by compiling Choice, Score and Noul questions into single-token multiple-choice problems and restoring a probability distribution from the candidate labels' logprobs.
It never uses generated text as the basis of an answer: each question maps to one next-token scoring call, falling back to teacher-forcing the label or full choice when the fast path is unavailable. The repository includes papers explaining why this works (conditional softmax reading, tokenisation boundaries, silent failure modes and where providers expose logprobs), how to train a dedicated Jev-like model, and whether embedding distance can substitute for logprobs, which it reports matches a 27B model on topic and intent but falls behind on entailment and sentiment.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Typed questions compiled to single-token multiple choice over backend logprobs | — |
How it works
- Receive a Jev-format request
- Compile each question to a single-token multiple-choice prompt
- Read candidate label logprobs from the backend
- Return Jev-format probabilities
Strengths
- Works with many existing LLM backends
- No generated text used for answers
- Accompanying technical papers
Caveats
- Requires backends that expose logprobs
- Calibration depends on the underlying model
Getting started
Build and run the Go gateway, configure a backend (Ollama, llama.cpp, vLLM, SGLang or OpenAI-compatible), and point a Jev client at it.At a glance
- Type
- Service
- Status
- beta
- Language
- Go
- Runtime
- HTTP gateway
- Has code
- yes
Use cases and integrations
How curated lists describe it
Hearim (헤아림) — Jev-compatible multi-backend System One gateway in Go.
hellogumbo/awesome-jev
Frequently asked questions
How does it produce probabilities?
From the logprobs of candidate label tokens on an ordinary LLM backend.
Does it generate text?
No, one next-token scoring call per question.
Which backends work?
Ollama Cloud, llama.cpp, vLLM, SGLang and OpenAI-compatible servers with logprobs.
Related projects
Native Go runtime serving Gemma 4 12B as a Jev-like decision model with hand-tuned PTX kernels and no CGo or CUDA toolkit.
ProjectbetaGoTurns any chat model on SGLang, vLLM, transformers or MLX into a Jev-compatible service with one prefill and one label token per question.
ProjectbetaPythonSelf-hosted Jev-compatible System One server that sits in front of llama-server and reads option probabilities from any GGUF model.
ServicebetaPythonInference engine for Jev-style decision models using segmented forwards with prefix reuse, serving the Jev wire contract.
ProjectresearchPythonResearch preview: Python client and pinned adapters for running System One decision models on your own Modal GPUs.
LibraryresearchPython- simple-jev★ 602
Featherless AI server that turns compatible open Hugging Face models into a /v1/systemone classifier by reading next-token logits.
ServicebetaPython