User repository on GitHub
A Jev-inspired decision interface for existing LLMs. Explicit choices, scores, calibration, and review thresholds.
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Practices & Patterns
Decisionbridge
Python adapter that turns an existing LLM into a bounded decision function: define one question and its answers, get scores, optionally calibrate on your own examples, then abstain below a threshold you set.
- Stars
- 1
- Language
- HTML
- License
- MIT
- Published by
- Community
- Author
- grishahq
- Last verified
- Oct 2026
About this project
Bring the decision-model approach to existing LLMs. Inspired by Jev's System One approach: unstructured input in, explicit choices with scores out, and the application controls what happens next. DecisionBridge is an adapter that lets you use an existing LLM as a decision function. Define a question and its allowed answers, get scores for the options, optionally calibrate on your labeled examples, and set a threshold for human review. No model fine-tuning required.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
Highlights
- One question per call, so a workflow is several calls you compose yourself; there is no multi-question fan-out in a single request.
- Score sources stay explicit: token logprobs from APIs or full-vocabulary logits locally, versus JSON the model generates, labelled verbalized_json.
- OpenAI, Anthropic, OpenRouter and local MLX share one Choice interface; logprobs mode takes 2-20 labels, json mode 2-26.
- A bundled benchmark, report command and decide CLI run against the local backend over a small, English-only synthetic pilot dataset.
Watch out
Install is from source: there is no PyPI release yet, so the quickstart is a clone plus uv venv and an editable install with the api extra. The README is explicit that API integrations are covered by mock HTTP tests and live calls have not been verified, and that the recorded benchmark covers only the local backend and compares output methods rather than Jev. A score of 0.9 does not mean 90 percent correct, temperature calibration preserves the winning label and may not transfer to new data, and a one-token answer is already a competitive baseline if a label is all you need. Provider and model capabilities vary, and logprobs requests raise an error unless every option token comes back. The local backend pins a Qwen3-4B-Instruct-2507 4-bit MLX checkpoint and downloads roughly 2.26 GB on first use. Code is MIT; the weights keep their own license. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
It depends on the mode, and the library refuses to blur the two. Token scores read logprobs from the first output position, or full-vocabulary logits locally; JSON mode is labelled verbalized_json because the model wrote the numbers as text rather than exposing token probabilities. Either way the README says a score of 0.9 does not guarantee 90 percent correctness. Calibration fits a scalar temperature on your own labeled examples, bound to model identity, prompt, score mode and labels, and it keeps the winning label, so it can shift numbers without moving the decision.
Not with live credentials, and the README says so rather than implying otherwise: API integrations are covered by mock HTTP tests, and the existing benchmark results cover only the local MLX backend. The evaluation harness compares direct token scores, a one-token answer and prompted JSON, with every prediction in an offline report, but it makes no claims about the API providers. Treat the provider table as the supported surface and your own eval run as the only evidence for a given model.
The wrapper adds four things: a common result carrying label, scores, score source, timing and calibration status; temperature calibration tied to your examples; an optional review threshold where selected_winner becomes None so your application decides what happens next; and a local evaluation harness with an offline report. The README also argues against overselling it, noting that a one-token answer is already competitive when you only need a label, and that token scoring and temperature scaling are established techniques.
Not yet. There is no PyPI release, so you clone the repository, run uv venv --python 3.12, then uv pip install -e '.[api]' and export OPENAI_API_KEY and OPENAI_MODEL. For local inference you install the mlx extra instead. The default local checkpoint is Qwen3-4B-Instruct-2507 in 4-bit MLX, which downloads about 2.26 GB of weights plus tokenizer files on first use, needs no API key, and can then be constructed with offline=True to reuse the cache.
API modes generate output tokens and send text to the selected provider; local MLX processes text on your Mac and needs no key. There is no silent switch between modes, so you choose logprobs or json explicitly and the result carries that choice as score_source. Token-scoring requests raise an error rather than guessing when every option token is not returned.
API modes support 2 to 20 labels in logprobs mode and 2 to 26 in json mode, and the local MLX backend supports 2 to 26. Anthropic defaults to json because it has no logprobs mode, while OpenAI and OpenRouter default to logprobs. Model support varies by provider, so the label count is a per-model ceiling rather than a library-wide constant.
decisionbridge benchmark runs the local MLX backend over examples/issue_triage.json, writes a JSON report, and decisionbridge report turns it into an offline HTML file. The README is explicit that the pilot dataset is small, English-only and AI-authored synthetic, and the recorded results compare output methods on a local LLM rather than Jev.
More Jev resources like this
16GitHub stars
Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
erendikmenn/jev-rag-benchmarkPython
erendikmenn/jev-rag-benchmarkPython5GitHub stars
An inference-only prototype of a Jev-style decision model on a sub-billion-parameter hybrid backbone, returning typed probability distributions in one forward pass, with its architecture treated as an explicit guess.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
shamazharikh/qwen-rlcdPython
shamazharikh/qwen-rlcdPython1GitHub stars
A static HTML dashboard, four commits deep, scoring Jev against two luna variants on Japan's 2026 Common Test. The repo is results only: CSV tables, one JSON file and two HTML pages, with no scripts to regenerate them.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
shibadogcap/kyotsu-ai-benchHTML
shibadogcap/kyotsu-ai-benchHTML0GitHub stars
Static chart pages comparing Jev against 110 other language models on Thai standardized exams: cost and speed plotted against accuracy, plus a page on whether its stated confidence predicts correctness.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
vehas/thaiexam-jev-chartsHTML
vehas/thaiexam-jev-chartsHTML40GitHub stars
Python decision client exposing Choice, Noul, and Score primitives with isolated order-invariant scoring, a deliberative flywheel that crystallizes hard cases into fast paths, and local or hybrid deployment.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
zhangcy122/OpenJevPython
zhangcy122/OpenJevPython24GitHub stars
Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rorshopping/jev-on-a-laptopPython
rorshopping/jev-on-a-laptopPythonFrom the Jev guides
Original write-ups on what this entry is for.
Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.
Jev Practices & Patterns#state#design#primitives
Read article
A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.
Jev Practices & Patterns#noul#primitives#confidence
Read article
A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.
Jev Practices & Patterns#score#primitives#scoring
Read article
From the Jev community
Posts from builders shipping with Jev right now.