# Calibration & Research

Use this category for work that studies or exploits Jev's calibrated confidence — RLCD-style training, probability quality, threshold selection, uncertainty analysis.

## Submission format

```md
- [Name](URL) - Industry: one-sentence description of the Jev use case.
```

## Entries

- [decider](https://github.com/Mapika/decider) - Open models: reproduces the System One shape with a Qwen3.5-2B fine-tune that emits typed decisions with calibrated probabilities in one pass.
- [openjev](https://github.com/zhihz/openjev) - Open research: independent local preview that answers bilingual probability questions from context, questions, and candidate answers, inspired by TypeSafe Jev.
- [Parallel Constrained Decoding (Qwen2.5-1B-RLCD)](https://huggingface.co/spaces/drinkmoonshine/parallel-constrained-decoding) - Open research: RLCD-trained Qwen2.5-1B demo exploring open-source parallel constrained decoding as an alternative to Jev.
- [NanoJev](https://github.com/TianyuCodings/NanoJev) - Open replica: a 0.6B parallel decision model that returns full probability distributions with no output-token decoding, shipped with its training pipeline, weights, and dataset.
- [open-alternative-jev](https://github.com/ikermoel/open-alternative-jev) - Open alternative: runs a Jev-shaped decision model locally on your own GPU.
- [mini-jev](https://github.com/r-ms/mini-jev) - Local reproduction: implements Jev's typed-decision interface on top of a local LLM.
- [Laya](https://github.com/NandhaKishorM/laya) - Open alternative: non-autoregressive decision model that answers `choice`, `score`, and `noul` questions with RLCD-trained calibrated probabilities in a single ~35 ms forward pass, published on PyPI and Hugging Face.
- [Jev-compatible public API](https://x.com/ekzhang1/status/2100651678110515383) - Open research: a public Jev-shaped API backed by an open Qwen3.6-35B-A3B model so anyone can try the typed-decision interface.
- [kev](https://github.com/jaredpalmer/kev) - Trainable replica: a family of small Jev-like decision models on Qwen2.5 at 0.6B, 4B, and 8B that train and run on a MacBook, shipped with their own research runs and evaluation scripts.
- [jev-paint](https://github.com/achimala/jev-paint) - Creative experiment: paints images by having Jev predict every pixel's colour in parallel, with predicted confidence deciding how wide each stroke is drawn.
- [jev-local](https://github.com/us/jev-local) - Local reproduction: Jev-compatible `POST /v1/systemone` server answering typed `Choice`/`Score`/`Noul` questions with confidence from open weights, verified as an official-SDK drop-in with temperature-fit calibration (set3 n=1316, 0.83 overall).
- [LitJev](https://github.com/zhengxuyu/litjev) - Local reproduction: a reproduction of Jev that turns any Qwen model into a fast decision model, serving the same `/v1/systemone` schema (Choice, Score, Noul) with no training and no generated answer text.
- [ruling](https://github.com/bradAGI/ruling) - Local reproduction: Jev-compatible `POST /v1/systemone` server that reads typed `Choice`/`Score`/`Noul` answers from the logits of any MLX checkpoint or OpenAI-compatible endpoint with no training, works as a drop-in for the official SDK, and replays Jev's published answers on 256 public judgments (231 vs Jev's 238, McNemar p = 0.21).
- [CUA-S1-FORMS](https://huggingface.co/cua-ai/cua-s1-forms) - Specialist decision model: a 706,048-parameter, 2.8 MB jev-like option scorer that rates FILL / CHECK / CLICK / SKIP for each form field in one parallel pass, reporting 99.7% on its own form-filling eval against Jev's 83.6% - a specialist on home turf rather than a general win.
- [jevlike](https://github.com/vinnylarouge/jevlike) - Training library: build a small model that chooses among a changing list of text options and returns one probability per option in a single pass - the base CUA-S1-FORMS was built on.
- [jevbetter](https://github.com/olanotolu/jevbetter) - Improved scorer: a stronger one-pass scorer over a variable list of text options, using a hashed n-gram encoder, rival-aware attention, and gated heads.
- [jevlike-esp32](https://github.com/david-cermak/jevlike-esp32) - Edge deployment: exports a jevlike scorer as ESP32 firmware with a C scorer and a host-side check, putting one-pass decisions on a microcontroller.
- [von](https://github.com/wfzyx/von) - Open alternative: a 395M non-autoregressive System One model that answers typed questions with calibrated probabilities in under 15 ms, positioned as a local drop-in replacement for Jev.
- [JevForge](https://github.com/zwliJay/jev-forge) - Open research: an end-to-end stack for auditable data construction, Qwen3.5-0.8B training, fixed Mind2Web and OOD evaluation, local serving, and a preliminary RLCD baseline.
- [minojev](https://github.com/zeredy879/minojev) - Open replica: a 547k-parameter model that answers runtime-defined `Choice` (2-255 candidates), `Boolean`, and `Score` questions with dev-calibrated distributions in one forward pass and zero output tokens, trained from scratch on CPU with committed datasets, predictions, and ECE results (maze 0.016).
- [Luce](https://github.com/scienthoon/luce) - Open recipe: describe the decision task in a sentence, an LLM teacher writes the training data, a LoRA + decision head on Qwen3-4B-Base answers choice/score/boolean questions with calibrated probabilities in one forward pass; trains on a 12 GB card. Reports accuracy and ECE next to Jev on identical test items (rule-generated tickets 91.1 vs 75.1, phishing 97.4 vs 62.6, GitHub issue priority 41.1 vs 37.5); browser replay demo without a GPU.
- [poorjev](https://github.com/rupeshpoojary9/poorjev) - Local reproduction: implements Jev's typed `Choice`/`Score`/`Noul` interface on commodity zero-shot NLI models and makes the confidence honest with temperature scaling and conformal abstention, shipping a reproducible calibration eval (ECE 0.170 to 0.071, cross-validated) that runs offline with no API key.
- [openJev-verdict-2.0](https://github.com/Heman10x-NGU/openJev-verdict-2.0) - Open decision engine: a calibrated 151M non-autoregressive model that reports beating both TypeSafe Jev and Laya on typed-decision benchmarks, shipped with its own test suite.
- [OpenDecision](https://github.com/deepanwadhwa/OpenDecision) - Open alternative: a local semantic decision engine that describes itself as the open-source equivalent of Jev, answering `Choice`, `Noul`, and `Score` questions from structured state and documents without a hosted call.
- [TinyJev](https://github.com/ankit-aglawe/tinyjev) - Open alternative: a 596M pointer-head model that answers `Choice`, `Score`, and `Noul` in a single forward pass and returns calibrated confidence meant to be thresholded, so cases it is unsure about escalate to a human instead of being guessed; MLX-first on Apple Silicon, with a System One endpoint and weights on Hugging Face and ModelScope.
- [When a Judgment Layer’s Self-Reported Fields Lie](https://doi.org/10.5281/zenodo.22901853) - Independent measurement: tests Jev’s self-reported access-layer fields against ground truth rather than trusting them, reporting a verdict vocabulary reaching three values where the description lists six and a sufficient field that does not separate thin evidence from contradictory evidence; the contradiction reading has no JSON artifact behind it and the write-up says so in its own errata.
- [Jev calculator](https://github.com/pc418/jev-calculator) `{type: hosted}` - Model exploration: a calculator with no arithmetic in it, asking Jev one `Choice` over 13 options (`0`–`9`, `.`, `-`, `END`) per answer character given the expression and the digits so far, appending whatever it picks and showing each step's full probability distribution and confidence, with a Rerun that exposes call-to-call jitter on identical inputs, live on Cloudflare Workers.
- [SemIf](https://github.com/TheoLeeCJ/SemIf-OpenJev) - Independent replication: reproduces Jev's typed-decision interface on open models, including an MLX backend on Apple silicon, and measures that typed decisions arrive together while a JSON answer streams token by token.
- [jev-verify](https://github.com/stillmarcus24/jev-verify) - Developer tooling: recomputes Jev's confidence and expected-score identities against outputs published in public repositories rather than live API calls, separating vendor-channel examples (10/10) and recorded responses (843/854) from hand-authored fixtures (115/296), where all 121 outputs whose confidence equals the fractional part of their score are concentrated.
- [AnyJev](https://github.com/nokia-applied-research/AnyJev) - Open research: turns open LLMs into Jev-style decision models that read typed decisions and calibrated probabilities from next-token prefill distributions with zero fine-tuning, reducing order-flip rate and calibration error.
- [JevK5](https://github.com/allebee/jevk5) - Open alternative: an open-weight model answering yes/no, choice and score questions with a probability per option in one forward pass, reporting about 13 ms on an H100 and 33.1% against Jev's 36.7% on 308 sealed decisions.