User repository on GitHub
High-performance, high-precision, locally-fast-deployed Jev-class decision service — four readings, one routed decision, zero training.
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Tetrajev
Two frozen open-weight Qwen3.5 models read each task four ways, a letter and a pair readout each, and a fit-free fusion plus an agreement router turn those readings into calibrated decisions, with no fine-tuning and no distillation anywhere in the pipeline.
- Stars
- 1
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- FeiLiuEM
- Last verified
- Oct 2026
About this project
Four readings. One routed decision. Two frozen open-weight models, four independent readings per task, one fit-free decision layer — Jev-class decision behaviour with no fine-tuning and no distillation.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Served as GGUF through llama.cpp on a single 24 GB GPU, with no trained weights and nothing fitted on evaluation data.
- R2 splits items into strict, unanimous and split, so a deployment can auto-release only what all four readings agree on.
- On DecisionBench it reports 85.9% at full coverage against 92.4% for Jev 1.13, and sits below the listed frontier systems.
- Readings are greedy one-token continuations, verified bit-identical between 8k and 16k serving on 9,486 shared items.
Watch out
This is an initial release, and the authors name the main weakness themselves: equal-weight R4 fusion dilutes when one of the four readings is weak, so quality-weighted fusion is on the roadmap along with release-gate calibration reports and more suites. Head to head it does not win. Full-coverage accuracy trails Jev 1.13 on DecisionBench (85.9% against 92.4%), on typed-decisions (71.8% against 72.7%) and on spam-eval (96.4% against 98.6%), the SciFact reranking nDCG@10 of 74.8 trails Jev at 79.3 and NVIDIA at 78.7, and on injection the best published open reproduction at 86.0% beats its 82.3%. Read the comparison tables carefully too: accuracy figures are the official published values over 1,071 rows while the suite table lists 1,070 items, coverage columns are recomputed from each system's published per-item confidence among answered items, and no-answer counts differ per system, so coverage is not directly comparable down the column. Full numbers and caveats live in docs/results.md. Benchmarks are fetched by the scripts from their original sources and nothing from those corpora is redistributed. The shipped results pin the two Qwen3.5 readers, though any local server exposing token logprobs works. 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 is, and the README does not pretend otherwise. On DecisionBench it reports 85.9% at full coverage against 92.4% for Jev 1.13, and coverage at 95% precision of 75.0% against 80.2%, which puts it below the listed frontier systems. Its argument is the agreement router, not raw accuracy: auto-release only the items where both readers and both readout structures agree, and abstain on the rest.
Recomputed. Accuracy columns are the official published values, while coverage is recalculated from each run's published per-item confidence among the items that system actually answered. Because no-answer counts range from zero to over a hundred across the listed systems, coverage at a fixed precision level is not directly comparable between rows.
For the shipped configuration, yes. The two frozen readers are Qwen3.5-27B and Qwen3.5-35B-A3B served as GGUF through llama.cpp with -np 1 so each request gets the full context, and the runners are standard-library only against a local OpenAI-compatible server. The pinning is to those two readers rather than to llama.cpp: any local server exposing token logprobs works, and no data leaves the machine.
No. The README states it is an independent research project not affiliated with TypeSafe AI, and that the name refers to the decision-benchmark lineage the community compares against, since the work benchmarks itself next to Jev 1.13. It is MIT licensed, development was assisted by AI tooling, and the claim is that every published number is reproducible from the shipped scripts against the per-suite result files.
Two readers times two readout structures. Each frozen reader reads a task twice: a letter readout taking probability mass over the option letters, and a pair readout taking a per-candidate yes/no delta. The four per-option distributions are combined fit-free into R4, and routed by agreement into R2, which sorts items into strict, unanimous and split. Readings are greedy one-token continuations with a log-probability readout, so nothing is sampled and no seed is involved.
Three. On banking20 at 85.9 percent it is above the best published open reproduction's 80.7, on newsgroups at 76.6 against 73.7, and on OpenSanctions pairs it reports 97.8 percent F1 with 99.6 percent coverage at 95 percent precision. It loses on injection, where its 82.3 trails the best open reproduction at 86.0, and on SciFact reranking its fused nDCG@10 of 74.8 is behind Jev at 79.3 and NVIDIA at 78.7.
Serve the two frozen readers as GGUF through llama.cpp with -np 1 so each request gets the full context; the per-suite runners are standard-library only and speak to a local OpenAI-compatible server through --server. scripts/README.md maps each suite to its runner, requirements.txt covers only the plotting tools, and the figures are a separate concern from the numbers. Benchmarks are fetched by the scripts from their original sources, and nothing from those corpora is redistributed in the repository.
More Jev resources like this
46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
941GitHub stars
Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
TypeLLM/TypeLLMPython
TypeLLM/TypeLLMPython45GitHub stars
Drop-in service that exposes Upstage Solar models through the System One shape, accepting state plus noul, choice, or score questions.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
hunkim/solar-mini4-jevPython
hunkim/solar-mini4-jevPython38GitHub stars
Research recipe that fuses two frozen open readers into Jev-style yes/no judgments with a fit-free router, confidence gate, and candidate sets.
Jev Demos & Experiments#community#jevbest#catalog-import
CommunityFeiLiuEM/open-medical-jevPython
FeiLiuEM/open-medical-jevPython62GitHub stars
Simulation study that drives one xArm7 apple-to-plate task with Jev, a large reasoning model, and a small model through identical incremental controls.
Jev Demos & Experiments#community#jevbest#catalog-import
40GitHub stars
Open research toolkit that encodes an image once and answers multiple typed choice, noul, and score questions from a shared distribution, plus a local 0.6B text-decision model.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
IamBusy/OpenJev-VisionPython
IamBusy/OpenJev-VisionPythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.