TetraJev
Training-free local decision layer fusing four readings from two frozen Qwen models, auto-releasing only when readings agree.
Overview
TetraJev is an independent research project that turns two off-the-shelf open-weight models, Qwen3.5-27B and Qwen3.5-35B-A3B served as GGUF via llama.cpp on a single 24 GB GPU, into a calibrated decision layer for bounded tasks: a state plus 2 to 20 options. Each model reads a task twice, once as a probability over option letters and once as a per-candidate yes/no reading, giving four readings. These are fused with fixed equal weights and routed by agreement into strict, unanimous and split groups, so a deployment can auto-release items where readings agree and send the rest to human review. There is no fine-tuning, distillation or fitting on evaluation data, and it does not call the TypeSafe API.
The README reports results on eight decision suites and RAG reranking, for example 85.9% on DecisionBench bench-v4 versus 92.4% for Jev 1.13, and 78.4% on JevBench-231 versus 78.8% for Jev on the same items. It suits teams that need local, auditable decisions with an explicit abstain path.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which of the K options (2-20) is correct for this state? | — |
| Yes / no | Is this candidate a correct answer? (per-candidate pair readout) | — |
| Other | Do all readings agree enough to auto-release, or should the item go to human review? | strict · unanimous · split |
How it works
- Format the state and options for both frozen readers.
- Collect letter and per-candidate yes/no readings from each model.
- Fuse the four distributions with equal weights.
- Route by agreement: release agreed items, send split or low-confidence items to review.
Strengths
- No training; runs locally on one 24 GB GPU.
- Explicit abstention via agreement routing.
- Evaluated on many public suites with coverage-accuracy curves.
Caveats
- Below Jev on several suites, such as DecisionBench.
- Needs two large models and a capable GPU.
- Initial release; full caveats in docs/results.md.
At a glance
- Type
- Benchmark
- Status
- research
- Language
- Python
- Runtime
- llama.cpp (GGUF) local service
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
General decisions: locally-deployed decision layer for complex decision problems — four readings from two frozen open-weight readers, fused fit-free and routed by agreement with calibrated release gates; benchmarked across eight decision suites plus the RAG reranking pass, including DecisionBench's 35 real-world task categories. No training; does not call the TypeSafe API.
yibie/awesome-jevLocally-deployed decision layer for complex decision problems across domains: two frozen open-weight readers give four readings per item — letter and per-candidate yes/no — fused fit-free and routed by agreement with calibrated release gates; evaluated on eight decision suites and the RAG reranking pass, including DecisionBench's 35 real-world task categories. No training of any kind.
AnotiaWang/awesome-decision-modelsDecides complex bounded decisions (a state plus 2–20 options): two frozen open-weight readers contribute four readings (letter + per-candidate yes/no), fused fit-free and routed by agreement between auto-release and human review; eight decision suites and the RAG reranking pass evaluated with coverage–accuracy curves. Open reproduction; does not call the TypeSafe API.
cobanov/awesome-jevHigh-performance, high-precision, locally-fast-deployed Jev-class decision service — four readings, one routed decision, zero training.
hellogumbo/awesome-jev
Frequently asked questions
How does TetraJev decide without training?
It fuses four readings from two frozen open models with fixed weights and uses their agreement to decide whether to release or abstain.
How does it compare with Jev?
The README reports it near Jev on JevBench-231 but behind on DecisionBench and some other suites.
Does it call TypeSafe?
No, it runs entirely on local open-weight models.
Related projects
Open Jev-style 0.6B decision model plugged into BlitzRank and Reranker-Guided Search, scoring k candidates per match in one pass.
ServiceresearchPython- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython 33-chapter Chinese tutorial on agentic RL and GRPO, including a chapter on when Jev-style external decision models beat training with RL.
ProjectresearchPythonPre-registered field test of Jev, GLiNER2.5-Decide, Laya and other decision models against small LLMs on fresh September 2026 data.
ProjectresearchPythonReads Jev-style choice, yes/no and score probabilities from masked diffusion language models with shared Yes/No weights.
ProjectresearchPythonOpen Jev-class decision models on Qwen3.5 with candidate-logit readout; the 4B model reports 80.1% on JevBench's 231 items.
ProjectresearchPython