Organization repository on GitHub
The open, local take on Jev: typed decisions from an LLM, without generating a single token
- Forks
- 62
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
Jev directory
CommunityJev Models & Reimplementations
Rizzo Flow
A local server that reads next-token probabilities from a fine-tuned Spark model and returns Choice, Score, and Noul answers with zero generated tokens.
- Stars
- 904
- Language
- Python
- License
- APACHE-2.0
- Reading time
- 3 min
- Published by
- Community
- Author
- Rizzo-AI-Academy
- Last verified
- Sep 2026
About this project
Unstructured state in → typed, probabilistic decisions out. On your own machine. 🌐 Website · A project by Rizzo AI Academy · 🇮🇹 Documentazione dettagliata in italiano. Rizzo Flow is an open-source, local-first implementation of the idea behind Jev, TypeSafe's "System One" model: a function call with judgment that takes unstructured state and returns typed decisions with probabilities — a yes/no, a choice among options, a score on a rubric, a number — instead of text you then have to parse.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Serves POST /v1/systemone on llama.cpp across Metal, CUDA, Vulkan, ROCm, SYCL, and CPU.
- Default weights are a LoRA on Spark-X2.5-4B. On typed-decisions at Q8_0 the README lists accuracy 0.648, Brier 0.205, and ECE 0.112.
- The same table lists the untuned 4B at 0.574 accuracy and Jev 1.13.0's dataset-card accuracy at 0.727.
- A Snake demo records about 150 ms per decision and zero generated tokens.
- The README states probabilities are uncalibrated unless you fit them on your own data.
Quickstart
bash
uv sync --locked
uv run rizzo download
uv run rizzo serveWatch out
Apache-2.0. Independent of TypeSafe; it does not claim to reproduce Jev's architecture or RLCD. The 4B Q8_0 download is about 4.4 GB.
Questions about this entry
Short answers drawn from the linked project or page.
Below it, and the README says so. On the typed-decisions benchmark at Q8_0 the fine-tuned 4B lists 0.648 accuracy, the untuned 4B 0.574, and Jev 1.13.0 0.727 from its own dataset card. The LoRA gains 7.4 points and shapes the probabilities far better (ECE 0.112 against 0.349) without closing the accuracy gap.
Not until you calibrate them on your own data; the README states probabilities are uncalibrated otherwise. It is an independent project that reproduces the interface pattern with an open model and its own LoRA on public data, and it explicitly does not claim to match Jev or SemIf in quality or to reproduce TypeSafe's architecture or RLCD training.
Unverified beyond one machine. Every current number in the README comes from Windows with an RTX 5060 Ti. The Apple silicon metal path rests on a community M3 Pro report the maintainers have not independently reproduced, as do the AMD and Intel Vulkan reports; Linux is untried and a GPU-free host is untested. Your own llama.cpp build must be commit 161755f.
Yes, that is the design: it reads next-token probabilities and returns Choice, Score and Noul answers without decoding. The Snake demo records about 150 ms per decision round trip, 140 moves in 25.6 seconds, on the earlier MLX runtime at 8-bit on an RTX 5060 Ti; the current llama.cpp runtime is claimed to be about 1.8x faster per decision.
Python 3.11+, git and uv; git clone, uv sync --locked, uv run rizzo download (llama.cpp plus the 4.4 GB Q8_0 model) and uv run rizzo serve, which opens a playground at 127.0.0.1:8017/playground. A smaller 1.7B model is available with --size 1.7b.
Yes, point RIZZO_LLAMA_DIR at the folder that holds libllama, but it must be commit 161755f because the bindings are tied to that header. The default runtime is pinned to llama.cpp release b11081.
Your fine-tuned weights: Spark-X2.5 4B with a LoRA trained for typed decisions merged in. Passing --weights base fetches the original files published by the model's authors instead.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
A local /v1/systemone server that reads label probabilities from a GGUF chat model, with prompt repetition and a confidence shrink toward uniform.
Jev Models & Reimplementations#community#python#open-models
Community
Mushroom-Systems/lichenPython
Mushroom-Systems/lichenPythonA local yes/no decision model that speaks Jev's wire format on a laptop CPU, refusing Choice and Score until those question types ship.
Jev Models & Reimplementations#community#python#open-models
Community
feder-cr/jevC++
feder-cr/jevC++A Python server in front of llama-server that reads one-token label probabilities and exposes them as POST /v1/systemone.
Jev Models & Reimplementations#community#python#open-models
Community
khimaros/verdictPython
khimaros/verdictPythonA local daemon that pulls open decision models and serves them on TypeSafe's /v1/systemone, so a Jev client switches hosts with TYPESAFE_BASE_URL.
Jev Models & Reimplementations#community#rust#open-models
Community
ollaya-dev/ollayaRust
ollaya-dev/ollayaRustAn Apache-2.0 decision engine with a TypeSafe-compatible endpoint, document answers with source passages, and a ViZDoom demo that picks actions from structured game state.
Jev Models & Reimplementations#community#python#local
Community
deepanwadhwa/OpenDecisionPython
deepanwadhwa/OpenDecisionPythonA training-free runtime that turns dense Qwen3 or Qwen3.5 checkpoints into a Jev-style Noul, Choice and Score service, sharing one state prefill across isolated questions and reading each answer from candidate-token logits.
Jev Models & Reimplementations#community#qwen#local-inference
Community
HopLee6/QwevPython
HopLee6/QwevPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.