User repository on GitHub
Toy local System One–style decision API (Jev-shaped). Not affiliated with TypeSafe.
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Tools & CLIs
Systemone Lite
A local, trainable implementation of the System One decision contract on a 0.5B model: you send state plus closed questions and get one discrete answer per question by scoring option tokens instead of generating JSON.
- Stars
- 3
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- fritzprix
- Last verified
- Oct 2026
About this project
Local typed decisions on a 0.5B model: you pass state + closed questions (yes/no, choice, score); it returns one discrete answer per question by scoring option tokens — not by writing JSON. Compatible wire shape with the public System One API. Independent project — not TypeSafe / Jev, and not a cloud drop-in. Bare-face self-play GIF — entertainment only. Held-out JSON is the claim surface. Weights: dwidlee/systemone-lite-0.5b · Data: dwidlee/systemone-lite-phase2 · Consumer GPU (numbers below: RTX 3060)
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Returns one discrete answer per question by scoring option tokens, so nothing has to be generated and parsed as JSON.
- Speaks the public System One wire shape over POST /v1/systemone, with a Python client and choice, noul and score helpers.
- Held-out numbers ship with sample sizes and standard errors: 61.6% on the phase2 test split, 50.7% on a 231-item JevBench public set.
- The README states it is an independent project, not TypeSafe or Jev, and not a cloud drop-in.
Watch out
The README itself calls the model 0.5B and says calibration is mediocre, so do not read confidence as a calibrated probability. Accuracy also drops on the full test split (chess around 42%, sokoban 38%), and uniform random on that JevBench set is about 32%, so the headroom is small. Only closed option sets are supported: aliases are ranked and JSON keys are mapped after scoring. It is not TypeSafe or Jev and no hosted endpoint is involved. 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.
They are the repo's own runs, printed with sample sizes and standard errors next to them: 4,700 held-out examples for the phase2 split and 231 items for the JevBench public set, where random guessing is about 32% rather than 50%. The README also says to treat small deltas as noise, and notes this is a published Hub revision rather than an official leaderboard row.
No. The request and response bodies match the public System One API, so the shape looks familiar, but the README says plainly that this is an independent project, not TypeSafe, not Jev, and not a cloud drop-in. You run the server yourself and point your own client at it.
Not on its own terms. The limits section says the 0.5B size is for demos and local research rather than a production decision service, calibration is mediocre, and only closed option sets work, with keys mapped after scoring. It is built for experiments, demos and benchmarking decision heads without a hosted stack.
Use the Hub dataset's test split, never train, for any accuracy claim. Only closed option sets work: it ranks the aliases you supply and maps JSON keys afterwards, so a question without a fixed option list is out of scope. After a clone and pip install -e ".[dev]", start it with systemone-lite --model dwidlee/systemone-lite-0.5b --port 8000 and curl the bundled fixture request, or point SystemOneClient at the same model id.
MIT-licensed, with weights at dwidlee/systemone-lite-0.5b and a training dataset at dwidlee/systemone-lite-phase2 on Hugging Face. It is an unofficial local approximation of the System One contract and is not affiliated with TypeSafe, Jev or any cloud drop-in.
It scores option tokens instead of writing JSON. The repo reports roughly 10-50x fewer tokens and lower latency than a greedy autoregressive JSON approach using the same weights, and a short one-question call runs in about 10-30 ms on the reference hardware.
They come from the action-v2-qwen revision measured on a consumer RTX 3060. The JevBench public set (231 items) shows 50.7% accuracy, ECE 0.245 and a p50 of 13.1 ms; the authors say to treat small deltas as noise since n=231 has an SE near 3 percentage points.
More Jev resources like this
34GitHub stars
Pure-HTTP automation for the TypeSafe register, magic-link login, onboarding, and API-key creation flow, with pluggable temp-mail backends, concurrent batch runs, and key verification exports.
Jev Tools & CLIs#community#awesomejev#catalog-import
Community
2951461586/Jev-Register-ToolPython
2951461586/Jev-Register-ToolPython36GitHub stars
Python toolkit plus 0.8B multimodal model that takes a message, document, or image and returns calibrated choice, yes/no, and score answers in one forward pass for local decision workloads.
Jev Tools & CLIs#community#awesomejev#catalog-import
Community
PsiACE/dohnutsPython
PsiACE/dohnutsPython39GitHub stars
Video tool where Jev scores subtitle segments and code assembles a short highlights cut plus a de-watered full cut with timestamped summaries.
Jev Tools & CLIs#community#jevbest#catalog-import
Community
cclank/jevclipPython
cclank/jevclipPython30GitHub stars
Forkable weekday radar that asks Jev one yes-or-no question per interest over each new arXiv announcement and publishes a page plus RSS feed.
Jev Tools & CLIs#community#jevbest#catalog-import
Community
Eliot5566/JEV-Paper-RadarPython
Eliot5566/JEV-Paper-RadarPython38GitHub stars
Async supervisory layer for Hermes Agent that locks Definitions of Done, proves mechanical criteria locally, and calls Jev only for unresolved semantic checks under a token budget.
Jev Tools & CLIs#community#jevbest#catalog-import
Community
keeltrace/hermes-nervePython
keeltrace/hermes-nervePython15GitHub stars
Two-tier decision engine pairing a 22.6M-parameter MiniLM encoder on the Apple Neural Engine with the on-device Apple Foundation Model as fallback for uncertain inputs.
Jev Tools & CLIs#community#awesomejev#catalog-import
Community
kerryrm/systemANEPython
kerryrm/systemANEPythonFrom the Jev guides
Original write-ups on what this entry is for.
The request shape barely changes across surfaces; the surrounding contract does. When to call the API directly and when a gateway, a Worker binding, or a router entry earns its extra hop.
Jev Tools & CLIs#gateway#vercel#cloudflare
Read article
Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.
Jev Tools & CLIs#open-models#local#onnx
Read article
Jev labels a record that already exists. Code moves, flags, or holds it. How the document, mail, and transfer tools keep the action out of the model.
Jev Tools & CLIs#classification#email#security
Read article
From the Jev community
Posts from builders shipping with Jev right now.