privatemode-decisions
Jev-style choice-plus-probabilities from GLM-5.3-Flash in one forward pass, in Edgeless Systems' confidential-computing API.
Overview
Privatemode Decisions makes an LLM choose from a fixed set of options and returns its choice plus a calibrated probability for each, from one forward pass and one token, without fine-tuning or parsing. It works like Jev but runs on GLM-5.3-Flash. On Privatemode AI, data stays encrypted while processed and the deployment's attestation is verified before anything is sent; any vLLM-backed OpenAI-compatible server also works, without the confidentiality guarantees.
It supports up to 1M tokens of context, images and up to 191 options per question, with calibration by default and fitting from a few labels. The README reports about 150 ms per decision including network, 0.894 on JevBench public items in its own run (versus 0.866 for Jev 1.13.0, noting official runs include sealed items on which systems score much lower), and on 29 labelled datasets a mean accuracy of 0.798 versus 0.775 for Jev with lower excess calibration error, though at higher cost per 1,000 decisions. The repository has a Python library and a web app.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which of up to 191 options fits the input? | — |
How it works
- Send state (text or images) and options.
- One masked forward pass reads option probabilities.
- Apply calibration (default or fitted).
- Use the choice and probabilities in your app.
Strengths
- Confidential computing with attestation.
- Image input and very long context.
- Reported better accuracy and calibration than Jev on its datasets.
Caveats
- Higher cost per decision than Jev per the README.
- JevBench figure is from its own unofficial run.
At a glance
- Type
- Service
- Status
- beta
- Language
- Python
- Runtime
- library, web app
- Has code
- yes
Use cases and integrations
How curated lists describe it
Gets a choice and a probability per option out of any OpenAI-compatible model in one forward pass, for Edgeless Systems' confidential-computing API or any endpoint you set.
MrJev/awesome-jev
Frequently asked questions
How is it confidential?
On Privatemode AI, data stays encrypted during processing and attestation is verified first.
Can I use my own server?
Yes, any vLLM-backed OpenAI-compatible server, without confidentiality guarantees.
How does it compare with Jev?
The README reports higher accuracy and lower calibration error on 29 datasets, at higher cost.
Related projects
Independent Jev-style server on DiffusionGemma and SGLang that classifies doodles, flowers and messages with Choice, Score and Noul.
ServiceresearchPythonOpen Jev-style text-and-image decision models built by a self-improving AI research loop, with an effort setting that picks depth.
ServiceresearchPythonSemIf (formerly OpenJev): reads typed option probabilities directly from open models like Qwen3.5-4B, including image inputs, on a home GPU.
ModelresearchPython- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
ProjectbetaPython