jev-systemone-lab
Stdlib Python dispatcher turning typed function signatures into one batched Jev request, with a browser harness against ground truth.
Overview
jev_demo is a dispatcher skeleton for TypeSafe System One plus a Playwright browser harness. The core maps Python type hints to question shapes (Literal becomes a Choice, list of Literal a set, bool a flag), builds one batched request for a spec and its functions, and returns a typed call. Free-form arguments such as ints, strings and dates never become questions; the function default stands. A no-match option is always added so the model is not forced to pick, and overall confidence is the weakest judgment rather than a product of all judgments.
The browser layer snapshots a live page into a state and an action space of visible, enabled controls tagged with refs, which become the Choice criteria, so the model cannot pick a control that is not on screen. On 12 hard near-miss test rooms with jev-1.13.0, the README reports 11 of 12 correct across scoped, whole-page, Chinese-language and 20-question variants at about 630 ms p50, and notes fan-out to 20 questions cost almost no extra latency. A calibration script checks whether confidence is worth gating on.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which on-screen control (by ref) should be activated, or none? | — |
| Choice | Which value of each Literal-typed function argument applies? | — |
How it works
- Snapshot the page into state and a ref-tagged action space
- Build one batched request from typed function signatures plus a no-match option
- Jev returns choices; weakest judgment sets call confidence
- Execute or gate on confidence; compare against ground truth
Strengths
- Core uses only the standard library
- Illegal actions cannot be expressed
- README reports 20 questions cost about the same latency as one
- Includes stress and calibration scripts
Caveats
- Small test sets (20 easy, 12 hard rooms)
- Skeleton rather than a full agent
Getting started
python play_chrome.py -n 10 (headed), python stress.py --set hard, python calib.pyAt a glance
- Type
- Project
- Status
- prototype
- Language
- Python
- Runtime
- CLI / Playwright
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
TypeSafe System One (Jev) dispatcher skeleton: typed signatures to batched…
wh000wh000/awesome-jev-live
Frequently asked questions
How does it avoid invalid actions?
The Choice criteria are refs of controls actually visible on the page, so nothing else can be selected.
What happens to free-text arguments?
They are never asked; Jev only selects closed-set values and defaults stand for the rest.
How accurate was it?
The README reports 11 of 12 on hard near-miss rooms with jev-1.13.0.
Related projects
- SkillRanker★ 134
Rust CLI that ranks which agent skills fit the next step from live session context with Jev, can abstain, and hooks into Claude Code.
CLIbetaRust Open 2B non-autoregressive decision model on Qwen3.5-2B for agent tool routing and DOM actions, with BF16, FP8 and INT8 weights.
ModelresearchPythonOpen infrastructure for training, evaluating and deploying System 1 decision models, with calibrated LoRA checkpoints on a 27B backbone.
ModelresearchPythonOpen stack to synthesise decision data, train and calibrate a Qwen3.5-0.8B scorer, evaluate on Mind2Web and serve Jev-style.
ServiceresearchPythonRAM-efficient Rust coding-agent harness, listed as using Jev decisions for memory pruning, browser navigation and voice routing.
ServiceproductionRustMCP server and libraries that drive Chrome over CDP and hand a decision model like Jev a bounded menu of page actions.
ProjectproductionTypeScript