User repository on GitHub
화재 피난자 개인별 판단 엔진 — 타입 안전 판독(Bonsai 2 27B / TypeSafe Jev) + 소셜포스 이동, FDS-GPU 화재장 위에서 역할·구조·경로 결정 · Meteor Simulation
- Forks
- 2
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Demos & Experiments
TypeEvacSafe
Fire-evacuation simulation where every occupant gets a Jev-style typed decision engine: a local Ternary Bonsai 2 27B, Qwen3.5-4B or the TypeSafe Jev API reads choice probabilities, while movement runs on validated social-force physics.
- Stars
- 2
- Language
- Python
- Reading time
- 2 min
- Published by
- Community
- Author
- using76
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Runs locally
Highlights
- Each evacuee carries one typed decision engine: Ternary Bonsai 2 27B (local, 7.2 GB), Qwen3.5-4B (local, 4.5 GB) or TypeSafe Jev over the API.
- Per-person heads: mode, target_exit (1-8), rescue_target, rescue_feasible, pace and survive, with only the heads selected by mode executed.
- A failed rescue_feasible answer demotes the action to evacuate, so an 'I will rescue' and 'rescue infeasible' contradiction never executes.
- Reads option-letter logits instead of generating text: softmax over A-P tokens, zero generated tokens, one forward pass per decision, no parse failures.
- Hall demo: 30 x 30 m, 2 MW fire, 100 occupants plus 2 firefighters, 211 frames at 0.5 s intervals, Bonsai and Jev runs side by side.
Watch out
Copy reviewed against the Korean and English repo README on 2026-10-15; it is truncated before the installation section, so setup beyond model sizes is not covered. The license is PolyForm Noncommercial 1.0.0, restricting commercial use, and the project is pre-release ahead of integration into Meteor Simulation's BULC.
Questions about this entry
Short answers drawn from the linked project or page.
A fire-evacuation simulation that rebuilds Jev's typed decision format on local LLMs and gives every simulated occupant their own decision engine, so people react to smoke, fire and crowding instead of always taking the shortest path.
It reads the next-token logits at the answer position, keeps the logits of the option letters A through P and takes a softmax, so each decision costs one forward pass, produces zero generated tokens and cannot fail parsing.
Ternary Bonsai 2 27B (local, 7.2 GB), Qwen3.5-4B (local, 4.5 GB) and TypeSafe Jev over the API; the demo video in the README compares Bonsai and Jev on the same hall fire.
The decision is split into heads, and only the heads selected by the mode run. If rescue_feasible answers no, the action is demoted to evacuate, so a person cannot rescue while the engine says rescue is infeasible.
CSV files with every person's coordinates and state at 0.5-second intervals, trajectory PNGs, MP4 videos and behavior-evaluation metrics, according to the README's feature table.
The local engines run on one GPU without an API, which the README says is required because building floor plans and occupancy data should not leave the machine. TypeSafe Jev is the API option for comparison.
It is a working research project being prepared for integration into Meteor Simulation's BULC fire-simulation evacuation feature, with a patent application in preparation; the license is PolyForm Noncommercial 1.0.0, so commercial use is restricted.
More Jev resources like this
Browser-only Laya demo suite: four Vite apps — a Traditional Chinese decision playground, Gomoku, Big Two and a 3D maze — run the Laya ONNX model on-device through Transformers.js, downloading fp16 weights from Hugging Face on first use.
Jev Demos & Experiments#community#catalog-import#madewithjev
Community
killkli/open-jev-layaPython
killkli/open-jev-layaPythonJev plays Minesweeper against the full wiring of an adult fruit fly brain: both see only revealed clues on the same seeded 9×9 boards, Jev judging every covered square safe or mined. Over 16 held-out boards the fly averaged 59.8 safe squares and Jev 59.1, against 52.6 for random clicks.
Jev Demos & Experiments#community#catalog-import#madewithjev
Community
enesyilmazcode/jevminesweeperPython
enesyilmazcode/jevminesweeperPythonA local proof of concept where Jev picks which of two conflicting signal phases gets the green light at a real Hong Kong intersection, using near real-time keyless open-data traffic sensors. A phase-timing state machine enforces minimum and maximum green, amber and all-red around the model's answer.
Jev Demos & Experiments#community#catalog-import#logicrw
Local macOS computer-use automation on Apple silicon using on-device laya typed decisions and CUA-S1-FORMS for filling, via the Accessibility API.
Jev Demos & Experiments#community#browser#local
Community
FluidInference/FluidUseSwift
FluidInference/FluidUseSwiftAn agent skill that reproduces Jev-style single-step logits over indexed browser controls using local Qwen weights for offline browser use.
Jev Demos & Experiments#community#browser#open-models
Community
APUS-AI-Lab/fast-browser-usePython
APUS-AI-Lab/fast-browser-usePythonA local app that turns a worded query into a physics pile of floating YC logos: retrieval over 6,241 companies by meaning, logo image, OCR'd logo text, and Jev re-applied tags, with Jev scoring the finalists into calibrated probabilities.
Jev Demos & Experiments#community#catalog-import#madewithjev
Community
aayan-dev/aayans-yc-indexorHTML
aayan-dev/aayans-yc-indexorHTMLFrom 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.