Playground
CommunityJev Demos & Experiments
Jev Determinism and Abstraction
A browser playground that turns prompts into abstract art through Jev's structured decisions, updating only the relevant fields of an 80+ characteristic renderer spec as you type, while comparing Jev with a single-call LLM, sandboxed custom JavaScript, and prompt-to-image generation.
- Stars
- 0
- Format
- Playground
- Reading time
- 2 min
- Published by
- Community
- Added
- Nov 2026
Highlights
- Runs four generation paths: Jev, a single-call LLM, sandboxed custom JavaScript, and AI image generation.
- Jev selects relevant fields of a renderer spec with more than 80 characteristics while you type.
- Score characteristics return probabilities over ordered tiers; relevance scores below 1.5 are omitted.
- Built with Jev, canvas-sketch, Vercel AI Gateway, and Vercel Sandbox.
- Local development runs with npm install and npm run dev after setting AI_GATEWAY_API_KEY.
Questions about this entry
Short answers drawn from the linked project or page.
Run npm install, copy .env.example to .env.local, and set AI_GATEWAY_API_KEY in that file, then start it with npm run dev. For local Sandbox previews you also run vercel link and vercel env pull.
Jev, which updates only relevant renderer fields as you type and reroutes on Compose when needed; an LLM that fills the complete renderer spec in one call; custom JavaScript written by Claude Sonnet 4.6 and run in Vercel Sandbox; and direct AI image generation.
Choice characteristics use the highest-probability option. Score characteristics return probabilities over ordered tiers and the result is their weighted mean, so motion importance of 0.47 came from 0x.65 plus 1x.25 plus 2x.08 plus 3x.02.
For the prompt a quiet blue gradient, motion importance scored 0.47 out of 3, below the 1.5 cutoff, so the motion, tempo, and rhythm fields were omitted and the artwork stayed still. The displayed confidence was the largest tier probability, 65%.
That path sends the prompt directly to an image model and shows the generated pixels without canvas-sketch or the renderer spec. Defaults include GPT Image 2.5 Flare, Recraft V4.1 Flash, FLUX 3 Image, and Muse Image 1.0.
On Compose, Jev checks whether the fixed renderer spec can express the prompt. If the spec cannot, the app reroutes to the custom JavaScript path, where Claude Sonnet 4.6 writes a prompt-specific renderer that Vercel Sandbox validates and runs in an isolated iframe.
The repository documents two verification commands: npm test to run the tests and npm run build to produce a build, both of which are listed together in its Verify section.
More Jev resources like this
Browser demo that sends one text to TypeSafe Jev and lights up 200 yes/no judgments as a live heatmap in roughly 150-300 ms, with seven 200-question sets, a threshold slider, and zero npm dependencies.
Jev Demos & ExperimentsPlayground#demo#browser#javascript
Communitycharles-forsyth/jev-decision-wall
Minecraft survival bot built on Mineflayer that asks Jev four typed questions every 600ms — threat level, action, whether to eat, flee direction — plus a tactical goal-selection loop, with simulated emotions and persistent memory feeding each request.
Jev Demos & Experiments#demo#games#madewithjev
Communityakash-kamat/jev-craft
A Hugging Face Space by ikaraoha titled JEV-9B — typed decisions, duplicated from autotrust/jev-9b-decision-demo. It presents an interactive browser demo of a JEV-9B model producing typed decisions, and is served from Hugging Face Zero hardware.
Jev Demos & ExperimentsPlayground#demo#playground#madewithjev
Community
A Hugging Face Space by Tonic that pairs MiniCPM-OCR with a Jev checker and serves the app from a Docker repository. Visitors can open the hosted page to try the OCR-plus-check flow without setting anything up locally.
Jev Demos & ExperimentsPlayground#demo#madewithjev#playground
Community
JevPilot is a real-time 2D autonomous highway runner and developer telemetry lab that steers a car with Jev or Clef System One decision models, charting latency, confidence, and steering choices in a live dashboard.
Jev Demos & Experiments#demo#games#javascript
Communityvijayprasath99/jev-pilot
Browser and terminal demo that ranks emoji against a free-text query by sending one Jev score question per emoji, then filtering by a tunable score floor and limit. Includes a bench harness over saved queries.
Jev Demos & Experiments#madewithjev#demo#typescript
Communitynischal94/emoji-sift
From 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.