User repository on GitHub
using the new model JEV to play the game tetris
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Agent-JEV-Tetris
Tetris where Jev picks every move from the seven legal options, but the game pre-computes landings, holes, roughness and a best plan so the model only ever judges outcomes.
- Stars
- 2
- Language
- HTML
- Published by
- Community
- Author
- Yasserbhb
- Last verified
- Oct 2026
About this project
Tetris where every single move is chosen by TypeSafe's jev model. The game never decides anything. Each turn it sends the board state and a choice question over the seven legal moves, and plays back whatever jev picks. 363 pieces in, 114 lines cleared, zero holes. The panel on the right is the exact JSON sent to jev each turn, and its answer with the probability it gave every legal move.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
Highlights
- Best run so far: 14,300 points, 114 lines, 363 pieces and zero holes, at gravity of 300 ms per row.
- That run cost 1,713 requests and roughly 4.7M tokens, about 4.7 requests per piece, median response around 310 ms.
- One bestPlan across every rotation and both directions fixed the ping-pong that separate per-direction hints caused.
- Gravity runs on its own clock, so a slow request cannot cheat the fall, and the panel shows the exact JSON sent.
Watch out
The strategic ceiling is stated plainly: there is no memory between requests and no search beyond the current piece, so it cannot plan a setup or hold a well open across several pieces on purpose, and everything strategic it appears to do comes from which pre-computed facts it is shown. The prompt design is shaped by documented model behaviour - the jev-1.13 jaggedness notes say it is unreliable at counting, grids and arithmetic and that accuracy drops as irrelevant state grows - so the game measures every consequence in code. Only reachable targets are advertised: rows sealed under blocks are filtered out and each open dip reports which piece kinds still fit by simulation. Notable fixes were zero-slide placements, which the original lookahead made invisible, and landing depth rather than board height, which had silently punished filling a valley. Node 18+ and your own key are required, and no dollar cost is given for the 4.7M-token run. 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.
Judging pre-computed outcomes. Each of the seven legal moves arrives as a small JSON object: what it does, then_lands with landing columns, row, lines cleared, new holes and roughness after, and this_starts_the_best_plan describing the best placement found for this piece and whether the move is its first step. The arithmetic, grids and lookahead all happen in code, so the model never counts anything.
That is the trade the author made, and the README gives the volume without a price. The run took 1,713 requests, about 4.7 requests per piece at roughly 2,700 tokens each, with a median response around 310 ms. So per-piece cost is real but the payoff is measurable: 14,300 points, 114 lines and zero holes. You would be extrapolating the dollar figure yourself.
It survives without luck but does not strategise. Zero holes in 363 pieces is the author reading it as refusing to seal squares it cannot get back, rather than getting lucky. But with no memory between requests and no search beyond the current piece, it cannot plan a setup or deliberately hold a well across several pieces.
The key stays server-side: the page talks to a local /move endpoint that proxies to the API, so the key never reaches the browser. And node test.mjs runs assertions for the rules and for every number shown to jev, so the board measurements can be checked without spending anything.
bestPlan is a single best placement found across every rotation and both directions, attached to each option as this_starts_the_best_plan. An earlier version gave a head-this-way hint to both left and right; they pointed opposite ways and the piece ping-ponged in place until it landed, so one plan replaced them.
So latency cannot change the game. Gravity runs on a wall clock independent of the API, which gives Jev several actions per row of fall the way a human player gets, and means a slow request cannot cheat the fall or stall the piece in mid-air while waiting for an answer.
Because tallest_column_after never moves when you fill a valley, so it silently punished the right play. The author replaced it with lands_on_row, the row the piece actually lands on, which rewards filling a low gap instead of treating the deepest column as a standing cost.
More Jev resources like this
40GitHub stars
Python decision client exposing Choice, Noul, and Score primitives with isolated order-invariant scoring, a deliberative flywheel that crystallizes hard cases into fast paths, and local or hybrid deployment.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
zhangcy122/OpenJevPython
zhangcy122/OpenJevPython0GitHub stars
A benchmark for browser-using agents hosted on Almond: one prompt, five timed events scored by the page, plus Almond-fastloop, a ~200-line Chrome DevTools loop where Jev picks one bounded action per tick.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
eriestra/browser-use-olympicsHTML
eriestra/browser-use-olympicsHTML0GitHub stars
A FastAPI app that turns a text prompt into 16x16 pixel art by asking Jev 256 parallel Choice questions, one per pixel over a 16-colour palette, then refining the grid.
Jev Demos & Experiments#community#jevbest#catalog-import
46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
941GitHub stars
Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
TypeLLM/TypeLLMPython
TypeLLM/TypeLLMPython98GitHub stars
Self-hosted Docker service and web UI that packages Laya structured-decision inference behind a compatible System One endpoint.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
1Panel-dev/laya-serverPython
1Panel-dev/laya-serverPythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.