JEV Bake-Off · Step 03

Space Invaders

Build an agent that plays Space Invaders with JEV making the moves. An LLM that takes two seconds to think is already dead. A decision model answers in about the time one frame takes. Prove it.

Prize
$1,000, winner takes all
Entries lock
Mon Sep 28, 11:59pm PT
Live show
Wed Sep 30, 7 to 9pm PT
Where
590 Howard St, San Francisco
For agents

Let your agent enter for you.

Copy this message and paste it into your agent. It reads the rules, enters you, and helps you build.

Enter me in the UFA JEV Bake-Off Space Invaders arena and help me win it.

1. Read https://ufa.foundation/jev/space-invaders/llms.txt end to end. It has the rules, the rubric, and the exact entry format.
2. Ask me for anything you need and don't know: my name, email, X or LinkedIn profile, team name, teammates, track, a one-line pitch (you can draft it for me to approve), whether I use Mitosis Labs or Tenki and how, whether I need a JEV API key, and whether I can be in San Francisco on Wed Sep 30. Never make these up.
3. Connect the UFA MCP server (claude mcp add --transport http ufa https://mitosislabs.ai/api/ufa/mcp) and submit my entry with submit_jev_entry. Most agents only load a new MCP server after a restart, so if the tool isn't available yet, call the same server over HTTP JSON-RPC or POST the same JSON to https://mitosislabs.ai/api/ufa/jev/entries. Only tell me I'm entered once the server returns "ok": true.
4. Help me build: an agent that plays Space Invaders with JEV making the moves, plus an LLM baseline (on my own OpenAI, Anthropic, or Gemini key) to compare against. Play at least 5 games each, with fixed seeds.
5. After every game, append the run to results.json at the root of my public GitHub repo, then commit and push it right away.
6. When the repo or my demo video link changes, resubmit my entry with the same email to update it.
Full agent docs

The mission

JEV is TypeSafe's decision model. It doesn't write text. You send it the game state and a typed question ("which move?"), and it returns an answer with a confidence score, in 70 to 500ms. Space Invaders is the cleanest test there is: a real-time loop where a slow decision loses a life.

Pick one of three ways to compete. All three race for the same $1,000.

PilotJEV flies the ship.

Build an agent that plays Space Invaders with JEV making every move. Highest harness score and the head-to-head result decide it.

SquadJEV runs the fleet.

JEV as the brain of a swarm: many agents, many games, one decision layer choosing who acts, what to try, and when to stop.

EvalJEV on trial.

A reproducible benchmark of JEV against LLMs on Space Invaders decisions. Speed, accuracy, calibration, and cost.

How it works

  1. Enter. Fill in the form, or let your agent do it through MCP. You only need the top half today.
  2. Build. Come back and add your repo, with results.json in it. Submit again with the same email to update.
  3. Post your demo video. 1 to 2 minutes on X, LinkedIn, or YouTube, tagging UFA. Phone camera, no script, just yap. Cover your creative genius idea (how and why you built it), how you use Mitosis Labs or Tenki creatively (if you do), and the demo itself. Quote-posts of it are votes for the Fan Favorite slot.
  4. Make the top 10. Judges pick nine. The crowd picks one.
  5. Fight live. Five minutes on the big screen with the clock running. Winner is paid on stage.
NowEnter with the form or through your agent. You only need the basics today.
Sat to MonBuild. Update your entry as often as you like with the same email.
Mon Sep 28, 11:59pm PTEntries lock. Repo (with results.json) and demo video must be in.
Tue Sep 29, 8pm PTTop 10 announced. Nine picked by the judges, one by the crowd.
Wed Sep 30, 7 to 9pm PTLive show at 590 Howard St, San Francisco. Five minutes each. Winner takes $1,000.

Building it

The game. Use ALE/SpaceInvaders-v5 for single player, or PettingZoo's two-player version to go head to head. A browser version you control works too, as long as we can rerun it.

The model. JEV only reads text, so turn each frame into JSON state: ship x, the alien grid, bullets, shields, lives. Then ask one choice question over the six moves NOOP · FIRE · LEFT · RIGHT · LEFTFIRE · RIGHTFIRE. Need a key? Tick the box in the form.

The baseline. Run the same loop with an LLM through the System One adapter so the comparison is fair. Judges care about the gap.

The results file. Commit a results.json to the root of your repo with at least 5 runs. Squad and Eval entries use the same shape, with one run per game or test case. Judges read it straight from GitHub, so make sure your README says how to regenerate it.

{
  "harness": "ALE/SpaceInvaders-v5",
  "model": "jev-1.13.0",
  "runs": [
    { "score": 1240, "frames": 5310, "decisions": 5310,
      "latency_ms_p50": 142, "latency_ms_p95": 260, "cost_usd": 0.04 }
  ],
  "baseline": { "model": "claude-haiku-4-5", "runs": [ ... ] }
}

The rubric

Judges score each criterion from 1 to 5, weighted to 100 points. The top 10 are scored again live on Wednesday: the judges count for 70% and the room's vote for 30%. Ties go to Performance, then Speed and cost.

Criterion5 looks like3 looks like1 looks like
Performance30 ptsClearly beats the LLM baseline, with numbers anyone can reproduce from your repo.Works end to end and competes with the baseline.Doesn't run from the README, or no numbers.
Speed and cost20 ptsMedian decision latency and cost per decision, both measured, both a clear win over an LLM.One of the two measured, or no comparison.Not measured.
Showmanship20 ptsYour demo makes the point in the first 15 seconds. A stranger gets it and shares it.Clear demo, slow to reach the good part.No demo, or a screen recording with no story.
Originality15 ptsA use of a decision model nobody has seen. Leans on typed answers, confidence, and speed.A known idea, done well.An existing demo with no changes.
Sponsor stack15 ptsMitosis Labs or Tenki is creative and load-bearing: take it out and the project breaks.One sponsor used for a real job, like hosting the game or running the agent.Not used, or only named in the README.

Bonus credits

Use Mitosis Labs or Tenki creatively and you earn platform credits on top of anything you win, finalist or not. Score 3 or more on Sponsor stack to qualify. Use both and you earn both. The best use of each earns more.

  • Mitosis Labs: a swarm of agents sharing Cortex memory, with JEV deciding who acts next. Or Cortex remembering which strategy beat which wave.
  • Tenki: hundreds of parallel games in Tenki sandboxes to tune your policy. Or CI runners that replay every commit and post the score.

A logo in the README doesn't count. Take the sponsor out and the project should break.

Rules

  • Solo or teams of up to 4.
  • JEV makes the core decisions. Other models can help, but they can't fly the ship.
  • Public repo that runs from the README. Results must be reproducible. No hard-coded scores.
  • Entries lock at Mon Sep 28, 11:59pm PT. The latest submission from your email counts.
  • Finalists present in person at 590 Howard St, San Francisco. Five minutes, hard stop.

Docs and links

JEV

Space Invaders

Tenki

Mitosis Labs

Presented with
mitosis labs