j. JEV / REWARD ARCADE
OPEN EXPERIMENT GitHub ↗
REINFORCEMENT LEARNING, WITH A JUDGE

Let JEV judge.
Let the agent learn.

A tiny agent. A world to figure out. Every move becomes a lesson.
Watch a reward signal turn trial and error into a winning run.

OBSERVE→ACT→JUDGE→LEARN
One decision at a time. ↲
RULES DEMORecorded CPU training · deterministic judge · no JEV API callsREPLAY / SEED 7
01 / THE PLAYGROUND

Key Quest 9 × 7

TRAINED POLICY
STEP 00FIND THE KEY
Agent Key Exit LavaCollect the key. Avoid lava. Reach the exit.
04 / THE LEARNING CURVE

Practice leaves a trace.

Last 20 episodes · success
EXPLORATION ε 0.04350 EPISODES
05 / THE RESULT

Same world.
Better decisions.

BEFORE—
→
AFTER—

Success on 60 evaluation episodes, same seeds and map. Evaluation makes no judge calls.

UNDER THE HOOD

The judge supplies the reward. The agent learns the actions. Tabular Q-learning updates from the judge’s outcome probabilities. JEV receives structured state, including computed route distances; the agent sees only position and key possession. This is a small, inspectable experiment—not a pixel-based or general-game benchmark.

About JEV Score ↗