Your move.
Play a decision, then compare with the model.
Four games. Sixteen saved situations.
Choose an action. See every probability.
Loading recorded decisions…
What would you do?
Select a move to reveal the model’s choice.
Your choice stays in this browser. This page reads saved outputs; it does not call a model.
The same input the model saw.
Real outputs. Open evidence.
These illustrations reconstruct synthetic game states from the audited v1.1 evaluation. Each situation is an independent recorded decision.
This situation: exact input and output
How to read this arcade
Choose one of the original action options, then compare it with the highest-probability action in the saved Open-Jev-27B-v1.1 output. Bars show the complete distribution in the original option order. They are probabilities over actions, not chances of winning.
The displayed boards are illustrations of supplied text states, not captured game viewports. No action is executed, and the situations do not form a continuous game. The selection includes disagreements with the supplied reference policies.
Snake uses a heuristic reference policy; tic-tac-toe uses exact minimax. The runner uses original box physics, and the platformer uses an original discrete tile simulator. Neither is a replay of a commercial game.
These sixteen examples do not establish a win rate or overall gameplay ability. The hosted classification workbench separately runs Open-Jev-2B on CPU.
What would you build?
Give Open-Jev a state, a question and a set of options.