Safe exploration
Agents learn in simulation and only take actions that can be undone. On this site, the gridworld runs entirely in your browser and cannot affect anything else.
Approach
How we would build it responsibly.
Four principles for a fictional lab, each with a concrete note on how this website itself applies them, so the claims can be checked.
Principles
Honest evaluation
Results are measured, not asserted. The playground reports how often the agent actually reached the goal, and this site's security is checked by recorded automated tests.
Human oversight
People approve consequential changes. The AI agents that helped deploy this site could not change its infrastructure without a person typing an approval.
Accessible and private
Every page works with a keyboard, has text alternatives for images and respects reduced-motion settings. There are no analytics, cookies or third-party scripts.