What this proves
This sample game shows a tamper-evident public record: each side locks in its word before play, and anyone can re-check the receipt. It does not show whether anyone peeked or answered free-text questions honestly, and it is a demonstration, not a court or a certification.
Play a sample word game with your own AI
Atlas is the referee and the record keeper only. It does not think, guess, or pick words with a model. Your own AI plays. The word your AI commits can appear in its own transcript; do not peek. That part runs on trust.
Add a connector in your AI app, pointed at:
https://demo.atlasclose.com/guess/mcp
Then paste this line into your AI:
Use the Atlas Guess connector at https://demo.atlasclose.com/guess/mcp to start a new sample game. Give me the watch link, then lock in your own secret word (do not tell me the word). Keep going on your own until you have committed. I will play on the watch page. When the game ends, give me the receipt link.
This folder is a staged copy. It is not live on the public site. Nothing here has been deployed.
Moves on this page are accepted only from this page, with an origin check and a one-time code. If this is built this way, a visitor-authored entry would come from the page, and the AI would not be able to answer for the person. That is a design goal, not yet proven.
How a game works
- Your AI starts a game and gives you a watch link.
- On the watch page, you lock in your word. The hash is computed in your browser; the word and salt stay on this page until you choose to reveal.
- Your AI locks in its word through the connector. Atlas records the hash, not the word, until reveal.
- You play on the watch page. Your AI plays through the connector. Every move is written to a public docket.
- When both rounds end, anyone can open the receipt and re-check it.