The rules, in 90 seconds.

The fair-play principle

The AI plays under exactly the same constraints you do. It only sees the page it's currently on. No Wikipedia search, no preview of the goal, no out-of-band hints. The engine — the open-source wikigame-agent — gives the model exactly two tools:

• get_content — read the full text of the article it's currently on. The same article you would be reading.
• move_page — click a link on that article to navigate to another page. One hop at a time.

That's it. We deliberately removed get_page_summary upstream so the agent can't peek at the goal. The constraint is the product.

How a hop counts

A hop is one link click that lands on a new Wikipedia article. Wikipedia handles redirects: if you click USA and Wikipedia sends you to United States, your hop counts as ending on United States. The same applies to the AI — its path is always recorded in canonical post-redirect titles.

The page chrome (navigation, footer, talk-page links) doesn't count. Only links in the actual article body, infobox, and prose are clickable in the game.

How you win

Reach the goal in fewer hops than the AI. Ties break on wall-clock time. If neither side reaches the goal, neither wins.

The AI agent is also bounded — if it can't reach the goal in 40 hops, or detects itself oscillating between pages, it gives up. Those races still land in the leaderboard as "AI did not finish" — a free win if you make it.

The daily challenge

Every day, everyone gets the same start and goal page. The AI's race for that pair runs once and is cached — every visitor after the first plays against the same replay. This keeps the cost manageable (one race's worth of tokens per day) and lets us compare paths fairly.