Wyrd

Wyrd is a browser fantasy adventure run by an AI game master that can't cheat, and can prove it. The problem: AI agents are starting to act for us, but we can't check what they do. In AI story games this shows up directly: the AI can hand you a legendary sword out of nowhere, or quietly move the treasure when you get close. Wyrd fixes that with three layers: 1. The AI proposes, code decides. Meta's Muse Spark can only act through tool calls (move, attack, sneak, throw, attempt...). A TypeScript rule engine is the only thing that changes the game, and it refuses anything impossible, with a visible REFUSED and a reason. 2. Secrets are sealed before you move. The relic's room and the trapped door are chosen at the start and locked with a SHA-256 fingerprint shown on screen. At the end, your browser recomputes the fingerprint itself. If it matches, the AI never moved them. 3. Creative freedom with fair rules. Players can click, type anything, or speak (English or Tagalog). Unusual ideas like "I tip the shelf onto the troll" become a code-judged attempt: the AI only classifies the idea, and the engine sets the difficulty, rolls the dice and picks the outcome (success, partial, fail). An Inspector shows every tool call the model made and the engine's verdict, so anyone can see the rules working. The game is the demo; the pattern (AI proposes, code decides, a seal proves it) applies to any agent that issues refunds, books trips or runs game economies. Live: https://wyrd-web.onrender.com Honest scope: the sealed secrets are provable; the dice are server-rolled and shown with their targets but not yet independently verifiable.