Think of the best storyteller you know. Someone who remembers every detail, adjusts on the fly, and never breaks the spell no matter what you throw at them. Storyteller is trying to be that.
When you start a session, a voice called the Narrator takes over. It describes the world, plays the characters, and reacts to whatever you say. Explore every corner of a room, and the descriptions get richer. The story bends around you without any special effort on your part.
The person who builds a story is called an Author. Authors use a separate set of tools — the Create Engine — to write stories, design characters, and set the rules. When they're done, everything packages into a single file called a .tome (or StoryPak).
A StoryPak runs on the Play Engine the way an app runs on a phone. Stories can't touch each other. One engine, many stories.
Describes the world, plays every character, reacts to what you say in real time.
Sets you up before a story starts. Handles settings, onboarding, and coming back after a break.
Manages all music and sound effects in real time — reacting to the story as it unfolds.
Remembers everything across every session you've ever played. Nothing is forgotten.
Runs all of them, second by second. The foundation every story is built on.
Before any of that can exist, one thing has to be proven: that an AI can run a real game correctly, in real time.
We're using a tabletop role-playing game called Open Adventure as our test case. Think Dungeons & Dragons. It's free and open-source, so there are no licensing hurdles.
It has real rules: combat, spellcasting, character abilities, a 240-term glossary where every mechanic is named and defined. That specificity is exactly what makes it useful as a proving ground.
A tabletop RPG has clear right answers. If the AI miscalculates a combat roll, you can point to exactly where it went wrong. Clear that bar, and the engine can handle any story system we throw at it later.
The output looks fine. The rules underneath are off. By the time you notice something strange, you're six steps from where the error started.
So before the AI ever sees the rulebook, we're cleaning it up. We converted all 16 chapters into structured files and built a machine-readable index of all 445 game terms: what each one means, and how each one connects to the others. That index is the term registry.
The registry works like a map of the entire game system. When the AI needs to resolve a rule, it walks the map — everything structured, nothing guessed.
445 terms, all written by hand, will have errors. Relationships listed in the wrong direction. Definitions with a missing mechanic. Fields left blank when the author moved on.
We built Sherlock to find them. It's an AI agent with one job: read the rulebook, check every term against the source text, and flag anything that doesn't hold up.
A human reviewer looks at every flag before anything changes. Nothing is applied automatically.
We process terms in batches, highest-priority first. The terms that connect to the most other terms go first — because a single error there ripples into every rule interaction that touches it.