Fiction
A story is a world. Characters know some things and not others. Relationships change. Pressure accumulates. Actions close one future and open another.
A world model can ask:
- What is true at this point in the story?
- What is each character likely to do next?
- Which tensions are building across scenes?
- Where has the story contradicted its own world?
- How does a proposed scene change the futures that follow it?
Current work: Torus
Torus is a world model of a 100-scene manuscript. Choose a scene and ask what may happen next or what to watch. The model runs in the browser, reads four visible scenes, and projects the next four across 97 possible starting points.
It turns a draft into something an author can question: not just a searchable manuscript, but a model of how the story is moving. In its held-out tail, the model's top three questions captured 11 of 13 material changes across eight anchors. A sequence-ridge control matched that recall and had lower mean error.
Open the Torus Story Atlas — review access
Fictional-world stress tests
We also built fresh Star Wars and Middle-earth event worlds to test whether the machinery works when GPT cannot simply remember the future. Across eight sampled comparisons, GPT-5 mini led none. The world model led three and a current-state control led five.
That second result matters. Beating GPT is not enough. A world model should also beat the simple answer: assume the world stays close to its current state.
See the fictional-world comparisons
Where this goes
The aim is a model an author, studio or game team can inhabit with their own agents: one that remembers the whole canon, understands the state of the world at any scene, and can explore what a change would set in motion.