Explorations
Earlier and side work: games, stories, history and tools we built along the way. Each one says when it was made and what it does and does not show. Our current work is in the Factory and the Research Worlds. These earlier results keep the scored comparisons from these studies.
Games and stories
Games and stories are worlds with state, actions and consequences. A world model can learn how such a world moves, then judge a move or a scene by what it expects to happen next.
Strategy game
Mandate of Heaven
Run a new AI lab from January 2020 through six years of real AI history. Each quarter you choose a move; a trained world model judges what your lab becomes.
The model learnt from 10,066 events across the AI industry. The game's own rules set money, costs and odds.
Technology history
Technology history scenarios
Pick a moment in technological history, describe a technology becoming available, and follow one path the model expects over the next 25 years.
One plausible chain of consequences, not a forecast of what did happen.
Fiction
Torus
A world model of a 100-scene manuscript. Choose a scene and ask what may happen next or what to watch.
Review access. In the held-out tail, the top three questions captured 11 of 13 material changes; a sequence-ridge control matched that recall with lower mean error.
Fictional-world tests
Star Wars and Middle-earth
Fresh event worlds that test the method when GPT cannot simply remember the future.
Across eight sampled comparisons, GPT-5 mini led none, the world model three and a current-state control five. See the comparisons.
History
Political history
Bismarck
Inspect hidden-history forecasts, then propose an alternate move for GPT to interpret using the saved model readout.
A hidden-history forecasting test across diplomacy, alliances and alternate choices.
Macroeconomy
Macro History
Pick a date in the public macro record, write the recommendation an analyst could have made then, and see the pressures each move shifts.
5,339 records through June 9, 2026. A dated laboratory, not a current feed.
Company history
Enron
Start at one of three real moments in the Enron record, choose a proposed action, and see two reads of the same fixed evidence.
The record stops at the chosen day.
Models and tools
Model catalog
World Model Atlas
Every evaluated model line and data project, with its own evidence date.
Review access. A catalog update does not mean every model was evaluated again.
Agent routing
World Router
A Codex skill that follows the state of the work and decides which model or group of models takes the next step.
Review access.
Agent teams
Organization Lab
Compare ways of organising the people who watch, fix and manage a company's agents.
Preferences from model and synthetic-inbox tests, not measured business outcomes.
Drug development
Project Confirm
A small world model predicted what happened next in FDA accelerated-approval programs it had not seen during training.
Review access.
Clinical trials
Clinical-trial scenarios
Historical conditional trial forecasts from a previous model family.
Review access. Supplied registry edits change the model response; these are not measured intervention effects.