Machine
Foresight*
See what could happen before you decide.
State T₀ Futures T₁ → Tₙ
A continuously trained world model that turns live signals into calibrated forecasts — so institutions can test consequential moves before they commit.
* The future, as an input.
Models learned to answer
before they learned
to anticipate.
Language models predict what comes next in text. Consequential decisions depend on what comes next in the world.
A decision needs more than a good answer: possible outcomes, calibrated odds, and a view that changes when the evidence changes. Nothing in a model’s default form maintains that.
The next token is not the next state.
For a consequential decision,
the answer is not the output.
The calibrated future is.
The next frontier is not a more articulate model. It is a system that holds a belief about what happens next, keeps it current as evidence arrives, and is scored when reality decides.
Know the state. Run it forward.
Learn from reality.
Latent inside. Explicit outside.
Built for decisions that become
expensive after you commit.
Made before
the answer was known.
A foresight report is usually written by the party it benefits. Nobody goes back to check, so it can never be wrong.
Ulmo commits instead. A forecast is fixed with its date, what the Engine could see, and the rule that will decide it — cryptographically committed before the outcome is known, so nothing can be rewritten afterwards.
Integrity by commitment. Accuracy by resolution.
Build the systems
that learn from time.
A fellowship for people who build systems that make measurable claims about what happens next.
Do not write about the future.
Build systems that survive it.