A world model makes the fork. Calibration makes it a signal.
Just-in-time world modeling is about generating the shape of the future when it is needed. The fork is necessary. The signal is earned.
A world model can simulate a state on demand; a forecast is a distribution that paid for itself. The distance between those two is the whole product — and the product is broken the moment the two are confused.
The recent push in just-in-time world modeling for planning makes a distinction worth keeping: a world model is a learned capacity to produce a plausible state of the world on demand, and “just in time” means it materializes that state when a decision asks for it, rather than maintaining an expensive representation continuously. That framing is useful because it splits a planning stack into two very different artifacts: the fork, and the signal. The fork is the set of futures that could be; the signal is the probability the fork carries. The fork without the signal is a story. The signal without the fork is a number with no story behind it.
The fork is what the sisters produce
DEINO’s sisters are, in one reading, world models: each personality imagines a plausible state of the future from the node versions it reads, and the ensemble of them is a fork — several coherent ways the world could go, each produced by a different lens (analyst, contrarian, disruptor, historian, institutionalist). That is genuinely useful on its own: a fork is how you explore what is conceivable. But a fork has no unit. It is a collection of futures, and a collection of futures is not yet a defensible probability about any of them.
The signal is what the oracle earns
The oracle’s merge is the step that converts a fork into a signal: the distribution is normalized in exactly one place, entropy in nats, scenarios sorted. That number earns its status only through the calibration loop — shadow runs against a realized timeline, scored by deino-eval in Brier and log-loss, gated by Gate 1b. The distinction is the product. A story can be wrong and still be a story. A signal has to be defensible, which means it must be a number that was tested against what actually happened. The fork is the world model. The signal is the forecast. The gap between them is where products die.
A world model is a fork. A calibrated forecast is a signal. The whole product is the distance between them.
- The fork is cheap and qualitative: simulate the plausible, get a good story about what might happen.
- The signal is expensive and quantitative: calibrate the fork, get a number with a unit.
- The fork can be wrong and remain useful — the signal, if it is wrong, must show it was wrong on the record.
Why we refuse the shortcut
It would be easier to ship the fork as the product: vivid scenarios, no calibration, no gate. We are pre-revenue, building the proof of concept, and the forecast-facing layer is gated precisely so the fork is not sold as a signal before the loop exists. The line is drawn in public: a simulated future is not a prediction, a cone is a distribution that earned its width, and the audit manifest seals how the fork became a signal — or does not. The refusal to collapse the two is what a regulator can stand on.
A future you can imagine is not a future you can defend. The imagination is the fork; the defense is the calibration; and nothing else is a signal.
— calibration note — the oracle