Everything runs in your browser. Nothing you build is sent anywhere.
There is no template to pick, no example to copy and no form to fill in. The box is empty because the question is yours. You look at the world, you write what should be different, and the engine runs a few hundred versions of what happens next.
Every run starts a few hundred futures from the same state, each meeting a different sequence of shocks and a different draw of the coefficients nobody actually knows. The median is one line among many; the branches are named with the share of futures in each; and the same seed run twice is what makes a comparison your lever rather than the noise.
Not a loading message and not an empty form. Every country on file, drawn and coloured by income per head rather than by anybody's growth rate. Turn on what the question needs: satellite imagery, the states of a country, its counties, its towns, the chokepoints, the trade flow, the live emergencies. Anything can be on at the same time as anything else, in the flat atlas, on the globe, or in the network view where the geography drops away and what is left is who trades with whom.
One box, one sentence, ordinary words. There is no placeholder in it and no list of suggestions under it, because a prompt put there by the product is a template wearing different clothes. Whatever you write is read into the model's own levers and weighed into the ways it could actually go, each with a probability, before anything runs. Nothing is laid out for you to check first; afterwards, Futures shows every way that was weighed and how it ended. If what you want to ask about is already happening, click it on the map instead and the app offers you the sentence.
Once with your change and once without it, on the same seed, so both runs meet the same luck in the same years. A few hundred futures each, stepped a year at a time, in your browser. The futures with your change are split between the ways it could go in their shares, so the spread holds a war that stays limited beside one that widens. The difference between the two runs is your sentence rather than the noise.
A year at a time, or a month, a week or a day. At every point it says what moved across people, stability, resources, capability and the economy; why, in the engine's own named terms; what else could be happening instead and with what share of the futures; and what would change it.
What changed, ranked. The chain from your lever to the number, level by level. The spread of futures rather than a single line. And the consequences nobody named, which is usually the part worth the run.
The largest changes since the previous step, three from each part of the world rather than ten from the economy. A change to the world lands on people, on what they can burn and eat, on whether a place is stable and on what it can build, at least as much as it lands on a growth rate.
Which of your levers is acting at this moment and whether it is still phasing in, and underneath that the named terms in the decomposition that differ most from the baseline. Those are the arithmetic the engine ran, not a description of it.
The families of future the ensemble separated into, each with the share of futures inside it and what that family looks like at this exact point rather than at the end.
Twelve candidate changes, each added to your scenario at this step and the whole model run again on the same seed, ranked by what it does to whichever outcome you are judging on. It is the model's answer rather than a rule of thumb about it, and it says so.
The engine solves a year at a time, because that is the interval at which national accounts, censuses and energy balances are actually observed. A shorter step shows the same annual answer arriving rather than appearing: a price moves almost at once, output and employment follow over quarters, population takes the whole year. The interface says this on every sub-annual step.
A change that redistributes rather than adds shows no effect on a world total, and the probe says so instead of printing a zero. Migration intake is the clearest case.
A grammar of verbs and nouns, a place index that knows adjectival forms, and a set of number and date readers turn what you typed into levers with a country, a size, a start year and a duration. It runs in your browser with no key and no request. The same code then weighs the reading into the ways it could go: what kind of event it is, who the parties are (rival blocs, nuclear arms, oil, alliances), what the live feeds say about them and what the sentence itself said set a probability for each way. If you have connected a model provider, the model does the weighing instead, and whatever comes back is checked against the same schema before anything runs; if it is slow or answers badly, the browser's own weighing stands.
Nothing is shown before the run. Afterwards, Futures lays out every way that was weighed, why it got its probability and how it ended, and any lever can be opened and changed.
Energy, food and semiconductors each have supply, demand and a price. Supply answers the price on a lag, and a fab takes four years to build, which is the whole reason a chip shock hurts rather than clearing overnight.
Bilateral intensity from partner size over distance, adjusted for the agreements and frictions on file. Tariffs, export controls, sanctions and freight all bite through it, and governments answer in kind with a probability rather than automatically.
Productivity, labour, capital and energy set what an economy can make. Demand is written as a sum of named contributions. The gap between them drives prices and employment.
A Phillips curve for inflation, a Taylor rule for the policy rate, a currency that mean reverts and responds to rate differentials and risk, and Okun's law for unemployment with an explicit automation displacement term.
Primary balance, debt dynamics, a risk premium that rises with debt and sanctions, and a financing crisis hazard that can fire in any year.
Births, ageing, migration driven by income gaps and stress, political stability, tension between blocs, and conflict hazard that rises with tension and falls where nuclear weapons are involved.
Growth, inflation and unemployment are not solved and then decomposed afterwards. They are built as a sum of named terms, which means the bars you click are the arithmetic the engine actually ran. Trend, technology, interest rates, fiscal stance, external demand, energy and food, input constraints, conflict, the global cycle, confidence and the unmodelled shock. They add up to the number they explain, every time, and each one opens into what drove it.
Open any term and you get the underlying series, the country's position and the partners or prices responsible.
Random draws are addressed by scenario seed, future number, year, country and channel through a hash rather than pulled from a sequence. Two scenarios run on the same seed meet the same luck in the same years, so the difference between them is your lever and not the noise. That is what makes an A and B comparison meaningful at a few hundred futures instead of needing tens of thousands.
Import tariffs, export controls on chips, compute, energy, food, minerals or capital goods, sanctions packages and trade agreements. Any of them can be imposed by a coalition so the whole bloc acts at once.
Oil and gas production, energy capacity investment, food and minerals supply, carbon price, water stress and the warming path.
Policy rate shifts off the rule, the inflation target itself, currency devaluation and capital controls.
Fiscal impulse, corporate tax, strategic sector subsidies, military spending and debt restructuring.
Migration intake and fertility policy, with the twenty year lag to the labour force that makes the second one so slow.
AI capability steps, automation cost, fab capacity, productivity shifts, armed conflict and pandemics.
The engine is JavaScript in a worker thread. The state of every country is printed into the page itself, so the first load cannot be broken by a permalink rule, a security plugin, a caching layer or a firewall. Nothing you type and nothing you run is sent anywhere. A saved scenario is stored only if you ask for it, and only on this site.
The world draws immediately and the baseline run takes about two seconds.
Open the simulator