World Modeling

Every forecast you run through FutureSearch draws on a world model: a set of shared drivers discovered across thousands of forecasts, which the forecaster commits to one view of and applies consistently.

A forecasting agent answering one question at a time has nothing keeping its answers coherent. Mutually exclusive outcomes can sum above 100 percent. The same question can come back with a different number an hour later, purely from variance. Human forecasters who keep their probabilities coherent across related questions are measurably more accurate on each one, and prediction markets get this for free through arbitrage. The world-modeling pass is how our forecaster gets it.

How it works

After the research agents finish and before you see a result, the pass reconciles the answer against the drivers underneath related questions. If your question about OpenAI's revenue depends on the same assumptions as a question you asked last week about its IPO timing, both answers now come from one view of those assumptions rather than two independent guesses.

The structure is discovered rather than imposed. The forecaster works out the drivers underneath the questions it is actually asked, keeps what it learns, and redraws the graph as the world moves.

The pass is automatic and always on. There is no parameter to enable or disable it, in the SDK, the MCP tools, or the app.

What it changes

We measured it on BTF-3, our pastcasting benchmark of already-resolved questions, across nine different base agents, from Claude Opus 4.8 to Claude Opus 5 and GPT-5.6, driven through our ReAct harness or their vendors' agent SDKs. It improved all nine, by 0.0003 to 0.0031 Brier, with four of the gains individually significant and two of those at p below 0.001. The smallest gains, on the agents that were already most internally consistent, are within noise of zero.

Those margins are what the frontier looks like: a gain of 0.003 is roughly the gap between second and fourth place on our leaderboard. The averages also understate the effect, because most benchmark questions have no related questions to reconcile against and the pass leaves them alone. On the questions where it does act, the average gain is about 0.0048, roughly twice the current gap between our full system and the strongest single-model agent we have measured.

It also keeps repeated questions stable, which is usually the difference you notice first in practice.

What you see

While a forecast runs, if your question overlaps with others, you will see a line such as "Drawing on 3 of your past forecasts and 5 public forecasts". When it finishes, you get a short note on whether the world model upheld or adjusted the raw forecast, along with the list of forecasts it considered.

Privacy

Your forecasts are private by default and stay out of the shared world model.

Your own questions can draw on your own past forecasts, but those forecasts will never inform anyone else's results. To contribute one to the shared pool, publish it: the globe icon on a forecast's results creates a permanent public page, and published forecasts join the shared world model.

You can see the effect on any published forecast. Anthropic's IPO pricing date shows what the answer was weighed against, and sits next to whether Anthropic trades before OpenAI and when OpenAI's shares begin trading, three questions the pass has to keep coherent with each other. More at published forecasts.

Consistency as a signal

Because the pass records every forecast before and after it acts, the total probability mass it had to move is itself a measurement. It turns out to predict how accurate an agent is without waiting for a single question to resolve. The best forecasters agree with themselves has the full analysis and an interactive explorer of the repairs on BTF-3.

Related docs

• forecast: the operation the pass runs on.
• decision: forecasting the outcome under each alternative of a choice you control.
• Publish and share a forecast: how to contribute a forecast to the shared pool.
• Published forecasts: real forecasts we have made public, grouped by question type.
• The shared world model behind every forecast: the announcement, with the history and the full results.