Featured public forecasts
Live from the system — real questions, real headline numbers. Each links to the full ensemble report: every analyst’s case, the red-team attacks, and the math.
How a forecast is made
Ask anything resolvable
Pose the question in plain language. An analyst pins down precise resolution criteria first — vague questions make dishonest forecasts.
Independent panels research
Multiple frontier models, run as independent forecasters, search live primary sources — filings, government data, markets, the news of the last hour. Fresh evidence beats stale training data.
Adversarial review
Red teams attack each case. Judges score what survives. A forecast that can’t defend itself doesn’t ship.
Deterministic aggregation
Panel views are pooled with transparent math, and every probability is recomputed and validated outside the models. If a headline number diverges from its own tables, it is rejected — not rounded.
Four kinds of questions
A single calibrated probability of YES, with a confidence band and the evidence that moves it.
A full distribution over dates — median, P10–P90 band, and milestone probabilities for the checkpoints you care about.
A probability for every outcome on a well-posed option set, including the scenarios nobody proposed.
A full distribution over a quantity — median, P10–P90 band, and the probability of clearing each threshold.
What makes it trustworthy
Forecasters cite what they read, and they read the live web — not last year’s memories. Post-cutoff evidence is the point, not a bug.
Panels are built to disagree: different models, independent runs, opposing briefs. The spread between them is reported, not averaged away.
Every aggregation step is deterministic and shown in the report. The headline number is recomputed from its own tables — or the forecast is thrown out.