Questions, answered.
A world model for science. Today it forecasts whether clinical trials will meet their main goal, seals each forecast the day it is made, and scores it when the trial's results are posted. Misses become questions about what science should test next.
About nine in ten drugs that enter human trials never reach patients. Each failure costs years, often hundreds of millions of dollars, and the time of the people who volunteered.
A model that can call trial results in advance, better than the odds, has learned something real about human biology that no single paper or lab holds. That is worth far more than any one forecast. It means money and patients can go to the treatments likely to work, and away from the ones likely to fail.
The same model can do something no forecaster does: say which experiment would teach it the most. Where it is unsure or wrong, the gap names the next best question to test. Run that experiment, update the model, forecast again. Each turn of the loop makes the next forecast sharper.
Protein structure had a public, scored forecasting contest for 25 years before AI solved it. Medicine has never had one. The goal is a model that chooses the most useful experiment in biology better than any lab can alone, and proves it in public, one sealed forecast at a time.
Randomized phase 2 and 3 trials of a treatment or prevention, from any sponsor, with a registered control arm and a main goal about benefit, that finished after September 2024 or finish by the end of 2027 and have not posted results.
Each night's new forecasts go into a file that is only ever added to, and a fingerprint of that file is recorded the same day outside our own servers. The forecast stays hidden until the trial reads out.
Then it is revealed with its file, so anyone can check the fingerprint and confirm it came before the result.
8,670 forecasts since Sep 29, 2026. New trials are sealed every night. See every trial.
Each probability is scored against the result (the Brier score), the probabilities are checked for calibration, and every score is compared with how often similar trials succeed. The model has skill only if it beats that baseline by more than chance.
It stays unresolved. It is never counted as a failure.
Something the model gets wrong or cannot pin down: a link nobody has measured, a number that rests on thin evidence, or studies that disagree. Each one comes with the evidence or experiment that would settle it. See the unknowns.
No. It is a working prototype of the whole loop on one question. Its predictions come from a model built from published studies; the experiments you can run on that page show how results would update the model.
No. Forecasts are research. They are not medical advice or investment advice.
Trial search runs in your browser and there are no accounts. Analytics leave out search terms; our host keeps ordinary server logs for a limited time. Privacy.