A world model for science.
OpenQuestion reads every trial, paper and dataset and forecasts what upcoming clinical trials will show. Each forecast is sealed the day it is made and revealed when the trial reads out, so the record can't be edited. Every miss becomes a question about what science should test next.
- Forecast clock
- Runningsince Sep 29, 2026
- Forecasts sealed
- 8,670as of Sep 29, 2026
- Resolved and scored
- 0
- Skill versus base rate
- Not yetnothing to score until trials resolve
Reading out soon
All 964 upcoming readouts →Real company-sponsored, randomized phase 2 and 3 trials finishing in the next four months (274 in all). Each eligible trial gets a sealed forecast; a new one is sealed the night after it appears.
- Sep 29, 2026Sealed Sep 29, 2026
- Sep 30, 2026Sealed Sep 29, 2026
- Sep 30, 2026Sealed Sep 29, 2026
- Sep 30, 2026Not forecastNo arm registered as a control
- Sep 30, 2026Not forecastNot a treatment or prevention trial
- Sep 30, 2026Sealed Sep 29, 2026
- Sep 30, 2026Sealed Sep 29, 2026
- Sep 30, 2026Sealed Sep 29, 2026
What goes wrong in science today, and what a world model changes
| Step | Today | With a world model |
|---|---|---|
| Predicting what works | Most drugs that reach human trials fail, often because the biology was misunderstood. | Forecast every trial before it reads out, and keep score in public. |
| Reading the literature | Over a million biomedical papers a year. Nobody can read them all. | Every claim, number and method from every paper, trial and dataset, read and linked. |
| Knowing what's known | Each lab keeps its own mental model of a disease. | One shared model per disease, every number pooled from all studies, with honest uncertainty. |
| Picking the next experiment | Driven by what a lab can already do, what gets funded and what gets published. | Ranked by how much each experiment would teach, per dollar. |
See all eight
| Finding what's unknown | Spotted informally, one scientist at a time. | Found automatically: links nobody has measured, numbers poorly pinned down, studies that conflict. |
| Running the experiment | Every lab has its own recipe, so results rarely combine. | Standard, machine-readable protocols that any qualified lab can run. |
| Updating what we know | Years, through papers, reviews and textbooks. | The model updates within minutes of a result and re-ranks what to test next. |
| Negative results | Mostly never published, so failures get repeated. | Every result counts. The model needs “ruled out” as much as “found”. |
How the loop works
The loop, explained with a loaf of bread →- 1 · ForecastPredict the trial
From everything known: earlier trials of the drug and its class, the biology of its target, and how often trials like it succeed.
- 2 · SealPublish a fingerprint
The forecast's fingerprint is published the day it is made. The forecast itself stays hidden until the readout.
- 3 · Reveal and scoreCheck it against reality
When results are posted, the forecast is revealed and scored against simple baselines. Hits and misses both stay on the record.
- 4 · LearnTurn misses into questions
A miss means the model is wrong somewhere. It becomes an unknown, with the evidence or experiment that would fix it.
What the misses reveal
All unknowns →When a forecast misses, the model looks for what it got wrong: a link nobody has measured, a number it could only guess, studies that disagree. Here is the full loop on one question, from a failed prediction to the experiments that would fix it.
Why does fat build up in the liver in type 2 diabetes?
Forecasts are research. They are not medical advice or investment advice.