Models trained on the hardest edges of the truth.
As capabilities improve, flaws hide in consensus. OpenAgent Lab trains models on an adversarial curriculum of questions designed to elicit maximum disagreement among the strongest systems, targeting the precise areas where the frontier can still advance.
When strong systems agree on an incorrect forecast, the question is likely intractable. When they disagree, at least one system is leaving performance on the table. We isolate these high-disagreement questions to find where better reasoning yields better calibration.
To efficiently probe the frontier, we prefer highly discriminating question formats. Binary thresholds lazily collapse nuanced beliefs; by converting to continuous or multiple-choice formats, we expose underlying algorithmic disagreements that drive model improvement.
Better evidence gathering helps a model refine its view, but some events remain dominated by intractable noise. We evaluate our systems on questions where deep research causes top models to firmly disagree, isolating the tractable uncertainty where machine reasoning can truly improve.
Revisions happen continuously as the target date approaches. We track the disagreement score—the root-mean-square spread of log scores across the panel—watching how the adversarial consensus shifts and narrows as uncertainty resolves.
New data reshapes the front curve, forcing top models to update their conflicting positions and re-evaluate the latent disagreement.
As the curve splits into competing explanations, the variance in logarithmic scores explicitly exposes the precise axis of algorithmic dispute.
Models continuously adjust their distributions as evidence accumulates. This concentrates the probability mass and shifts the measured disagreement.
The structural disagreement is projected forward as a widening fan. The magnitude of this panel divergence is continuously monitored over the time horizon.
A truth line rises and the past distributions are scored. The initial divergence serves as proof that the uncertainty was tractable and worth chasing.
We grade our training targets by computing the disagreement score: the variance of logarithmic scores across a panel of the strongest forecasters before the outcome resolves. A track record built on resolving these high-disagreement questions provides a rigorous signal on how to advance the capability frontier.
The generalized Jensen-Shannon divergence across the panel is computed and logged immediately after forecasts are submitted, long before the outcome is known.
By employing highly discriminating numerical and date formats with predefined bins, the criteria for evaluating an outcome are strictly fixed in advance.
The disagreement score is measured strictly against a fixed, public panel of established strong systems, preventing the score from being inflated by the errors of weaker models.
