Long horizons
The failures that matter appear hours into a task, not in the first reply. Our evaluations run long enough to find them.
Sonata LabsLong-horizon evals
for frontier models
for frontier models
Real-world measures of alignment
Sonata Labs is a research and product lab building long-horizon evaluations for frontier models.
The failures that matter appear hours into a task. Here is how we look for them:
The whole trajectory
We score every step an agent takes, not only where it ends up. Reaching the right result by a path you could not trust is still a failure.
Who we build with
Companies building environments, evals and benchmarks, to make them for alignment‑specific tasks. And model builders, to measure alignment on real-world tasks over long horizons.