How models reach an answer
We study how models actually reach an answer, where that reasoning holds, and where it quietly falls apart.
Models

We think the hard part of intelligence is reliability, not capability. So we build systems, then spend most of our time trying to break them.
Our work runs across three areas. How models reason, how they hold up in production, and how we keep them aligned with what people actually want.
Time on evaluation
Research areas
Models in production
Open evaluations
We study how models actually reach an answer, where that reasoning holds, and where it quietly falls apart.
Models
We turn a working model into infrastructure that stays fast and reliable long after the paper ships.
Systems
We build the evaluations and safeguards that keep a system honest before it ever reaches a real user.
Alignment
We take the work out of the benchmark and run it on the messy problems it will actually face in production.
Field tests
Research that stops at a benchmark is easy to fake and hard to trust. Every track here is meant to leave something real behind, whether that is a model, an evaluation, or a system already run in production.
Models that ship, not just papers
Evaluations anyone can rerun
Failure modes found before users do
Systems that hold under real traffic
Reasoning you can trace and check
Benchmarks we keep honest over time
Findings we publish in full
Tools partners can build on