2026 · 04
Long-horizon reasoning without forgetting
A memory schedule that keeps week-scale context coherent while holding compute flat.
Research
Notes from the long night
We publish what we learn — including the parts that did not work.
Papers & notes
Selected work.
2026 · 02
Citations as a first-class output
Training signals that make models prefer verifiable claims over fluent ones.
2025 · 11
The critic gap
Why self-critique improves precision but collapses recall, and how to reopen it.
2025 · 07
Evaluation you can argue with
Team-authored rubrics beat public benchmarks for predicting real-world usefulness.
Principles
Three commitments we do not trade away.
They shape what we publish and what we refuse to ship.
Legibility
If a result cannot be explained, it is not a result.
Contestability
Any claim the system makes must be challengeable by a person.
Restraint
Capability without a use we understand stays in the lab.