What we publish before we run it.
Publications, research notes, model cards, and evaluations. Every product links back here to the method behind it.
Browse publications
Models, methods, and limits.
- 2
- 1
Work & methods
We publish the models, the evaluations, and what failed. That is what makes the products correctable.
How to read- Compaction that fails closed: keep the latest requests, the active plan, and the open artifactsHarness
- Eleven SAST rules we own, and why a threshold can't silence themSecurity
- Measuring a stream: a "single-block" response is a reading of the transport, not the renderInfrastructure
- Visual grounding: describe a control, let a second call find the pixelComputer use
- What Deep Research cites, and why the citation is part of the messageResearch
A title is not proof.
To read a publication: identify the protocol, find the evaluation conditions, then look at the limits.
01 · Method
What problem is being studied? What protocol makes it reproducible?
02 · Evaluation
Under what conditions is the model evaluated? On what data?
03 · Limits
What failed? What does the measurement not allow us to conclude?
Start from a model
Every publication links back to the model and the evals behind it.
Explore the models