Interpretability
Mapping the internal structure of learned models — features, circuits and the geometry that connects them.
Research
Understanding systems from the inside out.
Our research asks how capable models are structured internally — and how that structure can make them more reliable, more legible and easier to direct.
Areas of focus
Reliability
Methods for predictable behaviour under distribution shift, adversarial pressure and long horizons.
Reasoning
How models plan, decompose problems and verify their own intermediate steps.
Evaluation
Measurement science for capable systems: benchmarks, protocols and the infrastructure to run them.
Publications
- Crystalline representations in learned modelsInterpretability
- Evaluation as a first-class artifactEvaluation
- Compositional control via shared feature axesInterpretability
- Step verification for long-horizon agentsReasoning
- Calibrated refusal under distribution shiftReliability
- A protocol for reproducible capability evaluationsEvaluation