Research
Our work runs along three connected lines, from formal foundations through to governance implications.
Formal foundations
What would it take to say that a system has goals, and to be wrong about it in a detectable way? We work on accounts of goal-directedness, coherence and optimisation with enough structure to generate predictions.
Measurement
A definition that cannot be estimated from data is not yet science. We develop measures of agency that can be applied to real systems, and study how they behave as capability scales.
Implications
If agency can be measured, evaluation and oversight regimes can be built on it. We work through what changes for AI evaluation, for deployment thresholds, and for governance.
Outputs
Papers and preprints
Nothing published yet — this section will list papers, preprints and talks as they appear.