The research programme
From the geometry of a single representation to the institutions a deployed system reshapes, four connected layers, studied with the conviction that they are one subject.
A The stack
Most research on AI systems picks one altitude. Mechanistic interpretability studies features and circuits; behavioural evaluation studies outputs; policy studies deployment. Each is productive, and each systematically underdetermines the others: circuit evidence does not settle what a system will do under distribution shift, behavioural evidence does not settle mechanism, and neither settles what an institution should do about a system it now depends on.
The institute's programmes are arranged as a stack because the phenomena are. Representations are read by circuits; circuits produce cognition and behaviour; behaviour earns deployment; deployment feeds back into training. Claims move up and down this stack constantly in public discussion, usually without the evidence that would license the move. Making those moves rigorous is the research programme.
B The four programmes
Representations
Features, superposition, activation geometry, sparse autoencoders, concept directions, persona vectors, and the structures through which models encode the world.
Probes · SAEs · crosscoders→ 02Computation
Attention and MLP circuits, attribution graphs, activation patching, causal tracing, model editing, and the discipline of causal over correlational evidence.
Patching · tracing · editing→ 03Cognition & behaviour
Situational awareness, evaluation awareness, self knowledge, unverbalised reasoning, chain of thought faithfulness, deception, personas, and introspection.
Evals · steering · introspection→ 04Deployment & institutions
Model entrenchment versus AI system entrenchment, infrastructure dependence, lock in, replacement and shutdown, and interpretability as institutional evidence.
Case coding · economics · audits→Every research object is filed under a primary programme in the papers index, and the field map arranges the institute's open questions across all four. The construct map documents, for each strategic game family, what is measured, the control a reading requires, and the claim the evidence never licenses. The entrenchment case coding page applies a fixed rubric to real sourced cases of hard to remove systems.
C Arenas
These are playable research prototypes that pool anonymous play data for exploratory analysis. They are prototypes, not benchmarks, and nothing about their pooled data is claimed here as evidence.