An observability and replay toolkit for studying world models. It makes a model’s internal activity available for closer inspection, from representations and uncertainty to the effects of an intervention.
Cache activations and inspect saliency, belief tracking, surprise, and uncertainty.
Revisit trajectories and compare alternative inputs or branches of imagined behavior.
Use causal tracing and related interventions to investigate which internal paths matter.
Asks whether the signals a model highlights are the ones that influence its prediction. The framework compares an attribution with the change produced when the corresponding information is altered.
A developing set of reproducible diagnostics for comparing model internals and behavior. The goal is to make evaluations easier to repeat across examples and model variants.
