Correlation: a feature or neuron lights up on "cat" inputs Steering: adding the direction pushes outputs toward "cat" Ablation: removing it breaks cat-related behaviour Counterfactual concept swapping: patch in "dog" activations where "cat" was and the downstream behaviour flips coherently (concept swapping in J-Lens and attribution graphs) [Verbalizable Representations Form a Global Workspace in Language Models](https://transformer-circuits.pub/2026/workspace/index.html) Simulation: you can rebuild the mechanism well enough to predict or reproduce the model's behaviour [Neural networks are not already clean — LessWrong](https://www.lesswrong.com/posts/Dkh7pXKjoNWkkxQDz/neural-networks-are-not-already-clean) Sometimes intervention of a mechanism at some layer might indirectly intervene another mechanism at some other layer which might actually be the actual thing that we care about, so ideally we would also do more proper localization of the mechanism by going through the space of interventions of mechanisms. And do confounder analyses by going through the space of confounders and trying to control for them. And there are sometimes "self-repairing" mechanisms after intervening, so trying to account for that too. [\[2402.15390\] Explorations of Self-Repair in Language Models](https://arxiv.org/abs/2402.15390) And making sure that the intervention itself doesnt create the mechanism itself.