Decisions

Mechanistic interpretability forces you to make choices about what kind of thing a “mechanism” is before you run your first experiment. These choices are often implicit — buried in method selection or inherited from a prior paper. The decision pages make them explicit.

Each page examines a fork where different views lead to different experiments, different claims, and different failure modes. They are organized around the practical question: given my research goal, which view should I adopt?

• Object vs. Role — Is the mechanism a specific component (head 9.9) or the functional role it plays (name-moving)? This determines whether your findings generalize across model seeds and architectures.

• Subspace vs. Structural — Is subspace proximity sufficient for mechanism identity, or do you need gauge-invariant characterization? This matters whenever you compare mechanisms across models that may differ by an internal symmetry transformation.

• Localized vs. Distributed — Can the mechanism be assigned to a compact circuit, or is it spread across the network? This determines whether circuit discovery methods can recover your mechanism at all.

• Single-Method vs. Triangulated — When is one strong experiment enough, and when do you need converging evidence from independent domains? This determines the strength of the claims you can make.

• Static vs. Process — Should the mechanism be described as a final-state object, or does its formation history matter? This determines whether you need checkpoint data and dynamics-domain evidence.

• Behavioral vs. Mechanistic — Is a finding about a model’s behavioral propensity or about an internal mechanism? This matters most for safety, where the gap between instrumental evidence and mechanistic conclusions is the central evidence deficit.