Worked Examples

Each worked example applies the mechanistic views framework to a concrete interpretability case, showing how multiple views combine to characterize the same phenomenon at different levels of description.

• Indirect Object Identification (IOI) — exercises seven views: object, role, subspace, structural, process, contrastive, and perspectival
• Induction Heads — exercises the object, role, structural, and process views
• Superposition — exercises four views: object, subspace, structural, and perspectival
• Refusal Direction — exercises the instrumental, subspace, and contrastive views
• Probing Classifiers — exercises the instrumental, subspace, and role views