Attribution & Activation Patching

The workhorse of causal interpretability:

1. Clean run — run prompt A ('John gave Mary a drink'), cache every activation 2. Corrupted run — run prompt B, corrupted to break the behavior (swap names, change the fact) 3. Patch — splice one component's activation from the clean run into the corrupted run 4. Measure — does the output recover toward the clean answer?

If patching layer-12 head 7's output restores the behavior, that component carries the causally-relevant signal. You've localized where the answer lives with a scalpel, not a guess. This is the difference between 'the head correlates with the behavior' and 'the head transmits the behavior.'