Where feature dictionaries give a parts list, attribution graphs give the wiring for one specific completion: nodes are interpretable features, edges are estimated causal influences, and the graph traces a path from the prompt through intermediate features to the output token. Anthropic builds them by substituting a trained replacement model whose components are interpretable by construction, then verifying key edges with interventions.
The companion paper On the Biology of a Large Language Model used attribution graphs to show Claude planning rhyme words before writing the line that leads to them, running parallel rough-estimate and exact-digit paths in mental arithmetic, and producing chain-of-thought text that played no causal role in its answer, the phenomenon behind chain-of-thought unfaithfulness.