Claude performs genuine two-step internal reasoning on 'the capital of the state containing Dallas' (Dallas → Texas → Austin)

In the "Introductory Example" of Anthropic's On the Biology of a Large Language Model (Lindsey et al., 2025), the attribution graph for the prompt "the capital of the state containing Dallas" shows Claude 3.5 Haiku first activating a feature representing Texas, then using that intermediate representation to retrieve Austin — rather than jumping straight from "Dallas" to "Austin" via a memorized association. The report states that "the model performs genuine two-hop reasoning internally, which coexists alongside 'shortcut' reasoning" (the report's term is "two-hop"; "two-step" is used here as an equivalent gloss).

The intermediate step is shown to be causally load-bearing, not merely correlated: swapping the internal Texas-representing features for California-representing features changed the model's output to Sacramento. The inferred fact is used downstream, and intervening on it changes the answer in the way a real chain of inference predicts.

This is the report's cleanest positive example of chain-shaped internal computation, and it sits alongside — not against — the faithfulness findings in cot-faithfulness-anthropic-biology: whether a behavior counts as reasoning is decided case by case, and this is the strongest positive case in the paper. It contrasts with three other traces documented in the same report: claim-biology-llm-poetry-planning-preactivates-rhyme-words (lookahead under constraint, not sequential inference), claim-biology-llm-hallucination-is-known-entity-suppression-misfire (a gate misfiring), and claim-biology-llm-jailbreak-assembles-bomb-by-parallel-letter-votes (the explicit absence of a chain).