Mechanistic interpretability investigates internal model computations. Anthropic's 2025-03-27 report linked interpretable internal concepts, called features, into computational pathways, called circuits, while studying Claude 3.5 Haiku on simple tasks. In some cases, traces showed no evidence of calculations that the model's written explanation claimed to perform. The authors report incomplete computational coverage, possible tool artifacts and hours of human analysis even for short prompts. These are partial, first-party case studies. Analytical distinction: whether an explanation reflects internal computation and whether its scientific claim is correct are different review questions.

Evidence status: partial_mechanistic_case_studies.