Mechanistic interpretability
Research that investigates internal model mechanisms to explain particular behaviours.
ConceptSources & confidence: High confidence
Checked 2 October 2026 · AI-assisted review
Checked the definition, attribution and stated distinctions against the listed source passages.
- Anthropic, On the Biology of a Large Language Model (2025)
Attribution-graph case studies and explicit limitations of approximating model computations.
What this check doesn’t establish
- The check covers this representative primary study, not the whole field or subsequent capabilities.
Confidence applies to the scope above, not every claim made by the underlying source. How confidence works
Mechanistic interpretability attempts to explain behaviour through structures and computations inside a model. Researchers can form and test hypotheses using interventions, features and circuit analyses.
This is not direct, complete access to a model’s reasoning. The cited attribution-graph research uses approximations, studies selected behaviours and describes gaps in what its methods capture. It does not establish a general fingerprint that identifies an agent’s character across arbitrary updates.
Record history
Catalogue source: knowledge/concepts/mechanistic-interpretability.md
Generated by claude-code/claude-opus-5 on 2026-07-26