Mechanistic Validity: A Theory of Validity for Interpretability Claims

A mechanistic explanation of a neural network is not a single experimental result, but a structured claim linking a theoretical construct, a measurement procedure, an intervention, and a scope of interpretation. We introduce Mechanistic Validity (MechVal), a formalized epistemology for validating these causal claims, drawing on validation methodology from causal inference, measurement theory, and the philosophy of science. The framework makes interpretability claims explicit, falsifiable, and comparable, formalizing the discovery, validation, and interpretation of causal mechanisms as separate evidential stages. MechVal evaluates claims through six layers: (1) description mode, (2) evidence family, (3) metrics, (4) criteria, (5) validity type, (6) verdict. A gap at any layer is visible in the final verdict. The framework is method-agnostic, applying the same criteria to circuits, SAE features, probing classifiers, steering vectors, transcoders, and crosscoders. Applying the framework to 13 published circuits, we show that the tier system discriminates among claims—from Proposed (probing classifiers) to Triangulated (induction heads)—with every verdict traceable to specific present or missing evidence. By making the evidential status of a claim explicit and the path to strengthening it concrete, MechVal provides a foundation for interpretability claims that are explicit, testable, and scientifically grounded.