A research program for evaluating causal mechanisms — what they are, whether the evidence warrants them, and when they transfer across systems.
Mechanistic Validity
A Validity Theory, Research Methodology, and Evidence Standard for Mechanistic Interpretability
A component can affect an outcome without performing the function assigned to it. A measurement can be reliable without measuring the intended construct. A causal result can be genuine yet depend on one intervention, dataset, or experimental system. Existing evaluations often treat these as isolated limitations; Mechanistic Validity treats them as connected failures in the inference from evidence to claim.
The framework extends construct, measurement, internal, external, and interpretive validity to mechanistic research. These dimensions form a dependency structure: the target mechanism must be coherent before it can be measured; the measurement must be trustworthy before it can support causal inference; and a causal result established in one setting cannot automatically support broader generalization or interpretation. Evidence cards connect each conclusion to the specific measurement and criterion that support it, preventing strength along one dimension from concealing an unresolved failure along another.
Mechanistic Views
An Atlas of Hidden Commitments and a Realism Criterion for Mechanistic Claims
Researchers identify mechanisms as components, functional roles, causal subspaces, structural relations, or processes — and these are not interchangeable descriptions of the same object. Each carries different assumptions about what a mechanism is, when two mechanisms count as the same, and what evidence can identify one. Disagreements that appear empirical can instead arise because researchers are making different kinds of claims from the same evidence.
Mechanistic Views makes those assumptions explicit. It defines a mechanistic view through five connected commitments — ontology, identity, evidence, formalism, and explanatory target — and organizes nine views currently operating in mechanistic research, each with a characteristic failure mode. The paper also proposes a realism criterion: a reported structure counts as a feature of the system rather than of the measurement when it is recovered across approaches whose failure modes are genuinely independent.
Mechanistic Reference
When Does a Mechanism Term Pick Out the Same Thing?
Mechanistic research routinely carries terms across experimental boundaries: from one model to another, from animals to humans, from cell cultures to organisms, or from laboratory interventions to clinical settings. Similar behavior does not by itself establish that the same underlying mechanism is present. Conversely, a failed replication does not always show that the original result was invalid — the mechanism term may simply have crossed a boundary its evidence could not support.
Mechanistic Reference develops a theory of cross-system reference for such cases. It distinguishes several levels at which mechanisms can be identified — including concrete components, functional roles, causal subspaces, structural organization, and processes — and specifies the evidence required to transport a claim at each level. Evidence adequate for similarity at one level cannot substitute for evidence required at another: behavioral equivalence, for example, cannot by accumulation establish structural identity.