Mechanism-Based Generalization

Mechanism-based generalization is an alternative to statistical extrapolation for establishing external validity. Instead of asking whether an effect size replicates across contexts, it asks whether the causal mechanism that produced the effect is still active in the target context. The approach shifts the burden of proof from data volume to theoretical clarity: if you know why an intervention works, you can reason about whether that why still applies elsewhere.

This approach is theoretically elegant but practically demanding. It requires a level of mechanistic detail that most studies do not provide. The actor-critic architecture in reinforcement learning offers a useful analogy: the actor proposes an action, and the critic evaluates it. In mechanism-based generalization, the mechanism is the actor and the context is the critic — the mechanism must be evaluated against the specific features of each new environment. Without a well-specified mechanism, mechanism-based generalization collapses into intuition dressed as theory.