From causal models to actionable learning models

The book can now be read as carrying a central lesson of causal inference into machine learning. Association alone does not determine the effect of an action. A model must expose the relevant mechanisms, specify what an intervention changes, and state what remains invariant. Conventional causal inference applies this discipline to a domain model. LINCS applies it more broadly to the organization of learning itself.

This shift creates a hierarchy of models rather than a competition between causal and noncausal learning. A domain sketch may represent biological, economic, physical, or social mechanisms. A learning sketch represents the system that estimates, composes, compares, and repairs such models. A foundry sketch represents the longer-lived process that gathers evidence, maintains provenance, revises declarations, and decides what enters durable state. Interventions at these levels answer different questions and require different certificates.

The frontier is therefore not merely to attach causal variables to existing predictors. It is to build learning systems whose own components and obligations are explicit enough to support controlled action. Such systems should distinguish:

The higher the intervention rises, the more demanding the transport and admission problem becomes. A future theory of structural intervention should make these levels compositional, relate their invariance assumptions, and characterize when a repair at one level can be soundly realized at another.