2.2 An actionable model of the learning system

Causal inference made a decisive methodological move: it replaced a purely associational description with a model whose mechanisms could be acted upon [ Pearl , 2009b , 2018 ] . LINCS applies that lesson to the learner. The sketch does not merely describe the external domain; it can describe the architecture of learning, the transformations it performs, the obligations those transformations must satisfy, and the observations by which their failures become visible.

A minimal actionable learning model may be written

where \(\mathbb S\) is the structural declaration, \(D\) is its current realization, \(\Lambda \) is the observer family, \(\mathcal R\) is the typed language of admissible interventions, and \(\mathcal A\) is the independent decision rule. An edit \(e\in \mathcal R(D)\) then induces

For a parameter update, \(\mathbb S_e=\mathbb S\). For an architectural or theory-level repair, the declaration itself may change and the old model must transport conservatively into the new one.

This is analogous to causal surgery but intentionally more general. The edit must name its target and its invariants, yet it is causal only when the target is a domain mechanism and the edit has intervention semantics. Adapter replacement, reward-representation repair, cover refinement, and warrant revision are structural interventions on learning systems without thereby becoming causal effects.