The boundary of the claim

lincs does not assert that every compositional failure has a unique cause, that every obstruction supports a repair, or that categorical structure eliminates uncertainty. It makes those boundaries explicit by distinguishing diagnosis from identification, intervention, repair, and admission.

Nor does it assert that every repair is a causal intervention. The broader claim is methodological: mainstream learning becomes more inspectable when its own architecture and obligations are modeled explicitly enough to support localized, invariant-preserving action. Causal semantics is an additional achievement when those modeled components are grounded as mechanisms.

That distinction matters as learning systems become modular, agentic, and distributed. A global score can improve while a local contract is violated; successful components can fail when composed; and a plausible repair can destroy the semantics used to justify it. This book treats such cases as central objects of learning.

The durable question running through the book is therefore not simply whether a model can reduce a loss. It is:

Which structural promise failed, how does that failure move, and what part of the modeled learning system may be changed while the rest is held fixed? What independent evidence licenses that repair without changing the promise merely in order to pass?

The chapters that follow develop a language, workflow, and family of systems for answering that question.