5 Learning Inside Fixed Doctrines

A doctrinal audit of established learning paradigms

No learning algorithm begins without a doctrine. Before observing data, it already presupposes which objects and maps may constitute a world, how observations compose, which solution constructions exist, and which equivalences make two candidate worlds indistinguishable. These commitments are often distributed across a model class, a loss, a feedback protocol, and the hypotheses of a convergence theorem. Chapter 1 collects them under one name: the learner’s doctrine.

Established methods become tractable by fixing a sharply restricted doctrine before learning begins. Reinforcement learning ordinarily fixes a Markovian, reward-enriched world in which Bellman operators can be formed; causal discovery fixes a world of graphs or structural mechanisms with intervention semantics; automata learning fixes a coalgebraic machine type; and Transformer language modeling fixes a sequential presentation and continuation-prediction interface. Their achievements show how powerful the right doctrine can be. They do not show that the doctrine itself was discovered from interaction.

The central claim of this chapter is therefore not that every learning algorithm is secretly doing category theory. It is that a learning paradigm becomes a query-relative UOCL specialization only after its structural doctrine, presentation, solution construction, and success quotient are made explicit. Most existing methods learn inside that declaration. ORACLE asks the additional question of how the declaration can be identified, compared, composed, and repaired.