ora-0087

5.9 What the fixed-doctrine analysis establishes

Table 5.1 reveals a common pattern, but UOCL is not valuable merely because it can rename familiar components. Its contribution begins by turning the older paradigms’ implicit doctrines into explicit, comparable declarations:

  1. It types the hypothesis category and its morphisms rather than naming only a parameter space.

  2. It separates the world from the presentation through which the world is revealed.

  3. It makes the query-relative quotient explicit, preventing prediction, control, causal, and explanatory identification from being conflated.

  4. It distinguishes assimilation inside a doctrine from accommodation of the doctrine itself.

  5. It asks whether learned answers persist under revision and transport, not merely whether instantaneous error becomes small.

  6. It separates an internal solver—Bellman iteration, conditioning, likelihood optimization, final semantics, descent, or planning—from the doctrine that guarantees the solver is meaningful.

Design principle

An established learner is a special case of UOCL only after its hypothesis doctrine, presentation, queries, solution construction, equivalence, update, and convergence notion have been declared. Its theorem proves success relative to that declaration. It does not prove that the declaration is the uniquely correct account of the world.

This chapter also reverses the comparison. Automata, MDPs, causal models, grammars, PSRs, and latent world models are not competitors to ORACLE. They are candidate doctrines, often extraordinarily successful ones. Their theorems demonstrate how much becomes learnable once the right world class and solver are fixed. Their systematic failures reveal possible doctrinal boundaries: the world may not be finite-state, Markovian, causally sufficient, autoregressive at the chosen representation, globally gluable, or adequately captured by the current latent quotient.

ORACLE therefore adds a new outer loop. The inner learner estimates a model, predictive state, value function, causal graph, grammar, or representation inside \(\mathfrak D_t\). The outer learner compares the residual obstructions against alternative doctrines and, when warranted, transports settled structure along

\[ \mathfrak D_t\longrightarrow \mathfrak D_{t+1}. \]

The difficult problem is not merely choosing a larger hypothesis class. It is declaring which evidence licenses that move and which earlier conclusions survive it.

Guiding question.

For each established paradigm, what is the weakest structural doctrine, smallest separating query doctrine, and minimal internal solver that recover its successful predictions? Which observable obstruction would justify leaving that doctrine while preserving the distinctions needed for later decision, causal repair, and cross-task transport?