Preface

A child confronts the problem at the center of this book. The world does not arrive already divided into objects, actions, causes, and rules. It arrives through many senses at once, changing as the child acts within it. William James memorably described this first encounter as “one great blooming, buzzing confusion” [ James , 1890 ] . The thought is almost frightening: each of us once faced the task of making an intelligible world from that confusion, yet none of us can remember how our brain managed to do it.

Evolution had millions of years to shape a learner capable of meeting this challenge. Its achievement was not to provide the newborn with a finished theory of a particular world. It was, rather, to equip the child with abstractions through which a theory could be progressively constructed by observation and interaction. The learner begins neither with a blank slate nor with a complete model, but with forms of organization that make experience learnable.

Spelke’s core-knowledge hypothesis gives empirical content to these forms of organization. Evidence from infants and other learners supports six early emerging systems: representations of cohesive physical objects and their mechanical interactions; approximate number; the geometry of navigable space; object form; goal-directed agents; and social beings with attention, emotion, and mental states [ Spelke , 2022 , 2023 ] . These systems are not finished theories of the world. They are automatic, domain-specific interfaces between perception and explicit belief that guide learning and continue to function throughout life. They make the blooming, buzzing confusion initially probeable.

Piaget gave this construction a dynamic form. A child assimilates a new experience when it can be incorporated into an existing schema; the child accommodates when the schema itself must change to make sense of experience [ Piaget , 1952 ] . Learning is therefore not simply the accumulation of observations. It is the continuing construction, application, and repair of structures that organize observations and actions into a coherent whole. Abstracted from its developmental setting, the Piagetian conjecture is that an intelligent learner progressively builds a compositional theory of its world.

This book begins where Infinitesimal Creativity left us. That book studied how a learner can vary, compare, and repair compositional structure through infinitesimal probes and double involution. Here we move one level deeper. Before a learner can creatively transform a theory, how can it discover the theory’s objects, transformations, compositions, and universal regularities from interaction? And before it can perform an infinitesimal repair, how can it discover which variations the world admits?

Discovery may also be more compact than a catalogue of the world. A child playing with construction pieces can learn a small algebraic theory of joining, juxtaposing, and separating whose composites generate indefinitely many assemblies. ORACLE must therefore distinguish learning a category of encountered structures from learning a sketch of generators and relations that presents those structures and transports across models.

Our categorical hypothesis is deliberately spare:

The stronger refinement developed in this book replaces the unknown category by an unknown tangent category. The learner must then identify not only a base category \(\mathcal C\), but tangent structure

describing admissible infinitesimal variation and its coherence. This need not be an innate commitment imposed on every UOCL instance. It is a structured hypothesis class whose additional probes make infinitesimal learning possible.

This is not the claim that an infant consciously knows category theory. Rather, categorical structure plays the role of an innate form of organization. Objects can persist, arrows can be composed, identities can be recognized, and different paths can be compared. Interaction then supplies a presentation from which a particular category may be identified: its objects, morphisms, composites, equations, and universal properties, but only to the extent that the learner’s admissible probes can distinguish them.

We call the resulting formal learner oracle: an Online Rational Agent for Categorical Learning. Here rationality means coherence with the categorical structure supported by evidence, not prior possession of an optimal policy. Assimilation extends a currently viable categorical model with new observations and composites. Accommodation repairs the presentation—or changes the class of models—when the new evidence cannot be coherently absorbed. The familiar name is appropriate: the learner’s knowledge is tested by the structural questions it can answer.

Spelke’s results also sharpen what this learner may know before interaction. The categorical prior need not be an empty doctrine of objects and arrows. It may include a small, typed family of core observers—for objects, number, space, form, agency, and sociality—that determine which distinctions can be made first. The category of the world is still unknown; core knowledge provides some of the probes through which its presentation is revealed. A central question for ORACLE is then how initially distinct systems become linked by composition without losing the invariants that made early learning possible.

This formulation is a categorical generalization of Gold’s identification of languages from presentations and of predictive system identification. It also generalizes universal imitation games, where inductive inference by initial algebras was contrasted with coinductive inference by final coalgebras. ORACLE permits the unknown target itself to range over a world of categories. A universal coalgebra is one important case, not the boundary of the theory.

Universal Online Categorical Learning supplies the algorithmic bridge from this semantic program to decision. UOCL specifies how hypotheses are selected, probes are chosen, evidence is assimilated, and categorical declarations are repaired under progressive revelation. Its open learners form compositional systems: modality-specific learners can run in parallel, be wired in sequence, or be fused by limits enforcing a shared object, spatial, linguistic, or social doctrine. Tangent UOCL asks the same learner to identify how its categorical world can vary. In some domains it may seek the stronger explanation of a differential category whose coalgebras generate that tangent geometry; this is a separate identification claim, not an automatic consequence of possessing derivatives. Universal Online Decision Learning is a second enrichment, obtained when a learned category is equipped with information, action, consistency, consequence, and observation maps. Their intersection—tangent UODL—creates the typed infinitesimal conditions under which DIAL can diagnose and repair a decision mechanism rather than merely assume those conditions in advance.

Exact categorical identification is not always statistically attainable or task-relevant. PACC UOCL therefore adds a controlled relaxation: a learned world is probably approximately categorically correct relative to a declared distribution of equivalence-invariant categorical probes. Ordinary PAC learning is recovered when the world and probe language are discrete; structural and coherent PACC additionally test composition, universal properties, horns, and filler spaces.

This generality contains familiar learning paradigms as deliberately restricted cases. System identification searches a category of dynamical models; automata and machine inference learn transition coalgebras up to observable behavior; reinforcement learning fixes an MDP doctrine and adds consequential action; causal discovery searches a category of causal models; grammatical inference and Transformer language modeling restrict the unknown world to classes of languages or conditional sequence laws. Prometheus and Odyssey instead compile document presentations into local, action-oriented predictive-state models and glue them into provenance-bearing Topos World Models. Chapter 5 makes these reductions explicit while distinguishing genuine categorical structure from the vacuous act of treating any hypothesis set as a discrete category. Online convex optimization supplies a compact test suite rather than the architecture of the book. Hazan’s progression from regularized leader methods through bandit feedback and boosting occupies the global-clock case; later parts treat non-linear information structures, homotopy-coherent repair, safety, and lifelong transport.

The four books form a systematic ascent from static modeling to dynamic discovery. Categories for Artificial General Intelligence gave a functorial semantics of cognition: morphisms represented compositional steps of thought, and functors represented systematic translations between conceptual systems. Its categories, however, were available before the translation began. Machine Learning by Enforcing Compositionality made learning depend on sketches—finite declarations by generators, relations, cones, and cocones. Learning became the completion and comparison of models satisfying a declared sketch. The model was unknown, but the type of sketch was fixed for the learning problem; its constraints need not select a unique model unless additional hypotheses make them separating.

Infinitesimal Creativity then asked how an identified compositional structure could vary and be repaired. Tangent categories typed the admissible infinitesimal directions, while involution algebroids and double involution supplied an algebraic mechanism for creative variation and reconstruction. The base structure on which those variations acted was still assumed to have been identified. The present book makes the remaining assumption explicit: the doctrine—the structural grammar specifying what kinds of sketches, models, morphisms, and universal constructions are admissible—may itself be unknown.

Here doctrine is used in a deliberately broad Lawvere-style sense. A doctrine is a structured theory of categories and their models, together with its admitted structure-preserving maps. Many doctrines can be presented by a 2-monad, an essentially algebraic theory, or a specified class of sketches, but the word is not restricted to any one of these realizations. Lawvere theories, monoidal categories, causal models, Markov decision processes, toposes, and tangent categories exemplify doctrines at different levels of structure.

The technical object that gathers the four books is a hypothesis fibration. Suppressing the separate transcript coordinate for the moment, write

The base \(\mathcal B\) records admissible doctrines and their declared revisions; the fiber \(\mathcal E_b\) contains the hypotheses meaningful under doctrine \(b\); and reindexing translates a hypothesis along a change of doctrine. In the full theory, presentation prefixes refine this base, as in the Grothendieck construction \(\int H_{\mathrm{doc}}\to \mathbf{Doc}\) and its pullback to the observed presentation path. An ORACLE learner is therefore a coherent section along that path. A global section \(\sigma :\mathcal B\to \mathcal E\), when one exists, is a stronger semantic ideal rather than an assumption made by every learning problem.

From static sketches to fibered presentations. In the second book, a learner worked relative to the selected sketch: it could complete or repair that declaration, but the ambient language of admissible declarations was held fixed. ORACLE distinguishes the growing presentation from its doctrine. A learner may begin under a coarse commitment such as “category” and move, when licensed by evidence, to a more structured doctrine such as “differential category” or “stable \(\infty \)-category.” Reindexing and transport record which earlier conclusions survive the refinement.

From fixed variation to learned variation. The third book’s infinitesimal mechanisms presupposed a base on which variation was defined. Tangent UOCL learns the base category and tangent structure jointly. In a tangent enhancement of the displayed hypothesis fibration, the learner must identify not merely directions in a preselected model but the projection, zero, addition, lift, and canonical flip that make those directions coherent. This is what creates the conditions under which DIAL’s infinitesimal repair has an identified rather than stipulated semantics.

Beyond translation. A functor in the first book mapped between already available worlds. A section of the ORACLE fibration is dependent: the category in which the selected hypothesis lives varies with the doctrine. The learning problem is consequently not only to translate between worlds, but to discover which world of models the interaction supports and which changes of world preserve established knowledge.

This also gives the Piagetian dynamic a precise address. Assimilation is update within a fixed fiber: completing a sketch, estimating a model, or moving within its learned tangent geometry. Accommodation changes the base object from \(b\) to \(b'\), then transports the settled part of the old hypothesis into \(\mathcal E_{b'}\). The two operations are not competing learning rules. They are vertical and base-changing components of one fibered learning process.

The tetralogy can therefore be summarized by four questions: Which compositional translation? Which presentation and model? Which admissible variation and repair? Which doctrine makes those questions meaningful? ORACLE provides the meta-theory that makes the earlier three programs operational: it asks how to discover the appropriate functors, sketches, and tangent structures, and how to revise all three when interaction exposes a doctrinal obstruction.

The present book assumes none of the earlier volumes’ terminology. Chapter 1 develops the working language of categories, Kan extensions, simplicial nerves, homotopy-coherent diagrams, stable \(\infty \)-categories, and tangent structure. Later chapters return to these constructions with sharper hypotheses and decision-theoretic interpretations.

¹1. The central organizing principle of the book is Piagetian: assimilate what fits the current compositional world; accommodate when the world itself must be revised. ↩

The present draft preserves the exact finite UODL and FTRL results as controlled mathematical witnesses, but places them after a new UOCL theory of effective categorical and tangent identification. In this sense the fourth book supplies the acquisition theory that makes the DIAL process feasible. It also records theorem boundaries: prediction is not realization, categorical identification is not causal identification, and a tangent comparison does not establish convergence without stability assumptions.