17.1 What the current theory establishes

ORACLE asks which categorical world can be identified from an interaction presentation and a declared family of probes. The first theorem ladder is therefore presentation, observational separation, identification, and realization. UODL enters only after the realized category carries the additional structure needed to make and compare decisions.

Online UDL is UDL evaluated on the least causally available past and subject to non-anticipation. Persistent online UDL retains the time-indexed semantic object and its coherent comparisons. At fixed diagram shape, its Kan invariance is precisely ordinary Kan invariance in a functor category. When the observation diagram grows, the Beck–Chevalley mate separates exact prefix transport from a genuine repair obligation. The decision nerve adds an intrinsic filtration by compositional depth, and skeletal continuity states exactly when incremental UDL recovers all-at-once semantics. Homotopy-coherent UDL then localizes presentations at declared semantic weak equivalences, replaces strict uniqueness by coherent spaces of choices, and turns a failed derived continuity comparison into a structured repair defect. This is the foundational departure from ordinary OCO: regret appears only after an observer compares the cumulative value of the causal branch with the full-information value of a right-Kan-consistent reference branch, whereas UODL first asks whether the two branches are coherently comparable.

The first convex UODL theorem is deliberately elementary, but it establishes the required pattern with no hidden identifications:

Its tangent lift commutes exactly under finite fixed-shape hypotheses, evidence and mechanism variations remain typed, and a nontrivial decision-null quotient is proved. The result supplies a stable base from which the harder right-consistency and repair questions can be asked precisely. This completes the compact OCO realization; no catalog of optimizers is needed.

The main theory begins when the classical chain is weakened. Bandit evidence can be reconstructed only after barycentric expectation. Decentralized and nonclassical information invalidates counterfactual replay on a fixed history. Homotopy-coherent repair replaces a single correction by a space of admissible revisions, while tangent stability asks whether infinitesimal transport survives iteration and limiting processes. Free convex completion then tests whether weak mechanisms can be composed into persistent stronger ones. The final ORACLE question is whether categorical discoveries, decision constructions, and their warrants transport across tasks and environments, accumulating as lifelong compositional knowledge rather than disappearing into the parameters of one predictive or decision rule.