sec-kan-insufficiency

4.8.3 Why a Kan extension is not yet an ORACLE learner

There is an equally serious objection from within category theory. Mac Lane’s famous maxim that every concept is a Kan extension suggests that the universal construction might already contain the whole ORACLE program [ Mac Lane , 1971 ] . Given an indexing functor \(i:\mathcal A\to \mathcal B\), boundary data \(F:\mathcal A\to \mathcal C\), and an ambient category \(\mathcal C\), the left and right Kan extensions construct universal ways of extending \(F\) along \(i\). What more could a universal categorical learner require?

The answer is contained in the word given. A Kan extension solves a conditional completion problem after the indexing shape, boundary data, codomain, admissible morphisms, and direction of universality have been declared. ORACLE begins without knowing which such declaration is warranted. It must discover or repair:

which observations denote the same object and which arrows compose;
which presentation shape records the relevant evidence;
which observer and query language determine empirical agreement;
which equivalences identify two realizations as the same world; and
whether new evidence extends the current model or invalidates part of the declaration itself.

Thus universal completion and categorical identification answer different questions. A Kan extension asks for the best extension under this declaration. ORACLE asks which declaration and which equivalence class of worlds are supported by interaction.

Proposition 4.8 Kan-insufficiency principle

Let \(M\) and \(N\) be inequivalent admissible worlds whose observer-visible restrictions induce equivalent boundary diagrams \(F_M\simeq F_N:\mathcal A\to \mathcal C\). Suppose their current left and right Kan extensions along \(i:\mathcal A\to \mathcal B\) agree on every registered query, but an admissible continuation or query separates \(M\) from \(N\). Then no learner whose inference is restricted to those current Kan extensions can identify which of \(M\) or \(N\) generated the presentation.

Proof ▼

Kan extension is determined, up to its universal equivalence, by the declared indexing functor and boundary diagram. Equivalent visible boundary data therefore yield equivalent current universal completions and identical answers to every query that factors through them. By assumption, the evidence or query that separates \(M\) from \(N\) lies outside this common observable completion. Consequently an extension-only learner receives no distinguishing witness.

This is another indistinguishability theorem, not a defect peculiar to Kan extensions. The construction cannot create evidence absent from its boundary diagram. Several further gaps follow.

: [leftmargin=!,labelwidth=3.2cm,itemsep=4pt]
Relativity.: Universality is relative to the chosen categories, morphisms, and 2-cells. A perfectly computed extension can be the correct answer to a wrongly declared learning problem.
Probes.: The extension need not select the next probe that separates the remaining hypothesis fiber.
Persistence.: Independently universal extensions at times \(t\) and \(t+1\) need not preserve settled knowledge. That requires coherent comparison maps, base-change conditions, and a repair policy.
Effectivity.: Abstract existence does not imply that an extension can be computed, estimated from finite noisy data, or represented within the learner’s resources.
Causal meaning.: Left and right universality provide extremal completions, but an observer and intervention semantics are still needed to interpret their agreement or discrepancy causally.

Kan extensions nevertheless remain central rather than dispensable. Their place in ORACLE is one layer of a larger architecture:

\[ \text{declaration discovery and repair} \longrightarrow \text{universal extension} \longrightarrow \text{observer-relative comparison}. \]

The first layer determines what is being extended. The second constructs the universal candidates. The third determines what their comparison warrants for the learner. Online persistence adds coherent transport among successive instances of all three layers.

One may reply that declaration discovery is itself expressible as a Kan extension in a higher category of presentations and doctrines. This may be formally correct; it does not remove the epistemic content. It relocates that content into the choice of higher category, observer, admissible repairs, and equivalence relation. Mac Lane’s maxim establishes expressive universality, not the empirical sufficiency of an undeclared extension problem.

Design principle

Every concept may be expressible as a Kan extension, but a Kan extension does not by itself identify which concept is warranted by the evidence.

Design principle

A categorical prior has substantive developmental content only if it predicts an identification, invariance, transport, or repair distinction not already implied by unrestricted sequence or latent-state prediction on the same evidence.

There are then three legitimate outcomes. A Transformer system may satisfy the ORACLE obligations, and a JEPA may identify a query-sufficient, action-coherent world model; in either case the theory supplies semantics for what has been learned. A system may satisfy local prediction while failing persistence, re-presentation, underdetermination, or repair, in which case ORACLE localizes the missing structure. Or the proposed obligations may predict nothing beyond ordinary sequence or latent prediction, in which case the categorical account is a redescription and must be weakened, revised, or rejected.

The elephant in the room thus becomes a theorem and evaluation program rather than a contest between neural and categorical slogans. The central question is whether frontier systems discover persistent compositional worlds, learn query-sufficient predictive quotients of them, or reconstruct locally adequate fragments of worlds on demand.

Guiding question.

Can two learners with matched compositional performance be separated by a finite family of re-presentation, continuation, repair, abstention, and transport tests derived from the declared categorical prior?