Neurosymbolic AI, Hybrid Architectures & the Three Mechanisms

The Inquiry: Has the neurosymbolic AI field — in three decades of research on the "integration question" (how to combine neural and symbolic systems) — identified, named, or formalized what we call Constitution as a distinct mechanism of model creation co-equal with training? And does training-based optimization conflate statistical rarity with structural prohibition in a way that no training regime can resolve?

The neurosymbolic AI tradition (Garcez & Lamb 2023, Marcus 2020, Kautz 2022) has produced a rich taxonomy of integration architectures — from loose coupling (Type 2: Symbolic[Neuro]) to deep embedding (Type 6: Neuro[Symbolic]). Yet all six types in Kautz's taxonomy assume integration of two components within a single system. The ontology engineering tradition (Gruber 1993, Guarino 1998) has been specifying constitutive constraints — what can meaningfully exist in a domain — for three decades, without recognizing this activity as a mechanism of model creation co-equal with training. If neither tradition has named this mechanism, then a three-mechanism taxonomy (Training, Constitution, Accretion) is a genuine contribution. If one has, the contribution shifts to integrating existing concepts under a unifying framework.

Falsifiable formulation: The thesis is falsifiable: if the NeSy literature contains a named concept equivalent to Constitution-as-mechanism, the novelty claim fails. If the rarity/prohibition conflation has been identified and resolved within training-based approaches, the architectural argument for a separate constitutional mechanism weakens.

The integration question vs. the mechanism question. The NeSy field's 30-year organizing question — "how do we combine neural and symbolic?" — has produced valuable architectures (F1, F3, F5) and compelling arguments for hybrid systems (F2). But the question itself presupposes that there are two things to be combined. The mechanism question — "how do models come into existence?" — reveals three distinct processes (F13), each with different temporal character (F15), different structural guarantees, and different coverage of what a complete world model requires. The NeSy tradition has been answering the wrong question — not a bad question, but an incomplete one.


Constitution as unnamed practice. The most striking finding is F8: the practice of Constitution has existed for 30 years under the name "ontology engineering." Gruber (1993) defined the core act — explicit specification of a conceptualization (F6). Guarino (1998) formalized the key property — ontological commitments as constitutive constraints (F7). But neither recognized this as a mechanism of model creation co-equal with training. The ontology engineering community treats specification as knowledge representation — a way of encoding what you already know. The mechanism perspective treats it differently: Constitution is how structural constraints come into existence. It's not encoding existing knowledge; it's creating the structures that define what knowledge can exist.

This finding resolves SCT-C01 in a nuanced way: Constitution-as-practice has been named (it's called ontology engineering). Constitution-as-mechanism has not. The novelty is the mechanism recognition, not the practice. This is analogous to how Deming didn't invent manufacturing inspection — he recognized it as a mechanism with specific properties and limitations, enabling the shift to quality-by-design. The three-mechanisms taxonomy does for knowledge creation what Deming did for quality: it recognizes that the mode of creation determines the properties of the output.

The rarity/prohibition argument. F9 through F12 establish the sprint's strongest novel contribution. Training cannot distinguish between statistical rarity and structural prohibition because both manifest as low-probability regions in the loss landscape. This is not a limitation of current models — it is intrinsic to what training is (optimizing a loss function over observed data). Constitutional uncertainty (F10, F18) — the uncertainty about whether a low-probability observation represents genuine rarity or structural impossibility — is resolvable only by specification, not by observation. Yuenyong's (2025) counter-argument (F12) — that LLMs can learn everything ontologies specify — fails precisely at this boundary: LLMs can learn that negative-time-duration transactions never appear in data, but they cannot learn that such transactions are structurally prohibited as opposed to merely unobserved.

The established epistemic/aleatoric dichotomy (F18) does not capture this. Epistemic uncertainty is about what the modeler doesn't know. Aleatoric uncertainty is about inherent randomness. Constitutional uncertainty is about structural constraints on what can exist — a domain property, not a modeler property or a random process property.


Accretion as the invisible mechanism. F14 establishes that Accretion has no NeSy equivalent. The NeSy field sees neural (Training) and symbolic (Constitution-as-component) but not operational accumulation (Accretion). This is because NeSy studies system architecture, not organizational knowledge lifecycle. The organizational memory literature (Walsh & Ungson 1991, Nonaka & Takeuchi 1995 — engaged in prior sprints S3 and S14) has studied what we call Accretion extensively but has not named it as a mechanism of model creation. The three-mechanisms taxonomy bridges these fields by recognizing Accretion as a third mechanism alongside Training and Constitution.

SAI as validation. Goldfeder/Wyder/LeCun (2026) validate the composed-systems thesis (F16) and the MILA workshop (F17) suggests growing interest in modular, specialized architectures. But neither addresses the governance question: how do you coordinate specialized modules? Who decides authority? How do you prevent incoherent joint behavior? SAI provides the existence proof for the gap — composed systems need coordination, and training alone cannot provide coordination guarantees. This is where Constitutional structure (authority hierarchies, constraint propagation, accountability chains) becomes architecturally necessary, not just organizationally convenient.