Where Does Intelligence Really Reside?: Beyond Individual Models Toward a Field-Based Explanation of Intelligent Systems
Where intelligence resides in a composed artificial-intelligence system may be underdetermined, ill-posed, or relative to a specified capability and intervention regime. The same language model can participate in systems with different tools, evidence sources, memory, missions, evaluation procedures, and institutional authority; conversely, similar observable outcomes can arise from materially different causal and organizational arrangements. Attributing capability solely to the model producing a visible response can therefore obscure conditions of its realization or legitimate exercise. This paper develops a bounded framework for capability attribution by separating four questions: which entity bears a capacity, which mechanisms causally realize it, which operating or deontic constitution makes its exercise possible or valid, and which independent evidence warrants the attribution. These distinctions permit system-level attribution without assuming that intelligence has a unique internal location, and without conflating competence and permission. The analysis examines the October 2025 proposal of Large Language Fields (LLFs) as named, reusable, versioned, mission-relative relational constitutions, alongside Large Signals Fields (LSFs), introduced in November 2025 records as a generalization to heterogeneous signals. The earliest identified LSF record retains an author-acknowledged creator-name variant, a documented archival discrepancy. The genealogy is compared with distributed and extended cognition, modular language-model programming, programmable guardrails, compound AI systems, agent interfaces, and context engineering. Because these antecedents already demonstrate the significance of model-external structures, the potential originality of LLF lies in its specific constitutional formulation and its proposed relations of alignment, conflict, transfer, and continuity across changing carriers, not in discovering such dependencies. Documentary chronology, conceptual novelty, and independently established scientific validity are therefore treated as separate questions. A typed Field schema is related to an earlier scientific qualification regime, keeping in-system evaluation distinct from the external evidence required to justify a scientific claim. Competing accounts include model sufficiency, an adequately specified conventional-system explanation, incremental Field-based explanatory or engineering utility, and cases in which localization remains non-identifiable. Unexecuted comparative protocols use strong conventional baselines, neutral checklists, terminology-swap controls, predefined effect thresholds, and explicit equivalence or rejection conditions. They allow conventional explanations to match or outperform Field-based descriptions without treating such outcomes as failures of scientific method. The paper contributes a defeasible discipline for attributing capability and a source-critical account of a proposed analytical unit. It does not establish a new physical law, empirically demonstrate Field-specific superiority, or defend a universal ontology of intelligence.
Authors
- Rogério Figurelli (ORCID: https://orcid.org/0009-0002-5726-1765)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24430664
- Primary Topic
- Cybernetics and Technology in Society
- Type
- preprint