A Proposal for an Architectural Definition of Intelligence

Performance definitions of intelligence can tell us what a system accomplishes: whether it solves problems, pursues goals, or succeeds across environments. Bernard Jennings asks them to answer a question they were not built to settle. What internal organization would make those performances instances of intelligence rather than capable execution?

His working paper, The Architecture of Intelligence: Definition, Spectrum, and Ceiling, answers with a revisionary definition.¹ Intelligence is the capacity of a system to resolve its own coherence demands while producing behavioural residue beyond what its stable competence plus input would predict. The definition has three operational components. A system must register a coherence demand as its own, own the resolution process, and produce the residue. Randomness, historical novelty, and computational complexity do not count as residue by themselves.

This is a more careful proposal than “architecture instead of behaviour” suggests. Jennings connects a claimed internal architecture to a behavioural signature. When the architecture cannot be fully observed, applying the criterion remains interpretive and empirical work.

What the paper claims, and what it leaves open

On its own terms, the framework places chess engines, classical optimization systems, and language models before the newer reasoning architectures among the settled negative cases at inference. Their stable competence and input are sufficient, in Jennings’ account, to explain the output. He treats current reasoning models with inference-time loops, recursive architectures, and multimodal foundation models differently. His present assessment is still negative, but the verdict is contested because the architectural evidence is incomplete.

That evidential status matters. Zenodo identifies the document as a version 01 working paper. It specifies a theoretical criterion and a programme of possible falsification; it does not itself demonstrate residue empirically in the disputed systems. Jennings explicitly assigns that work to AI research, cognitive science, and the other disciplines he addresses.

The ceiling claim carries a further dependency. The eight Coherence Resolution Modes and the case for their completeness are derived in companion work that this paper summarizes and inherits. A determinate ceiling follows only if those modes exhaust the resolution space. The companion essay on Jennings’ ceiling examines that completeness problem in depth. Repeating it here would obscure a different issue that arises before any verdict on the ceiling.

Intelligence does not confer authority

Jennings is asking what properties a system must have to count as intelligent in his architectural sense. Intelligence Architecture asks how a system output becomes a signal, recommendation, authorization, or trigger inside an organization. Those are two different architecture questions.

They can vary independently. A simple classifier may fall outside Jennings’ intelligence spectrum yet exercise considerable organizational power when its threshold automatically removes cases from review. A system that met all three of his conditions could still be assigned an advisory role with no authority to release an action. Nothing about a system’s internal status creates a decision right or obliges another actor to follow its output.

The NIST AI Risk Management Framework offers a useful normative counterframe.² It does not allocate power according to a machine’s intelligence status. The voluntary framework calls for clear risk-management roles and lines of communication (GOVERN 2.1), executive responsibility for decisions about development and deployment risks (GOVERN 2.3), and differentiated responsibilities for human-AI configurations and oversight (GOVERN 3.2). In operation, it includes processes by which end users and affected communities can report problems and appeal outcomes (MEASURE 3.3), as well as mechanisms to override, supersede, disengage, or deactivate systems (MANAGE 2.4 and 4.1).

The framework is guidance, not evidence that these arrangements will work in every institution. Its contribution to this comparison is normative. Appeal, override, and deactivation are organized powers and duties. System design can make them technically possible, but the system cannot grant them to itself.

Reading a deployment on two axes

A responsible assessment therefore keeps two axes visible. The first concerns the system: what operations it performs, what evidence supports an attribution of intelligence, and where the verdict remains unsettled. Jennings supplies a demanding theoretical proposal for that inquiry.

The second concerns the decision arrangement. Who defines the purpose? Who may release an action based on the output? Who has the time, information, and authority to override it? How can an objection still reach the decision, and who bears the consequences? Whether a system is described by a strong or weak theory of intelligence does not answer these questions. They turn on the authority an organization actually attaches to its output.

Footnotes

1. Bernard Jennings, “The Architecture of Intelligence: Definition, Spectrum, and Ceiling”, Zenodo working paper, version 01, 22 May 2026. The record identifies the publication as a working paper. The manuscript describes its definition as revisionary, inherits the eight modes from the companion paper Coherence Resolution Modes (CRM) in Autonomous Systems, and leaves empirical tests of its criteria to future research. ↩

2. National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework (AI RMF 1.0), January 2023. The framework is voluntary, non-sector-specific, and use-case agnostic. The cited subcategories are normative risk-management outcomes, not an effectiveness study. ↩