Jennings and the Machine's Ceiling

Before considering the argument, fix its status. Bernard Jennings released The Architecture of Intelligence: Definition, Spectrum, and Ceiling on 22 May 2026 as an independent working paper, version 01, through Zenodo; an SSRN listing followed. It is very recent, not peer-reviewed, and not a canonical account of intelligence. The precision of its terminology is a feature of a proposed framework, not evidence of scientific consensus.¹

The proposal can still be tested against a hard case. In 2018, David Silver and colleagues reported AlphaZero in Science. A general reinforcement-learning algorithm was trained separately through self-play for chess, shogi, and Go. The systems began from random play and, according to the paper, received no domain-specific knowledge beyond the rules. A policy-and-value neural network worked with a general-purpose Monte Carlo tree search. The resulting systems defeated leading programs across the three games.²

AlphaZero is unquestionably capable at the task. Jennings’s question is whether capability of this kind should count as intelligence at all.

The proposed test

In Jennings’s framework, a system belongs on the intelligence spectrum only when three components coincide.

The first is demand-ownership. A task arriving from outside is not enough. The system must register its own pattern gaps, conflicts, or inefficiencies as demands requiring resolution. Jennings locates that registration in what he calls “Category 2 emotional architecture”: internal state-change and control signals. In this paper, the term is functional and does not assert phenomenal feeling. The proposed architecture itself remains part of the theory to be evaluated.

The second is resolution-ownership. The relevant resolution must be generated internally rather than retrieved or produced by simply executing an externally specified path. Jennings assesses this demand by demand and permits degrees. A system could own some resolutions while merely executing others.

The third is residue beyond encoding, intended as the behavioral handle on the first two. Encoding means the system’s stable competence, called layer 1. It includes built structure, parameters, learned dispositions consolidated through training, and the configured inference or operating procedure. Residue is present when layer 1 plus the input does not fully predict the system’s behavior in detecting and resolving a coherence demand.

This definition deliberately excludes three easy substitutes. Random variation is not residue when the stochastic mechanism is part of layer 1. Historical novelty is not residue when the stable procedure determines the new result. Computational difficulty is not residue merely because prediction would be expensive.

The paper thereby creates a substantial empirical burden. Jennings acknowledges that complete specification of layer 1 is often unavailable in practice and that applying the criterion requires interpretive work. More importantly, he states that the paper does not empirically demonstrate residue in any case. It specifies the question and offers theoretical classifications; empirical work is left to the relevant disciplines.¹

Running AlphaZero through the criteria

Start with demand-ownership. Silver et al. document an externally supplied game, its rules, and a learning signal tied to win, loss, or draw. Their study does not investigate whether AlphaZero registers its own pattern conflicts as demands in Jennings’s sense. The supplied objective is therefore not evidence of demand-ownership. But the AlphaZero paper was not designed as a direct empirical test of Jennings’s proposed internal architecture either.

Resolution-ownership is harder to read off the system. Self-play, learning, and search generate moves without retrieving them from a database of human games. In ordinary language that can look like independent problem solving. Jennings demands something narrower: a resolution arising from a demand the system registered as its own, rather than from execution of stable competence. Playing strength and generativity do not measure that property, so the Science results cannot settle it.

On residue, Jennings does settle the case within his own framework. At inference, he assigns the trained policy and value network and the configured search procedure to layer 1; the board position is the input. He therefore classifies AlphaZero as having zero residue and places it outside the intelligence spectrum. Training, in his account, produces layer 1 rather than constituting the live layer that could generate residue.

That is Jennings’s framework verdict, not a finding reported by Silver et al. Novel-looking moves, stochastic exploration, and expensive search would not satisfy his definition. Yet the working paper also performs no empirical comparison between a complete layer-1 prediction and AlphaZero’s observed behavior. The example shows what the definition entails. It does not independently validate the proposed boundary.

Where the ceiling gets its support

Jennings argues for more than a floor. His ceiling is a system that can integrate and fully calibrate a complete set of eight Coherence Resolution Modes.

The intelligence paper does not independently derive the completeness of that set. It inherits the eight modes from Jennings’s companion working paper on Coherence Resolution Modes and makes the ceiling explicitly conditional on that argument.³ An additional constraint type, another resolution mechanism, or a further admissible combination would defeat the completeness claim.

This is commendably falsifiable, but the number eight does not certify itself. A typology may be finite because its primitives were defined to produce a finite set; whether those primitives cover the phenomenon remains open. The companion and main paper also belong to the same young research program. They are not independent replications of one another.

The ceiling is therefore a structured hypothesis with named failure conditions. It is not an empirically established upper limit on intelligence.

Capacity, purpose, and accountability

Only a modest part of the framework needs to travel into organizational practice: operational capacity should not be confused with purpose or accountability.

AlphaZero demonstrates how much performance can grow under a narrow objective. Training, search, and compute increase the range and speed with which a system can evaluate action inside the game. They do not decide why the game should be played, which consequences outside its rules matter, or who is accountable for deployment.

That is an editorial inference for Intelligence Architecture, not an empirical result established by Jennings. It yields four separate questions for an organizational system:

• Which capability and throughput does the technology increase?
• Who sets the objective, boundaries, and permitted means?
• Who may use or stop an output as a premise for action?
• Who observes consequences and owns changes to the arrangement?

A system can sit outside Jennings’s intelligence spectrum and still be immensely useful or dangerous to an organization. Conversely, satisfying his three criteria would not legitimize a deployment. A classification of intelligence cannot replace evidence of effect, decision rights, or responsibility.

The paper is most useful with its provisional status attached. It offers a provocative definition, explicit routes to falsification, and a disciplined separation between fast execution and ownership of a demand. AlphaZero makes that separation vivid. It also exposes the remaining evidential gap: assigning a famous system zero residue in a theory is not the same as empirically establishing where the proposed ceiling lies.

Footnotes

1. Bernard Jennings, The Architecture of Intelligence: Definition, Spectrum, and Ceiling, version 01, independent working paper, 22 May 2026. The publication status comes from the Zenodo record and the document; neither identifies a peer-reviewed publication. ↩ ↩²

2. David Silver et al., “A General Reinforcement Learning Algorithm That Masters Chess, Shogi, and Go Through Self-Play”, Science 362 (2018), 1140–1144. The earlier open preprint describes the chess/shogi version of the approach. ↩

3. Bernard Jennings, Coherence Resolution Modes in Autonomous Systems, working paper, 2026. The main paper expressly makes its ceiling conditional on the completeness argument attributed to this companion work. ↩