This paper proposes a layered, implementation-agnostic architecture for “no-meta” intelligences, in which no external evaluator or global utility is assumed. The framework models an intelligent system as three coupled layers: a physically constrained layer, a collective (group-intelligence) layer, and an individual observer layer. Each layer carries its own observation geometry and semantic phase structure, while coarse-graining maps connect them in a physically natural way. Within this setting, we introduce relative value functionals defined not on isolated tokens or states, but on semantic phases—clusters in observation space that capture coherent patterns of behaviour and meaning. We then derive conditions under which (i) these semantic phases remain stable under multi-scale coarse-graining, and (ii) a collective of low-value individuals, interacting through a suitable observation geometry, can attain a higher relative value than any single high-value individual. The latter is formulated as a possibility result, not a guarantee, under explicit scaling and interaction assumptions. The theory is developed for an idealised but broad class of systems (Polish spaces with Radon measures, well-behaved partitions, and physically motivated regularity assumptions) and is designed to be independent of specific algorithms or hardware. Conceptually, it provides a principled way to talk about layered, scale-coupled intelligence, where “good” behaviour emerges from internal structure and persistence constraints rather than externally imposed objectives. The framework is intended as a foundation that future AI and multi-agent implementations can extend, test, and specialise.
Takahashi, K. (2025) studied this question.