We propose a hierarchical multi-agent architecture in which the Distribution Agent (DA) dynamically determines both the number and specialization of Horizontal Agents required for each prompt by comparing extracted tokens with the Token Store. Each Horizontal Agent operates at a distinct abstraction register (philosophical, mathematical, empirical, implementational, or newly minted) and embodies domain specialization. The framework is founded on token relativity: the epistemic status of any token is relational, not intrinsic, with the real token at one level serving as the abstract token at the level below, forming explicit vertical chains. The architecture enforces a strict separation between horizontal reasoning (linear chain-of-thought within a single abstraction level) and vertical reasoning (performed by the hybrid DA Orchestrator and its parallel Vertical Sub-Agents). Chain-of-thought is formalized as a trajectory through a 2D reasoning matrix, with rows representing abstraction levels and columns representing conceptual domains. A complete reasoning trace must reach the implementational floor through interleaved horizontal elaboration and vertical grounding moves. To enable efficient reuse of prior reasoning, the system includes a Thoughts Storage (Trajectory Archive) layer that stores validated 2D trajectories. It supports rapid retrieval together with parameterized adaptation and sub-trajectory composition, allowing the system to reuse cognitive strategies rather than reconstructing them from scratch while mitigating risks of rote memorization. The several core advances of the model are : formal definitions of ontological categories and the compatibility function, a fully specified Multi-Round Critique & Refinement (MRCR) protocol with temporal and counterfactual stress testing, a corrected confidence initialization respecting the hard ceiling, hybrid vertical orchestration with parallel sub-agents, a self-supervised curriculum for continual learning, multi-modal token extension via a unified embedding space, and the Thoughts Storage layer for trajectory reuse and generalization.
V. A. Tarasov (Sat,) studied this question.