OUROBOROS-AGI: A Candidate Organizational Architecture for Recursive Multi-Agent Reasoning

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Large language model capability is commonly discussed in terms of model scale, training data, inference-time computation, context length, tools, and sampling. A separate question is whether the organization of inference itself can produce measurable effects when those resources are controlled.This pre-experimental architecture note introduces OUROBOROS-AGI, a protocol family for testing whether structured organization of distributed LLM inference—particularly typed persistent epistemic state, explicit evidence-governed state transitions, and state-dependent re-entry—produces measurable effects beyond strong resource-matched sampling, refinement, verification, and aggregation baselines.OUROBOROS is not presented as a demonstrated AGI architecture, a new theory of intelligence, or an empirically superior multi-agent system. Its components have substantial precedent in blackboard architectures, truth-maintenance systems, argumentation frameworks, self-refinement, multi-agent debate, role-based agent systems, mixture-of-agents approaches, and structured deliberation systems.The central experimental challenge is therefore not whether multi-agent organization can be constructed, but whether its proposed mechanisms produce effects that survive strong competing explanations.The framework treats minority-hypothesis retention, role structure, communication topology, arbitration structure, and state-dependent re-entry as experimentally removable or variable components. Its primary falsifier, F*, is the strongest reasonable non-social allocation of matched models, tools, and inference resources to independent sampling, isolated refinement, verification, and aggregation. For each experiment, F* must be specified prospectively rather than reconstructed after outcomes are known.The manuscript also defines counterfeit-success conditions intended to distinguish genuine organizational effects from additional computation, prompting, verification, human intervention, representation advantages, search, refinement depth, selection effects, context truncation, minority-recovery artifacts, and status-label scaffolding.No performance advantage is claimed. The purpose of this note is prospective: to timestamp the candidate architecture, competing explanations, falsification conditions, provenance boundary, and pre-experimental commitments before associated empirical results are used as evidence for the architecture.The central question is:Does this organization of inference buy anything that an equally resourced, intelligently designed non-social inference process does not?