"This work provides the first publicly verifiable, hidden-ground-truth benchmark suite for LLM-free cognitive architectures at million-agent scale. We present an integrated, modular cognitive architecture composed of the NeuroCogniSwarm (NCS) brain, a seven-faculty CAF-MAS suite operating on a million-agent swarm substrate, and the 19,553-line Ultra-Elite AGI Monolith, which provides global-workspace broadcasting, hierarchical planning, multi-agent debate, and meta-cognitive oversight. The system contains no large language model; all reasoning, planning, and decision-making are performed by the cognitive modules themselves. We evaluate the architecture under a methodology designed to resist self-deception: hidden ground truth, multi-seed reproducibility, and held-out generalization testing of candidate improvements, and report results that survive that methodology, including those that fall short of ceiling. The architecture attains perfect scores on a 25-task compositional and metacognitive battery (THE CRUCIBLE, 25/25), a 60-point fifteen-faculty stress test including derived second-order theory of mind (60/60), and strong performance across twenty randomized seeds on five of seven autonomous cognitive faculties, with four at ceiling (100%). Causal discovery and fallacy detection are measured at 45% and 40% respectively, with the algorithmic causes of the remaining gaps characterized. The combination of breadth, hidden-ground-truth verification, and disclosed limitations is the central contribution: a precise, faculty-by-faculty measurement of proximity to general intelligence.
Mohamed Salih (2026) studied this question.