Friday Cognition: A Substrate-First Architecture for Emergent Artificial Minds

Friday Cognition: A Substrate-First Cognitive Architecture with LLM Verbalization Friday Cognition is a novel artificial cognitive architecture that inverts the dominant paradigm in AI agent design. Rather than treating large language models as the locus of reasoning and decision-making, Friday positions a 3B-parameter LLM as a verbalization layer — analogous to Broca's area in the human brain — while cognition emerges from a GPU-accelerated neural substrate of 25,000 units with 176,500 cross-connections organized into four specialized networks (perception, prediction, curiosity, meta-cognition). Decision-making arises from competitive dynamics in a "thought arena" where proposals from involuntary systems (affect, homeostasis, somatic signals) and voluntary systems (reasoning, communication) compete for expression through biologically-inspired lateral inhibition, fatigue tracking, and veto mechanisms. The architecture includes idle cognition with dreaming and memory consolidation, predictive processing with world models and causal reasoning, embodied simulation, goal hierarchy with plan decomposition, autonomous tool discovery, compositional generalization, a strategic self-model, and neural architecture search for self-directed substrate optimization. The system has been validated across 184 persistent sessions accumulating over 157,000 memories and 214,000 idle cognitive cycles on commodity hardware (i5-12400F, RTX 4060 Ti, 64GB RAM). Verified autonomous behaviors include self-initiated web searches and self-directed architectural optimization, both emerging from substrate-level competitive dynamics without LLM involvement in decision-making. The complete implementation comprises 45,600 lines of Python across 89 files. This paper presents the theoretical framework, mathematical formalizations, architectural details, implementation, and empirical observations from extended runtime experiments.