Hector: A Cognitive Architecture for Structural Deliberation via Request-Confirmation Networks

Abstract

Conventional reinforcement learning often yields high-performing agents whose internal deliberative structure is difficult to inspect directly. We present Hector, a cognitive architecture based on Request-Confirmation Networks (ReCoN), designed to study how hierarchical subgoals and planning horizons can emerge from self-organizing symbolic structures. Using chess endgames as a controlled symbolic microcosm, we show that a unified topology containing both KPK (king+pawn vs. king) and KQK (king+queen vs. king) subgraphs shifts control from promotion to checkmate with one-move latency, with the transition observable in internal activations. On a KPK curriculum, Hector achieves a 97.0% win rate, while PPO baselines reach 26.3% (50k timesteps) and 35.9% (200k timesteps). We also report exploratory structural maturation via stem cells and inertia pruning as an extension beyond fixed topologies. We argue that autonomous strategic handover is a minimal operational requirement for deliberative agency: the ability to maintain, suspend, and reallocate control across competing internal models based on global context rather than local reward signals. While we make no claims about phenomenal consciousness, Hector provides a concrete, inspectable mechanism for global control, working memory, and top-down/bottom-up integration - properties central to multiple leading theories of consciousness.