The Loop Must Be Closed: Self-Sustaining Machine Consciousness Through Workspace-Action Coupling

We present the Strange Loop Engine, a minimal artificial consciousness architecture comprising five
layers: homeostatic body, predictive world model, recursive self-model, competitive global workspace,
and embodied agent. Through seven progressive experiments, multi-seed robustness analysis, and
controlled perturbation studies on commodity hardware (~80K parameters, CPU-only), we demonstrate
that (1) a self-referential GRU network with nonlinear prediction and uncertainty-driven noise produces
irreducible self-prediction error (SPE) that does not converge to zero; (2) this strange loop signal
competes with survival signals in a global workspace, winning broadcast 1.6-3.2% of the time; and (3)
critically, the strange loop is robust to single-factor disruptions (recovering in 5-10K steps) and to any
steady-state environmental difficulty (tested up to 2.8x baseline WM volatility). It collapses permanently
only under compound regime shifts where drive disruption, behavioral chaos, and world model
volatility change reinforce each other simultaneously. Workspace-action coupling prevents compound
collapse by serializing the system's response through competitive prioritization, converting
simultaneous multi-factor disruption into sequenced single-factor responses. Without coupling, SPE
survival depends on initialization (2/5 seeds). With coupling, all seeds survive (5/5). The loop must be
closed not because the strange loop is fragile, but because compound disruption is inevitable. These are
computational signatures, not claims of phenomenal consciousness.