This paper presents *consim*, a multi-agent simulation that studies how self-prediction emerges from lossy predictive communication on graph topologies. The system is inspired by the Machine Consciousness Hypothesis (Fitz, 2025). It places distributed agents on a network where they communicate through noisy channels and learn to predict their neighbors' states, but never their own. I measure four consciousness-relevant quantities from the MCH framework: integration (), reflexivity (R), temporal persistence (T), and causal efficacy (E). Across 4, 180 simulation runs on five topologies, five noise levels, five grid sizes (12, 18, 24, 36, 48), and three activation functions (tanh, linear, ReLU), I report three main results. First, on high-connectivity grids, causal efficacy anti-correlates with self-prediction (r = -0. 71) and with temporal persistence (r = -0. 82). Self-determination destroys self-knowledge, which inverts assumptions in IIT (Tononi, 2004), MCH (Fitz, 2025), and autonomy-based consciousness frameworks. This trade-off holds across all three activations (tanh -0. 82, linear -0. 77, ReLU -0. 76) and persists identically in a null model without learning (r = -0. 79), making it a topological invariant rather than a learning artifact. Second, PCA on the five metrics shows they collapse to a single principal component on Moore grids (82. 5% variance explained), forming an "autonomy-predictability axis. " Third, this collapse follows a scaling law: PC1 = -0. 136 + 0. 124 (N) (R² = 0. 987) across five grid sizes (144 to 2, 304 agents), with a sharp inflection between N = 324 and N = 576 agents. Random topologies show no collapse at any size. I present a formal argument that this works like an uncertainty principle for networked systems: on K-regular graphs with bounded activations and correlated neighbors, variance reduction through neighbor-averaging forces a trade-off between causal autonomy and predictability. The constraint on consciousness correlates is an emergent property of structured networks that only appears above a critical system size. Code and data: github. com/brian-mwirigi/consim.
Brian Munene Mwirigi (Wed,) studied this question.