Does Network Structure Matter More Than Model Capability in Multi-LLM Systems? A Controlled Experiment on Adaptive vs. Fixed Topologies

We investigate whether structural change in a multi-LLM network produces performance gains independent of individual model capability. Two three-node networks, both using identical models (qwen2.5:3b via Ollama) on identical hardware, were compared on a world-consistency contradiction detection task. Experiment A used a fixed sequential topology. Experiment B used an adaptive topology in which connection weights (flow_weights) — termed Adaptive Artificial Synapses (AAS) — were updated after each outcome. The adaptive network outperformed the fixed network in all five trials. Mean accuracy difference on questions 51–100: +17.6% (t(4) = 9.60, p = 0.0007, Cohen's d = 4.29, 95% CI: [+0.117, +0.235]). All three pre-defined significance criteria were met. This suggests that connection dynamics — independent of node-level capability — contribute meaningfully to system-level performance. Code and data available at: https://github.com/piperendervt-glitch/sdnd-proofThis work is governed by the safety principles documented in constitution.md, available at: https://github.com/piperendervt-glitch/sdnd-proof