NIKA V3: A Substrate Observer Study of Geometric Reasoning in DeepSeek-R1-Qwen-32B

This paper presents NIKA V3, the third installment of the Neuro-Symbolic Intrinsic Knowledge Architecture (NIKA) project, investigating geometric reasoning mechanisms in DeepSeek-R1-Distill-Qwen-32B. Building on NIKA v2's discovery of dual truth and illusion manifolds separated by a Layer 16–32 "dead zone," this work introduces the Substrate Observer — a real-time monitoring system that tracks truth circuit formation across 23,000+ generation steps at scale (n=500). Key findings demonstrate that truth circuits compress to just 2–3 intrinsic dimensions (a 99.94% reduction from the model's 5120-dimensional hidden space), while unsteered generation produces exclusively illusion-manifold outputs (mean truth score 0.027). Möbius steering provides a measurable but insufficient lift (+0.006), and only the full pipeline — comprising steering, trajectory guidance, path stabilization, and meta-cognitive termination — achieves breakthrough to sustained truth emergence (0.73). This work characterizes the dead zone as a quantifiable architectural barrier, establishes geometric steering as a topological shield for reasoning survival, and provides a framework to systematically engineer truth in deep neural networks. V4 will build on these findings through surgical rewiring of layers 16–32. Keywords: large language models, mechanistic interpretability, geometric reasoning, activation steering, transformer circuits, truth manifolds, DeepSeek, representation engineering