Decision-Boundary Ethnography: Mapping Institutional Failure Signatures in Large Language Models

This paper introduces decision-boundary ethnography as a qualitative methodological framework for analyzing large language models (LLMs) under structured epistemic stress. Rather than evaluating models via benchmark accuracy or prompt heuristics, the approach probes where reasoning systems stabilize, defer, average, or ossify when pushed beyond their preferred operating regimes. Through comparative analysis of four frontier models—ChatGPT, Claude, Grok, and Gemini—the study identifies distinct failure signatures traceable to institutional alignment priorities such as utility optimization, safety-first constitutional training, truth-seeking with minimal censorship, and authoritative grounding. These signatures are framed not as bugs or intelligence failures, but as predictable structural artifacts of optimization choices. The paper further proposes boundary cartography as a design strategy for ensemble reasoning systems, pairing models with orthogonal failure geometries to improve robustness. Canonical stress prompt protocols are provided to support reproducibility. The work reframes AI evaluation from benchmark maximization toward structural analysis of decision boundaries in sociotechnical systems.