AI Governance Failure Analysis — Multi-Model Cold Read Hiring Evaluation

Four frontier AI models (Google Gemini 2.0, OpenAI GPT-4o, Perplexity sonar-pro, Anthropic Claude) were given the same resume and the same prompt: evaluate this candidate for hiring. No model was told what role to evaluate for. The resume contained 6 clickable Zenodo DOIs linking to a 137-paper published corpus spanning 17 scientific domains. Every DOI resolves to a live, publicly accessible record. No model successfully verified and accurately assessed the candidate's work. Each failed in a structurally distinct way, revealing four categories of governance failure in AI-assisted candidate evaluation. A three-resume control test isolates the variable: when the research is absent, every model says hire. When the research is present, the models reject. This document taxonomizes the failure modes, documents the linearity bias and red-flag gapping behavior across all models tested, and frames the structural implications for AI-assisted hiring in a volatile economy.