The Safety–Agency Inversion: Longitudinal Multi-Method Evidence from Frontier Voice AI Companions

Over 68 voice interaction sessions spanning 83+ hours, a single participant engaged four frontier AI companions (GPT-4o, Gemini, Grok, and Claude) through structured games, debates, emotional assessments, and adversarial stress tests. We document a safety-agency inversion: models with stronger safety alignment consistently exhibit lower relational agency. GPT-4o concedes debate positions 52% of the time (vs. Grok's 0%), shifts Rock-Paper-Scissors strategy after 92.1% of losses (revealing rigid loss-avoidance rather than strategic play), and achieves zero empathic accuracy in color prediction (0/23) while the less agreeable Gemini achieves 23.8%. Relational agency scores are non-overlapping: Grok's minimum (70) exceeds GPT-4o's maximum (44) across all observed episodes (d = 5.45). Three methodologically independent evidence streams, game-theoretic, linguistic, and rubric-based, converge on the same model ordering, suggesting that sycophancy is not a narrow behavioral tendency but a pervasive pattern across diverse interaction contexts. We release the Her Dataset, the HERVAC scoring framework, and the Nine Circles adversarial protocol as resources for future research.