Unsupervised Detection of Speculative Regimes and the Empirical Audit of Agentic Responsibility in Financial Markets

The rise of autonomous trading agents powered by Large Language Models introduces new systemic risks to financial markets through synchronized decision-making and emergent herding behavior. We introduce the Agentic Regime Auditor, an unsupervised framework for real-time detection of transitions between "Responsible" and "Speculative" market states driven by algorithmic homogeneity. By combining Gaussian Mixture Models with the Hurst exponent---a measure of long-term memory and fractal efficiency---we identify distinct behavioral regimes without requiring labeled training data. Our empirical analysis of BTC-USD and SPY markets from 2024--2026 reveals three key findings: (1) speculative agentic regimes exhibit statistically distinct roughness and heavy-tailed returns, validating the existence of identifiable "unsafe" states; (2) a Fragility Gap emerges where Bitcoin shows 4x higher regime entry probability (P = 0.04) compared to regulated markets (P = 0.01$), while traditional markets exhibit greater persistence once destabilized; (3) these transitions are computationally tractable for deployment as regulatory circuit breakers. Our results demonstrate that fractal-volatility phase space clustering can effectively audit collective agentic behavior, enabling regime-adaptive governance mechanisms for AI-driven financial systems.