Aegis: Uncertainty-Aware Governance for AI-Generated Signals

We introduce Aegis, a two-layer governance framework that makes AI-generated financial signals risk-aware and auditable. The framework integrates a pluggable Signal Generator with an Intelligent Gatekeeper that filters outputs using model-derived uncertainty quantification (UQ) and a volatility-based regime proxy. In backtests on Taiwan equities (2003--2025), Aegis reduced maximum drawdown from 34% to 4% and improved Sharpe from 1.3 to 1.7. By exploiting the monotonic relation between interval width and error, the Gatekeeper converts UQ into transparent inclusion rules without requiring exact calibration. As a simple, reproducible, and extensible framework, Aegis establishes a baseline verifiable risk agent for future generative-finance systems.