Abstract—This study presents the first long-term, production-scale empirical characterization of Secure Reliable Transport (SRT) protocol metrics for Reinforcement Learning (RL)-based adaptive streaming control in low-latency live video applications. Production data spanning 722 hours (approximately 30 days) were collected from an IEEE 802.11ac wireless network, yielding 64,340 aligned multivariate samples. These data were analyzed using statistical methods and machine learningbased anomaly detection. Key findings include: (1) strong positive correlation between average Round-Trip Time (RTT) and jitter (r = 0.912, 95% CI: 0.909–0.915), indicating coupled latency dynamics suitable for state-space reduction; (2) Isolation Forest detected 4.8× more anomalies than the Interquartile Range (IQR) method (6,425 vs. 1,339), demonstrating superior multivariate pattern recognition; (3) three distinct anomaly regimes were identified: congestiondriven, timing-driven, and bandwidth-limited; (4) a datadriven three-level operational monitoring framework (Green/Yellow/Red) was derived for five SRT metrics using convergent statistical evidence; and (5) comprehensive RL validation demonstrated effective reward signal translation: rule-based (+82%) and Q-learning (+54%) controllers outperformed static baselines at hourly resolution; at 5- minute resolution, Q-learning achieved +138.1% improvement with 98% oscillation reduction; in closed-loop emulation, Q-learning achieved +81.3% improvement with minimal red-zone violations (0.23%). These findings establish empirical foundations for future RL-based SRT controllers, bridging network measurement research and adaptive streaming development. Limitations regarding single-network generalizability and recalibration for heterogeneous environments are discussed.