Within the complex European Air Traffic Management (ATM) network, Letters of Agreement (LOAs) are critical regulatory instruments primarily put in place to ensure operational safety during tactical sector handovers. However, as the industry transitions toward Trajectory-Based Operations (TBO), these vital safety documents remain trapped in publicly unavailable text formats, masking systemic flight inefficiencies and hindering compliance automation. To bridge this gap, this paper introduces a data-driven, reproducible trajectory processing pipeline designed to systematically reverse-engineer LOA constraints directly from historical flight data. The methodology employs a two-stage framework leveraging both classification and regression layers. Target entry/exit clusters, isolated via an unsupervised Hierarchical DB-SCAN (HDBSCAN) framework, are first modeled using an Extreme Gradient Boosting (XGBoost) classifier. The resulting predictions feed into a decoupled ensemble of multi-target XGBoost regressors to evaluate positional and efficiency targets. Model performance is validated against a 20% test split, and the final outputs are interpreted through game-theoretic SHapley Additive exPlanations (SHAP) analysis and a rule extractor to isolate primary operational drivers. Evaluated over a full 28-day AIRAC cycle, the architecture achieves exceptional spatial coordinate projection accuracy, demonstrating high 𝑅2 scores of 0.933 for latitude and 0.951 for longitude at sector entry points. The empirical results successfully reconstruct 95% confidence spatial boundary ellipsoids and identify (un)written hidden rules routinely executed by air traffic controllers. Ultimately, these discovered constraints are integrated into an interactive, open-source visualization dashboard, establishing a scalable digital twin framework to support real-time safety monitoring, flight plan compliance verification, and data-driven airspace redesign simulations. ... Read more