Abstract
This paper specifies the Planetary Health Axis System (PHAS), An active learning framework integrating data, large language models, and human experts for planetary health. The system is a pipeline of eight stages: (1) leakage-free preprocessing, mixture-of-experts imputation, and credibility-weighted data integration; (2) extraction of a dynamic, context-dependent knowledge prior from the scientific literature by a domain-tuned large language model; (3) nIDE non-invertible, asymmetric dimensionality expansion with median-kernel aggregation; (4) SCGL conditional structural graph learning on the lifted representation under a generalized-sparse, locally-dense, hub-mediated structure with acyclicity; (5) an active-learning, human-in-the-loop closure that converts learned structure into staged, verifiable queries; (6) non-discounted long-run-average optimal control subject to generalized planetary-boundary constraints, yielding KKT shadow prices and propagated top-level weights; (7) numerical solution by relative-value or policy iteration, with evaluation of the current state against the constrained optimum; and (8) full-system path/flux attribution together with unconditional and conditional distributional extrapolation. Together, the eight stages form a coupled estimation-and-control procedure that converts heterogeneous planetary-health evidence into verifiable structure and, ultimately, into boundary-respecting, decision-facing policies.