The degree to which machine learning synthesizability scores align with ground-truth thermodynamic metrics (convex hull energies, selectivity scores), measuring whether ML models correctly estimate the likelihood that a hypothetical material can be experimentally synthesized.