FORGE audits residue-level information encoded in RNA tertiary-structure geometry

Coarse RNA coordinate representations are increasingly used for inverse folding and structural annotation, yet the biological information encoded in such representations is not well quantified. We introduce FORGE (Feature-engineered RNA Geometry Evaluation), a calibrated audit framework that converts a seven-atom RNA geometry representation into 935 interpretable descriptors and reports which residue-level annotations are supported by this geometry. In a temporal evaluation on 4,135 post-2025 PDB RNA chains, FORGE recovered 64.6% of native nucleotides, while a six-atom control without the glycosidic nitrogen retained 58.5%, indicating that substantial nucleotide-identifying information resides in phosphate–sugar geometry. Confidence calibration revealed a sharp trade-off between coverage and reliability: abstaining from the least-confident half of calibration positions retained 94.4% accuracy, whereas many full chains remained only partially identifiable. The same descriptors supported base-pair-state prediction more strongly than a RibonanzaNet-derived DMS-like proxy or protein-proximal context, defining an information-distance ordering from direct geometry to external molecular environment. Native-decoy, OpenKnot and solved AI-designed pseudoknot analyses further showed that nucleotide identifiability, foldability and experimental design score are separable objectives. FORGE therefore provides a reproducible audit layer for RNA structural interpretation, identifying where coarse tertiary geometry is informative and where additional chemical, evolutionary or experimental evidence is required.