Architectural Disagreement as a Detector of Representation Ambiguity in Molecular Property Prediction

We propose that the absolute predictive disagreement D(m) between two Graph Neural Networks with different inductive biases is a robust, generalizable, and mechanistically interpretable detector of representation ambiguity in molecular property prediction — the Dark Chemical Space. Across ten biological targets including EGFR, BRAF, JAK2, VEGFR2, CDK2, and DRD2, D(m) outperforms MC Dropout (rho=+0.031, p=0.29, non-significant) and Deep Ensemble (2x weaker) as an uncertainty estimator on scaffold-split data. The signal is stable across training seeds (std < 0.05), mechanistically localized in the prediction head, and generalizes across 2D and 3D architectures. On BRAF, a combined selection criterion achieves 94% precision at top-50 candidates. On EGFR, 16 underestimated actives are recovered that standard models deprioritize by up to 100-fold in IC50. The Strange Molecule Hunter v2, a reusable Python tool, converts D(m) into actionable candidate prioritization. All results are reproducible across six open Kaggle notebooks, including explicit retraction of a mechanistic claim from NB4 that did not survive multi-seed validation.