Materials foundation models fail together: Correlated errors and collective blind spots

Universal machine-learning interatomic potentials (uMLIPs) are increasingly combined to screen candidate crystals, but the reliability gains from consensus and disagreement depend on diversity among model errors. We quantify error dependence across 15 public uMLIPs using density functional theory (DFT) convex-hull distances for 248,623 structures from the Wang–Botti–Marques benchmark. The mean pairwise error correlation is 0.39, corresponding to 4.3 effective models. Because rank-1 removal cannot distinguish shared model bias from shared reference effects, we applied two additive-reference controls; correlations remained 0.44 and 0.42, with effective model counts of 3.7 and 3.9. A stratified permutation null that preserves each model’s coupling to the DFT reference while destroying cross-model correlation showed that false discoveries shared by single-model, consensus, and agreement-based screening occurred 141 times more often than expected ( p < 0.001 ). Agreement-gated screening admitted 1620 false discoveries versus 7 under the null. The median consensus did not outperform the best individual model and produced a 27.6% false-discovery rate among predicted-stable structures. Cross-model disagreement was nearly uninformative (area under the receiver operating characteristic curve, AUC, 0.51), whereas the consensus margin reached AUC 0.75 and conformal-normalized margin 0.85; standard split-conformal intervals reached 0.68. Removing the 30% of candidates closest to the threshold reduced the retained false-discovery rate to 17.8%, but was exactly equivalent to tightening the stability threshold at matched shortlist size. These results show that correlated errors can limit ensemble reliability and that threshold-aware scores provide a practical ranking of screening risk. We release a reproducible reliability atlas and benchmark.