Pre-screening solid-electrolyte candidates before NEB/DFT: a reproducible benchmark and its documented limits

Pre-screening pipeline that ranks solid-electrolyte candidates before NEB/DFT: Materials Project structures (2026.09.28), CHGNet NEB barriers, 3D percolation of the mobile sublattice, distance to the hull. On a 15-material benchmark, 5/5 known superionics are surfaced and 3/3 poor conductors rejected. NEW IN V2: 101 published room-temperature ionic conductivities (OBELiX, one DOI per value) are joined to the computed labels without recomputing or fitting anything; on the 17 comparable compositions the rule classifies 12 correctly (6 good conductors surfaced, 6 of 11 insulators rejected) and the 5 errors are named. The reported accuracy is 0.706, an order of magnitude not a precision, and the benchmark-only 6/7 is labelled optimistic because it has no true negative. Between-potential spread: bias +0.016 eV, MAE 0.066 eV, Spearman 0.944, 1 verdict flip in 12. Labels below 0.10 eV are flagged low_barrier_unverified because a second potential does not separate the measured false positives there (0 of 4 rejected). Surrogate pre-ranker: Spearman 0.655 +/- 0.021 (n=260 labels, 5-fold, 5 seeds, out-of-fold). Sodium extension: 3/4 known Na superionics measurable by NEB+MD; the real beta-alumina cell is now ordered by the pipeline (Na6Al66O102), which transfers the existing <= 0.095 eV bound to a pipeline-derived model without recomputation -- still not a 4/4 claim. Limits are stated explicitly: machine-learned proxies rather than DFT, one hop per material, no conductivity measured by this work.