Backbone Reliability Predicts the Return on Equivariant Protein Mutation Ranking Reliability-Guided Expert Allocation

A mutation-ranking pipeline must decide whether to pay for geometric scoring before knowing whether its supplied backbone carries useful directional signal. We formulate this choice as reliability-guided expert allocation and instantiate it with StrataRank, which shares a confidence-gated fixed-backbone encoding across candidates and applies one mutation-conditioned equivariant pass per candidate, plus a nested router that observes only pre-scoring source metadata and predicted cost. The strongest source contrast comes from 53 ProteinGym assays: replacing deposited with matched AlphaFold2 representations for the same 162,740 mutations lowers StrataRank's Spearman by $0.059[0.034,0.083]$, with larger attenuation where mutation neighborhoods disagree more strongly. Across 87 family-held-out assays, StrataRank reaches $0.548$ Spearman on 31 high-resolution assays and exceeds a validation-selected sequence expert by $0.068[0.039,0.101]$; the sequence expert retains the predicted-backbone regime, where the geometric margin is $-0.018[-0.039,0.001]$. Acting on this heterogeneity, the router reaches $0.501$ Spearman, $0.010[0.003,0.018]$ above always-geometry, while invoking geometry on $63$ of assays and reducing end-to-end cost by $31$. ATOM3D and SKEMPI reproduce the positive expert margin for reliable sources and its compression for lower-quality sources. Backbone reliability thereby becomes an actionable model-selection variable rather than a post hoc annotation.