Cross-Library Audit of DFT Magnetic Moments for Curie-Temperature Ranking: An Empirical Residual-Gating Protocol

Reliable magnetic-materials screening requires more than a model that performs well within one computational catalogue. We develop and audit a composition-anchored library residual (CALR) protocol for transferring density-functional-theory (DFT) magnetic-moment information across heterogeneous databases. CALR harmonizes moment percentiles within each library, fits a composition-only ridge anchor, and admits a bounded moment residual only when a nested, composition-disjoint bridge test supports improved rank correlation. AFLOW and JARVIS moments are evaluated against experimental Curie temperatures from NEMAD, with a frozen Materials Project snapshot as a third source. On 484 AFLOW-JARVIS bridge compositions, neither directional gate opens: cross-fitted changes in Spearman correlation are +0.0079 (95% interval −0.0027 to 0.0187) and +0.0060 (−0.0125 to 0.0238). The Materials Project-to-JARVIS bridge gives Δrho = 0.0101 (−0.0036 to 0.0250), also returning the composition anchor. Ungated transfer improves one direction but produces negative transfer in the reverse; CORAL and density-ratio weighting underperform the anchor. A fixed-hash semi-synthetic control shows that CALR can activate for a known transferable residual and close for zero, non-transferable, or reversed residuals, although finite-sample false openings remain. CALR is therefore an auditable diagnostic for cross-library magnetic screening and negative-transfer risk, not formal risk control or validated permanent-magnet discovery. We re-ran the Materials Project analysis with current cell atom counts (nsites), traced every aligned key to a selected DFT identifier and experimental DOI, froze a percentile map on common compounds before gating, held out chemical families from training, resampled element-set groups through the full fit/select pipeline, and compared CALR with same-budget target-domain models. The MP gate remains closed after the nsites correction. We state a single protected deployment estimand: Spearman ρ of the frozen ranking versus experimental Tc on composition-disjoint target-library keys. Structure-resolved and aggregation sensitivities leave the headline |M|–Tc Spearman near 0.43. Family-grouped cross-validation lowers the composition-ridge OOF ρ from 0.697 to 0.461. The 484-key bridge is underpowered for the observed residual (forward 80% power requires Δρ ≈ 0.015–0.020). Materials Project reuses 100% of the NEMAD labels already seen in the AFLOW/JARVIS workflow. CALR still equals the composition ridge on every real direction; we state quantitative conditions under which its extra cost would be justified, and the independent library experiment that would be required to claim a practical benefit.