Outside-contour comparison, retrospective internal specification before scoring this transfer endpoint. The Alaska pair had already been used for a different endpoint (curvature axes). No public preregistration. Retain every eligible AK16 production-line segment. No selection on magnetic values, model errors or their signs. Epochs 1977.542 and 2008.535 approximate mid-campaign dates. The 2008 survey supplied a leveled, IGRF-reduced grid. Use AK16 MAG_DGRF, USE_FLAG=1. Neither residual is reduced twice. For old coordinates use the Alaska NAD27 to WGS84 mainland approximation (EPSG 1176; Clarke 1866 compilation); repeat using coordinates as archived. Cross-check against paired coordinates in the 2008 report. Work entirely in UTM 4N coordinates (km), including the main-field vectors.
Use the Brazilian 12-km lattice spacing, origin (500,7560) km in UTM 4N, 24-km training disk, 4-km source spacing, 6-km source depth, centered training data and source kernel, ridge 0.03 times the largest singular value. Sample each second row and column in the 800-m grid (1.6 km). Require at least 60% footprint coverage, and 80% of evaluation positions within 3 km of training positions. Require >=5 binned median readings spanning >=4 km north in an uninterrupted production-line section, with predominantly northward extent (north extent >=2 times east extent). Split at native header changes, rejected rows, fiducial reversals, jumps >4 km and tile boundaries. Aggregate by 0.25 km along north coordinate within segments; re-evaluate at 1 km. Primary model uses terrain-relative coordinates (6 km below nominal 2008 drape, AK16 height 243.84 m vs nominal 2008 304 m); this flattens relief locally. Additional sensitivity: infer ground from FLTSURF minus HEIGHT, bin at 1 km, interpolate projected ground over the training lattice using linear interpolation inside the historic ground-sample hull and nearest ground outside it; fit to an approximate undulating surface at ground + 304 m and predict at AK16 FLTSURF. No 1977 magnetic value may enter the 2008 fit. Compare fixed, full, projection only, intensity only and flat; center each segment. Report pooled profile RMSE, segments, unique native line labels, spatial results by 24 and 48 km and line clusters. Block ranges descriptive. Compare actual forcing (field rotation, intensity change, and RMS model discrepancy relative to fixed-model error) to the Brazilian site; a polar comparison cannot identify SAA causality. Geometry sensitivity and horizontal-datum alternative must be reported irrespective of outcome.

Post-score extension (exploratory). The median Brazilian rotations were similar. Compute the exact min-to-max rotation among nearby tile centers; retain every distant tile with rotation in that interval. Repeat using only the common 0.95 km level and the local range at that level. Never extrapolate a score when the overlap is empty. Use surface rotations for this cross-site comparison and the altitude-specific vectors for the forecasts. Relate relative RMSE change to the RMS change in prediction and the cosine between the update and the original prediction error. This is an algebraic description of archived paired predictions, not an independent estimate of a new mechanism.

Library-history extension. Inspect the preserved S16 originals and result archive; verify all three SHA-256 hashes, repeat the archived 20-km orientation statistic at every previously eligible center, and retain its original discovery-only status. For the separate metadata screen, evaluate both epoch-specific surface IGRF vectors at the documented intersection centroid for every row in the preserved 60-pair Alaska selection table. Compare to the exact local Brazilian rotation interval; no peak search within footprints and no magnetic-data access for P3. This screen is exploratory and does not represent every Alaska survey.
