Out-of-area transfer check for the Uruana–SAA manuscript
Retrospective internal specification, established before inspecting the new magnetic prediction errors. Uruana results and archive extents were known. This is not a public preregistration.

Use the 2008 ANP 0032 mag_var grid and historical 1009 PGW residual profiles. Do not retune the 2008 inversion, epoch, 12 km tile lattice, 24 km training-disk radius, 4 km source interval, equivalent depth (6 km), damping (0.03), minimum point count (5), east–west span (4 km), or grid coverage (60% in disk; ≥80% points within 3 km). Keep GPS/survey levels identically for both models. Group by production-line header, altitude and tile. Preserve native profile order; gaps >4 km or discontinuities in native row order break segments. Check the 75 known primary segments separately from their stored point predictions. They are not new tests.

Primary distant set: every historical observation in an eligible segment must lie ≥100 km from Uruana. Its 2008 training disk must not overlap the interior of any training disk previously used in the Uruana–SAA profile-transfer analyses. Require at least 48 km between its center and each of (a) the prior 25 pilot origins, (b) the regional 153-segment model tile centers, and (c) prior fixed / shifted origins in the north-sector study. In addition require every historical point to be at least 1 km from the union of stored prior scored point locations. Use all eligible distant segments. No selection by measured amplitude or error. Check a subset excluding reused line labels. Shared campaigns make this geographic transfer, not a new acquisition or a test outside the SAA. Report modeled SAA contour membership at both survey epochs using tile centers at the surface.

Score both 2008-frozen and 1971.5 projection/induction models using the point-weighted pooled within-segment RMSE; count improving segments and 24 km blocks, compute block-balanced mean squared errors and omit one block at a time. Include untrained zero/flat profile. Additional comparisons: (1) remove a uniform east–west error slope per model estimated solely from the 75 previously scored segments, then score distant data; (2) project observations and predictions onto the orthogonal complement of a constant and linear east-coordinate term separately inside each segment to test whether information survives removal of any line slope: a retrospective shape diagnostic, not a prediction of that nuisance slope. This is also applied to the local 75 for comparison. (3) report individual 48 km blocks and altitude strata to avoid treating nearby points as independent. Analyze projection-only and amplitude-only as secondary mechanism checks. Report absolute RMSE as well as percent changes irrespective of their sign. Flag insufficient geographic or profile support.

WMC is an alternative 1009 historical processing/coverage product. Inspect documentary provenance and support; use it, if documented as total field, as a secondary residual-profile sensitivity by subtracting the historical IGRF at the archived coordinates/height before assessing identical predictions. Treat both products as one underlying acquisition. Do not use WMC where support or channel meaning prevents it.

Archive audit before scoring. Monteiro (CPRM, September 2002), project 1009 Leia-me.pdf pp. 1–2, identifies the PGW coordinates as SAD69 and the WMC magnetic column as bearing an arbitrary unknown addition. Do not interpret WMC as observed total field or subtract IGRF from it. Exclude it from this transfer comparison. Use the IBGE SAD69-to-SIRGAS2000 spatial grid for primary co-registration, with the GRS80/WGS84-compatible geographic axes expressed on the 2008 UTM 23S grid. Compare the same 75 local segment memberships with/without correction, plus the same corrected distant memberships evaluated at unshifted positions as a sensitivity. This addresses reference-frame choice while preserving the v2 comparison.

Follow-up sensitivity after the distant result: resample entire 48 km blocks with replacement (20,000 draws, random seed 20261103). Pool each draw's pointwise sums of squares, then form its RMSE reduction. Repeat for 24 and 96 km blocks and with a half-block shift on both axes. Report descriptive percentile intervals rather than a design-based significance or an independence claim. Also compare the two subsets at their shared archived elevation of 0.95 km. Do not interpret an elevation-stratum contrast as a geological attribution.
