Version 5 | Matching the historical field rotation outside the South Atlantic weak-field contour
The new global selection uses the existing historical survey epochs rather than the year of a report or publication. Read PROTOCOL.txt for the sequence of exploratory decisions. Observed error scores use the 75 nearby and 1,528 distant Brazilian segments inherited from v4. No marine profile is treated as an observed source for these scores.

Reproduce from the root of this unpacked supplement (only local files; Python, NumPy, SciPy, pandas and Matplotlib):
  export OPENBLAS_NUM_THREADS=1
  python matched_rotation/geography.py
  python matched_rotation/transfer_bearing.py
  python matched_rotation/candidates.py
  python matched_rotation/check.py
  python matched_rotation/figure.py

GEOGRAPHY. Use the bundled IGRF-14 and geodetic zero height with east, north, up components. The angle is atan2(||v1 × v2||, v1 · v2). The 23 nearby surface model tile centers span 12.53219598 to 12.67237574 degrees at 1971.5 and 2008.5. A cell is exterior when its intensity is at least 28 µT at BOTH dates. This operational contour precedes the map analysis. The global regular latitude-longitude grids use cell centers offset by half a grid step from −90° and −180°. Matched counts are sampling-density dependent; approximate areas sum spherical cell areas (radius 6371.0088 km). These are not tests of global rarity, frequencies of crustal anomalies, or counts of independent comparators. Natural Earth 1:110m land polygons omit small islets. An explicit check of São Pedro e São Paulo at 0°55′02″ N, 29°20′42″ W gives 13.27°, outside the narrow angular rule but within the +/−1° padding. Do not interpret the padded land-mask result as a claim covering every islet. Source for the spot check: Ministério do Meio Ambiente, Ilhas oceânicas brasileiras, vol. II, https://www.gov.br/mma/pt-br/composicao/smc/doceano/publicacoes/biodiversidade-costeira-e-marinha-brasileira/2009-ilhas-oceanicas-brasileiras-volume-ii-da-pesquisa-ao-manejo.pdf.

DIRECTION. At each 1971/2008 pair, use the 2008 field unit vector u and orthonormal tangents d, i towards increasing geographic declination and downwards inclination. The historical vector's tangent bearing is beta = atan2(B1971·d, B1971·i). It describes the backward model update. Delta_D and delta_I instead refer to the forward geographic change from 1971 to 2008. The replacement unit vector in the local Brazilian frame is cos(alpha) u + sin(alpha) [cos(beta) i + sin(beta) d]. Here alpha is the actual local angular separation at the modeled heights. The historical/2008 strength ratio used in the kernels is likewise the actual local value at the original model heights. Only beta comes from the offshore location. Because the tangent basis is tied to the local horizontal and vertical, a change of the map's grid convergence does not change the bearing. Rotating the moment and the projecting field uses the same dipole-source model as v4. The representative node is nearest to the median bearing (all 15 occupy less than one degree); its geographic coordinates are in joint_nodes.csv. These are hypothetical field directions on the existing observations, not historical marine measurements. Matching angle and relative strength does not match absolute field, inclination, crust, or acquisition conditions. The 15 surface candidates are spatially correlated. The historic error patterns and the previous cone analysis were known; this is an explanatory retrospective comparison.

SCORING. Use point-weighted within-segment RMSE, separately by geography. Curvature removes the best linear east-coordinate trend from observed-minus-predicted values in each segment. Every centered historical and frozen prediction is checked against the previously archived datum-corrected value at all 17,354 observations. The geographic frozen versus historical selection is preserved. Block tables identify a relevant distinction: the nearby historical direction wins against the representative offshore alternative in five of nine blocks on full shape and all nine on curvature. Omitting any one block leaves the direction of the pooled nearby full-shape advantage unchanged. On distant segments the pooled ordering reverses; the historical direction wins on full shape in 46 of 74 individual blocks. None of these group counts is an independence-based significance test.

RESULTS. sensitivity.csv reports grids, contour definitions, one-year perturbations applied separately to each endpoint (Brazilian angle gate held fixed), approximate areas, latitude and longitude ranges and the highest grid-center rotation on mapped exterior land. matched_nodes.csv contains all 879 qualifying 0.125° points. local_forcing.csv defines the Brazilian rotation, strength ratio and geographic direction. joint_nodes.csv applies the local ratio envelope to the full global sample. joint_refinement.csv.gz uses a separate 0.0125° grid in the region around the 15 coarse joint matches; it is a local refinement, not a second global census. bearing_scores.csv includes leave-one-block pooled scores. blocks.csv, segment_sse.csv.gz and representative_profiles.csv.gz retain multiscale scores and centered predictions. bearing_scores.improved_blocks always denotes blocks where the historical local direction beats the representative marine alternative. Figure_6.png and panel.npz give the manuscript figure and its map input (map display 0.25°; joint markers 0.125°). Verification for the independent spherical harmonic code is in verification.json. The local and distant ranges quoted in the manuscript use all 15 bearings from the specified global grid.

DATA SOURCES. IAGA/NOAA IGRF-14, https://www.ncei.noaa.gov/products/international-geomagnetic-reference-field and https://doi.org/10.1186/s40623-025-02360-0. The supplied independent algorithm and coefficient table come from IAGA-VMOD/ppigrf, https://github.com/IAGA-VMOD/ppigrf, retained with the MIT license and the corresponding Git blobs. Code 5d542cb338bd7732950384befe5cfafd95b20c82; IGRF14.shc 782112655bfab891f8d24d3b2a0225e72f0ad6fc. To compare two implementations of the same interpolated epochs, the independent evaluation uses the exact same fractional position within each five-year coefficient interval. The software comparison does not constitute independent scientific measurements. The land mask is the public-domain Natural Earth 1:110m land layer, https://github.com/nvkelso/natural-earth-vector/blob/master/geojson/ne_110m_land.geojson, Git blob 04811d72fff2701ec67587e30ad8942675b511e3. Brazilian data identifiers and sources are documented in distant_transfer, outside_saa and the root README.

OBSERVATIONAL AUDIT WITH ALTERNATE YEARS. Conditions are documented separately in CANDIDATOS.txt. No alternative survey acquired during different years is used to compute the fixed-epoch global results or the directional transfer scores.
