Outside-contour epoch transfer, version 4
South NPR 1977 (AK16) and Howard Pass / east Misheguk Mountain 2008, Alaska.

Run from the extracted archive root with NumPy, SciPy and pandas:
  export OPENBLAS_NUM_THREADS=1
  python outside_saa/transfer.py
  python outside_saa/score.py
  python outside_saa/match.py
  python outside_saa/validate.py

Origin and independence. U.S. Geological Survey sources: Connard et al. (1999), https://doi.org/10.3133/ofr99503 (historical digital lines and channel definitions; AK16 in Table 1); Brown (2012), https://pubs.usgs.gov/of/2009/1256/ (digital magnetic grid, 2008 flight specifications and linked operational report). Appendix D of the latter gives paired NAD27 and WGS84 coordinates. The magnetic grid is WGS84 UTM 4N. The survey report documents pointwise IGRF subtraction, leveling and kriging. The Alaska pair had previously been selected for different magnetic curvature / orientation analyses. Independent acquisitions outside Brazil do not imply independent selection for this project. The 1977 year is known; 1977.542 is the nominal midyear epoch inherited from the archive protocol, not a verified flight date. The 2008.535 epoch represents its June–August campaign. The historical survey is described as 6–10 mile N–S lines and the 2008 survey as 1.6 km N–S lines.

Input files preserve original USGS bytes for ak16.gz and the 2008 GXF renamed HowardPass_Alaska-2008-AeromagMap.gxf in the prior archive. No external downloads are needed. prior_grid_checks.csv holds six grid-only checks from the older P2_PRIMARY_20KM_DOUBLE_SADDLES.csv, independently calculated using GDAL, PROJ and an Alaskan Albers projection. Checks use grid point counts and intercepts of 20 km local quadratic fits and do not select or score any test segment. Validate.py also checks a published WGS84 lat/lon-to-UTM coordinate and a NAD27 coordinate pair. The mainland approximation (EPSG 1176; https://epsg.io/1176) differs by about 26 m from the contractor’s illustrative datum pair in Appendix D. Treat NAD27 for the AK16 geographic coordinates as an assumption based on the compilation’s Clarke 1866 projection. An unshifted-coordinate alternative reruns segmentation and scoring. Native tie records, rejected rows and undefined magnetic values (1e30 sentinel) do not enter the magnetic scores. Neither MAG_MERGE nor MAG_COMP is used.

Geometry. GXF SENSE=1 centers start at 430000,7540000 m. Four-character base-90 data include dummy and run-length codes; scale and offset map values to nT. Every second row and column is used for the source fit. 1977 HEIGHT is constant 243.84 m. Primary geometry uses its difference from the 2008 nominal drape, 304 m. The terrain sensitivity estimates the 2008 surface by adding 304 m to interpolated AK16 ground (= FLTSURF minus HEIGHT); FLTSURF is a USGS compilation flight surface, not necessarily directly measured. Elevation inputs may include nonmagnetic locations / ties. Outside the convex hull the method uses nearest-neighbor heights (extrap column in each tile). That reconstruction is a sensitivity, not a substitute for measured 2008 GPS flight paths. Fixed and updated predictions use identical coordinates and flight surfaces within each case.

Files. pred_* contains one centered observed value and five centered predictions per bin. segment identifiers are case-specific. segments_* reports original line labels, tile IDs, modeled absolute field intensities (nT at surface), angular change at height, training coverage and the fraction of training nodes outside the historical ground hull. The model evaluates IGRF vectors at each tile's median inferred 2008 height and at the nominal earlier height difference; the kernels use the chosen within-tile flight surfaces. A 250 m bin spans grid north, not slant distance. Since segment means are removed, these tests predict profile shape, not absolute field strength or its mean change. Bins, segments and nominal line parts are not independent replicates. The native family groups line IDs sharing an integer prefix. See PROTOCOL.txt for selection criteria.

comparison.csv summarizes both Alaska and v3 Brazilian scores. angle is the median surface IGRF rotation over unique eligible tiles. shift is the RMS change in the predicted centered profiles in nT. bound is 100 times shift divided by the frozen-field RMSE: the triangle inequality bounds the absolute percent gain or loss in the RMSE. cosine = sum[(observed − frozen)(updated − frozen)] divided by the product of the two vector norms. A value of +1 means that the correction points directly toward the observations; it does not imply that its amplitude is optimal. For r = shift / frozen RMSE and c = cosine, (updated RMSE / frozen RMSE)^2 = 1 + r^2 − 2rc. These are descriptive diagnostics on one dataset. clusters.csv uses whole groups (20,000 resamples per grouping); spatial percentiles are sensitive to the origin and small number of blocks. rotation_overlap.csv includes an explicit zero row when no distant data meet the local angular limits at 0.95 km. Blank scores mean that a comparison was impossible.

History and catalog extension. Run:
  python outside_saa/inventory.py
  python outside_saa/s16_audit.py
The following are separate from the 1977 profile-transfer endpoint: (a) earlier curvature-axis tests (S16/P2); (b) the P3 protocol, still unexecuted in the archived records; (c) a new metadata-only rotation screen at the overlap centroids of the 60 listed catalog pairs. The 60 rows share surveys; they are not 60 independent magnetic tests. The old discovery pair had been excluded from the P2 ranking. The S16 replay uses its original 20 km support and native MAG_DGRF filtering, including the original choice not to filter on USE_FLAG. Both raw ZIP checksums and the result package checksum match the previous freeze. This replay verifies archived measurements on previously selected centers; it does not add a separate replication. The decoded AK_165 and AK_550 tables have FID but no native LINE identifier, which prevents unmodified use of the line-segment test applied to AK16. See HISTORICO_ALASCA.txt for the crosswalk to the earlier analyses. candidate_rotations.csv is sorted by angular distance to the local Brazilian interval, with original metadata_rank preserved. It screens only intersection centroids; the actual fit tiles may vary within a footprint.
