Sensitivity to survey processing and slope estimation
For each of the 23 sectors tested in section 4.2, this audit uses five archived 2011 total-field products and two estimators in both survey years. Measurements come from the same acquisitions, so these are processing sensitivities. The conditional geographic results count modeled background fields combined with measured Brazilian scalar residuals; they do not estimate the prevalence of observed anomalies or the instrument error.

Reproduction from the root of Supplementary Material S1 (Python, pandas, NumPy and SciPy; ppigrf is bundled)
  export OPENBLAS_NUM_THREADS=1 OMP_NUM_THREADS=1
  python processing_audit/audit.py --workers 4
  python processing_audit/stress.py --workers 4
  python processing_audit/verify_stages.py
Optional native re-extraction:
  python processing_audit/extract.py /path/to/3087.zip

The regional_specificity and background_transplant directories implement the previously defined sector eligibility, transfer geometry, IGRF synthesis, and native bin/row weights. stage_bins.csv.gz stores the five products for the exact 8,009 eligible bin positions. The coordinates, GPS heights, line IDs, windows and raw data used for transport are in nested_gradients/control_bins.csv.gz. The optional extractor reads the contractor's original Mag-3087.XYZ and checks bin medians against the supplied MAGBRU/MAGCOM coordinates and values. source_check.json identifies the archive member, byte length, CRC and number of records inside the selected windows. Processing labels are based on Microsurvey Aerogeofísica e Consultoria Científica Ltda. (2011), Levantamento aerogeofísico de Minas Gerais, Programa 2010/2011, Área 19 Formoso–Buritis–Lagoa Grande, Relatório final, vol. I, sections 3.1.2.1–3.1.2.5 and Table 22. Source: https://geosgb.sgb.gov.br/downloads/index.php?dir=3087-Area19-Formoso_Buritis_Lagoa . The XYZ export includes MAGBRU (uncompensated), MAGCOM (compensated), MAGCOR (diurnally corrected), MAGNIV (leveled) and MAGMIC (microleveled). Use the archived total-field values. MAGIGRF contains the result after subtracting the provider's reference field from MAGMIC. Before binning, the absolute difference between MAGMIC and MAGIGRF+IGRF is at most 0.0011 nT across the selected raw rows with all three fields present.

Calculation. The 2011 input is the median in each original 250 m bin. The 2008 input averages native nodes with the same northing within each band, channel and width. Each bin or occupied row has equal weight in OLS. Theil–Sen takes the median slope over every distinct pair of occupied northings. The median is computed AFTER subtracting the native main field and adding the destination main field point by point (medians of slopes are nonlinear). For 2008 the geographic change is applied to the nodes before forming row means. The 2011 estimator uses original northings for both native and relocated fields. Both dates must pass at the same center. A geographic destination is in the all-ten intersection only when it passes both dates for all five archived 2011 stages under both estimators. Both native 2008 channels and all widths are tested in every strict calculation.

results/native.csv and slopes.csv: complete native 2011 pairs, worst signed margins in each survey, joint margins, and per-window slopes. 23 sector rows per configuration in native.csv. Negative (target) and positive (north) are relative to grid north. 2008's score does not change between 2011 stages for a fixed estimator.
results/world.csv: strict joint matches out of 64,800 per stage, estimator and sector. common.csv: intersection of all ten outcomes at each destination. relaxed.csv: count after dropping the most adverse whole 2011 line AND 2008 corridor width for each alternative. Counts use equal-angle cells and are not area fractions.
results/locations_world.csv.gz and locations_parallel.csv.gz: destination coordinates, shared mask, selected single-configuration masks and both surface intensities. geography.csv reports counts and weighted within-class fractions: spherical latitude-strip area for the world; uniform longitude spacing for the single parallel. Class transitions have their own denominator.
results/one_bin.csv: recomputed native signed slope in the one affected 2011 line/window after omitting each single retained bin. The 348 possible omissions per configuration meet the original minimum count and span. An entry <=0 loses that sign; every other window retains its native data. Overlapping windows, estimators and processing stages do not supply independent replicates or error bars.
results/verification.json: numerical spot checks of independent ppigrf synthesis and scipy Theil–Sen at four destinations for four sector templates, plus baseline count checks. PROTOCOL.txt distinguishes the primary comparison from follow-ups after examining native margins and the first processing scan.
