RESOLUTION AND CONDITIONAL SIGN-STABILITY AUDIT
Post hoc analysis, 5 November 2026. Manuscript Sections 7.1–7.2, Figure 10.

PURPOSE AND DESIGN TIMING
The pre-existing 2008 equivalent-source model improved nearby historical
profiles but predicted positive target slopes on all 14 lines observed as
negative in 2011. This audit tests the limitations of that reconstruction.
It is not a prospective prediction, geological inversion or independent
validation of the observed anomaly. PROTOCOL.txt was frozen before these
new tests, after the original discrepancy was known. Its SHA-256 is supplied.
AMENDMENT_1.txt documents the subsequent weak-field amplitude correction.

REPRODUCTION (from the extracted supplement root)
  export OPENBLAS_NUM_THREADS=1
  python resolution_audit/run.py
  python resolution_audit/summarize.py
  python resolution_audit/verify.py
  python resolution_audit/figure.py

No network or original large survey downloads are needed. Required packages:
numpy, scipy, pandas and matplotlib, with versions listed in the root README.
Input subsets and the IGRF evaluator/coefficients are in amas_test/input and
amas_test/geo.py. All paths are resolved relative to the scripts. The primary
inverse is depth 6 km below the 2008 plane and relative damping 0.03.
run.py recreates numerical outputs; summarize.py creates the descriptive
comparator table; figure.py creates Figure_10.pdf and Figure_10.png.
verify.py checks the archived outputs and independently reconstructs the
primary real-data matrices and objective; it does not rerun every simulation.

To reproduce the superseded initial amplitude normalization separately:
  python resolution_audit/run.py --initial-normalization
This writes initial_reproduced/, leaving the amended outputs unchanged.
The original initial_synthetic_* files and initial_verification.json are
retained as the decision record. CSV content, not compressed-byte hashes,
should be compared after a rerun because gzip metadata can differ.

KNOWN-SOURCE EXPERIMENT
Training: 663 central nodes of the available interpolated 2008 grid, nominal
height 2.2 km, sampled on the native 1.5 km grid inside a 24 km radius. These
are not 663 independent raw flight observations; original line spacing is
6 km. Targets: 110 bins on 14 actual 2011 lines at their GPS heights. The
experiment does not reconstruct the original survey interpolation step.

Truth: one sphere, two equal spheres separated by 4.5 km, or a chain of
seven equal spheres 2 km apart along bearing 146 degrees. The centers use
all nine x/y combinations (-2.3, 0.7, 3.7)/(4.3, 7.3, 10.3) km relative
to the survey center and depths 2, 4 and 8 km below the 2008 plane. Radius
is 0.25 km; all observation points are outside the spheres. Source locations
are distinct from inverse nodes. Exterior fields are analytic dipoles.
Each geometry is tested with an induced moment following the epoch field,
and fixed directions rotated 45 or 90 degrees from the 2008 direction toward
a horizontal perpendicular. There are 81 geometries and 243 total scenarios.
Induced moments are prescribed proportional to the main field: susceptibility
and demagnetization are not inferred. Equivalent magnetization values are
archived for transparency and do not establish geological plausibility.

Truth scalar anomalies use first-order projection with the same center-field
convention as the inverse. Exact vector norms using the position-dependent
background are also evaluated. The initial 10 nT training-RMS normalization
produced some fields outside the weak-field approximation. The amendment
therefore rescales every scenario as necessary to cap the anomalous vector
at 0.1% of the main field at every training and target point. No geometry
was discarded based on recovery. The amended target exact-versus-linear RMS
is below 0.027 nT in every case, with no eligible truth-sign changes. The
resulting training-signal RMS varies from about 0.010 to 4.16 nT. Noise levels
are relative experimental controls, not estimates of actual survey noise.

Inversion: 169 signed equivalent sources on a 4 km mesh; depths 4, 6 and
9 km; relative ridge damping 0.01, 0.03 and 0.10. All nine settings are
applied without noise. The primary setting also uses 20 realizations each
of white and exponentially correlated noise (6 km correlation length), at
5% and 20% of scenario training-signal RMS. A common seeded noise bank
(112002026) is reused across cases. This yields 21,870 metric rows and
306,180 archived line slopes. These are related sensitivity cases, not
independent statistical replicates. The primary noiseless result appears
in both the sensitivity and noise tables and must not be counted twice.

Sign agreement includes truth slopes with magnitude at least 0.1 nT/km.
For noiseless induced cases at depths 2/4/8 km, correct signs are
254/367, 285/367 and 346/377 (69.2%, 77.7%, 91.8%). A background-only
comparator gives 80.4%, 66.2%, 82.8%; it outperforms the inverse for the
shallowest cases. The comparator is a descriptive addition after inspection,
not a predeclared confirmatory endpoint. Of six induced scenarios with all
14 true slopes <= -0.1 nT/km, only one predicts all 14 slopes negative.
All remanence, inverse-setting and noise outcomes remain in the archive.

CONDITIONAL REAL-DATA ANALYSIS
For centered training matrix A and observations d, the objective is
J(c) = ||A c - d||^2 + lambda^2 ||c||^2. At the unconstrained optimum c0,
J(c0 + delta) - J(c0) = delta^T H delta,
where H = A^T A + lambda^2 I.

For each of nine corrected-channel inverse settings, run.py finds the
least-objective solution satisfying all 14 total-field target slopes
<= -0.1 nT/km. Cholesky whitening turns the excess objective into a squared
Euclidean norm, minimized under linear inequalities. Feasibility, active
constraint KKT conditions and the objective identity are verified. Analytic
individual-slope envelopes are reported for objective factors 1.1, 1.5,
2 and 5. Individual extrema need not occur at one common coefficient vector;
the simultaneous constrained solutions are supplied separately.

Primary result: baseline/constrained training RMSE 1.006907/1.116244 nT;
objective +8.3765%; fitted-grid change 0.471426 nT RMS. The new coefficient
vector norm is 1.017362 times the original; the perturbation norm is
0.195710 times the original. These two measures are not interchangeable.
All nine settings permit simultaneous sign reversal at objective increases
0.5–16.8%. A separate table contains the 18 unconstrained settings (both
2008 channels), all predicting positive target slopes on all 14 lines.

The sign constraints use known observed signs. This is a sensitivity
counterexample, not independent prediction or an exact same-fit alternative.
Objective factors are descriptive fit–regularization budgets, not confidence
regions or acceptable-error thresholds calibrated from survey noise. Signed
coefficients and their norms do not establish geological admissibility.
The conclusion concerns conditional prediction stability within this family;
it does not establish the cause of the real 2008/2011 discrepancy.

OUTPUT MAP
conditional_models.csv: objectives, fit errors, coefficient measures and KKT.
conditional_slopes.csv: observed, baseline and constrained line gradients.
conditional_coefficients.csv: source positions and both coefficient vectors.
conditional_predictions.csv.gz: baseline/constrained sample predictions;
  training rows are datum-restored anomalous values (main field excluded),
  target rows are total-field values including the modeled regional field.
objective_envelopes.csv: individual slope extrema at each objective factor.
survey_channel_sensitivity.csv: all 18 unconstrained settings, 252 line rows.
synthetic_scenarios.csv: truth geometry, amplitude and approximation checks.
synthetic_metrics.csv.gz: all model/noise cases and aggregate metrics.
synthetic_slopes.csv.gz: every truth/predicted target-line gradient.
synthetic_summary.csv: summaries by depth, moment, inverse and noise controls.
synthetic_comparators.csv: primary noiseless model versus background alone.
verification.json: execution counts and independent kernel/ridge checks.
initial_*: superseded strong-field normalization, retained for transparency.
