Resolution and conditional identifiability audit
Protocol fixed on 5 November 2026 before computing these experiments

Status: post hoc audit motivated by the already observed failure of the 2008
equivalent-source model to predict the negative 2011 target slopes. The initial
18-model check is known: both 2008 channels, depths 4/6/9 km and relative ridge
0.01/0.03/0.10 all give positive slopes on the 14 target lines. This protocol is
not a preregistration of the study and no independent discovery is claimed.

Question 1: does the archived reconstruction operator recover target-gradient
signs from off-grid synthetic sources when observations use the available 2008
grid locations and the actual 2011 bin positions/GPS elevations?

Use the 663 available central training-grid nodes within 24 km, the 2.2 km
training plane, all 14 2011 target lines, and the archived 169-source inverse
mesh. The 2008 grid is interpolated from a survey with nominal 6 km line spacing;
its nodes are not independent original flight measurements. These experiments
audit transfer from the available grid, not the original gridding algorithm.

Generate three source geometries independently of the inverse-source mesh:
one uniformly magnetized sphere; two equal spheres 4.5 km apart east-west;
seven equal spheres along a 146-degree bearing with 2 km center spacing.
All spheres have radius 0.25 km and are below every observation. Their external
vector fields are analytic dipole fields. Center offsets are the full Cartesian
product x=(-2.3,0.7,3.7) km, y=(4.3,7.3,10.3) km from the fixed study center.
Depths below the 2.2 km training plane are 2,4,8 km. This makes 81 geometries.
Use induced moments parallel to the center field, and two fixed remanent
directions rotated 45 or 90 degrees from the 2008 field toward its horizontal
perpendicular. Induced strength/direction changes with epoch; remanence stays
fixed. Treat induced cases as the matched-physics resolution experiment;
remanent cases separately test the assumed magnetization direction.

Normalize the demeaned noiseless 2008 scalar anomaly to 10 nT RMS. Use the
first-order total-field projection, matching the approximation in the archived
inverse operator. Report implied uniform sphere magnetization, maximum vector
anomaly/background ratio, and the exact-minus-linearized total-field discrepancy.
Do not interpret any of these synthetic sources as identified Uruana geology.

Primary reconstruction: depth 6 km, relative ridge 0.03; no synthetic 2011
values select reconstruction parameters. Sensitivity: all nine combinations
of inverse depths 4/6/9 and relative ridges 0.01/0.03/0.10, noiseless only.
For the primary configuration add 0.5 and 2 nT RMS noise, both spatially white
and exponentially correlated with a 6 km correlation length, 20 seeds per
condition. Common seed bank, default_rng(112002026), fixed before evaluation.
Noise is demeaned and normalized to the stated RMS before adding it.
These noise levels are controlled experiments, not calibrated measurement errors.

Report training RMSE, target-profile RMSE after a separate line datum,
gradient RMSE, sign agreement, and all-14-negative recovery for cases where
all 14 true total-field slopes are <= -0.1 nT/km. For general sign agreement
use only truth slopes with absolute magnitude >=0.1 nT/km; also archive every
unthresholded slope. The cutoff excludes numerically marginal signs and is
fixed before viewing outcomes. Summarize by true depth, geometry and moment
assumption; retain every condition, including adverse outcomes.

Question 2: can the 2008 grid distinguish positive from negative 2011 slopes
inside the archived equivalent-source family under explicit fit/roughness budgets?

For each of the same nine inverse configurations, start from the corrected
2008-channel ridge solution a0. Use the exact quadratic ridge objective
J(a)=||A_centered a-d_centered||^2 + lambda^2 ||a||^2. Construct the model
minimizing J subject to all 14 predicted 2011 slopes <= -0.1 nT/km. Report
unconstrained/constrained J, training RMSE, coefficient-norm ratio, and all
slopes; verify constraints and convex optimality numerically. This is a
constrained demonstration, not a prediction or an independent validation.

For each line also compute the analytic least excess objective required to
reach zero slope, using its response under (A' A + lambda^2 I)^(-1).
Evaluate fixed objective envelopes J<=1.1,1.5,2,5 times J(a0), retaining all
budgets. These envelopes are descriptive sensitivity choices, not noise-based
confidence regions. Signed equivalent-source coefficients are mathematical
parameters; a norm bound alone does not establish physically admissible rock
magnetization. If reversal requires substantially worse fit or larger norm,
report that result rather than claiming non-identifiability.

Verification: reproduce the archived primary 14 slopes; verify the synthetic
forward kernel by an independently written vector formula and the equivalent
uniform-sphere moment relation; compare the primary noiseless inverse to a
direct ridge solve; check ridge-objective identities and constrained feasibility.
Archive code, raw metrics, slopes, summaries, protocol and verification report.
