Weak-field validation amendment, 5 November 2026

Timing: written after the complete originally specified audit. Its 243 scenarios,
21,870 metric rows and 306,180 slope rows are retained as initial_* outputs.
The initial 10 nT training-RMS normalization produced a maximum anomaly/main-field
vector ratio of 1.007 and a maximum exact-minus-linearized target RMS discrepancy
of 1,445 nT. Some near-surface cases therefore violate the first-order regime.
Their sign recovery cannot be presented as a small-anomaly resolution benchmark.
No geometry, position, depth, magnetization direction or inverse parameter is
selected or removed based on sign agreement or recovery performance.

Repeat every original scenario, multiplying its aggregate moment by
min(1, 0.001 / maximum_vector_fraction) at the combined training/target positions.
This imposes ||b||/||B|| <= 0.001 in every case. Retain the radius-0.25-km sphere
representation and archive its implied uniform magnetization; the prescribed
moments are test sources, not inferred rock properties. Record the actual
training RMS of every scenario. Repeat the same noise bank at 5% and 20% of each
scenario's training anomaly RMS. These are signal-to-noise experiments rather
than a calibrated uncertainty model for the real 2008 survey. Noiseless cases
remain the primary resolution diagnostic. Use exact vector addition, as well as
the first-order forward model, to audit sign classifications at the fixed
0.1 nT/km eligibility threshold. Report any classification differences.

The conditional real-data optimization is unchanged. Add the relative norm
of the coefficient perturbation and the RMS change of fitted 2008 values,
in addition to the already specified coefficient-norm ratio. Archive the
baseline and constrained coefficients and predicted profiles for all models.
This distinguishes a small change in vector norm from a small change in the
vector itself. It does not turn the descriptive objective envelopes into
confidence intervals or establish geological plausibility of signed coefficients.
