Linear repair did not produce valid alternatives
None of the four new calibrations passes the original training constraints. The numerical diagnostic completed in 59.49 seconds, using four native batches. Each linear system had rank 16 and a small linear mismatch, but the nonlinear training errors reached approximately 4–19 m against the unchanged 1 m limit.
The independent audit recomputes exact differences from the saved positions, checks all 27 physical coordinates and the linear equalities, and verifies the charged work. It uses no new native predictions or linear solves. All four nonlinear failures are retained.
The linear corrections were substantial: their largest normalized coordinates ranged from 0.091 to 0.248. Agreement in the tangent calculation did not survive that movement through the nonlinear calibration model. This method is not a qualified way to construct training-compatible alternatives.
The zero sampled spread in the generated results table means that no new point qualified, leaving only the known truth in that table. It is not zero uncertainty. The earlier accepted original-observation probes and their 13.33 m / 5.90 m sampled north spreads remain valid. No target violation has been established for those original observations, and no useful outer bound has been proved.
No aircraft results, training observations, physical bounds or frozen study allowances changed. Full-history uncertainty and the aircraft error model remain open requirements.