Why the short controls fail their stopping check
The six saved calibration tangents retain all 26 or 27 parameters, but each has only 12 numerically distinguishable directions. Constant-speed tracks cannot separately determine all calibration terms.
| Model | Removal profiled | Existing correction length | Minimum-norm length | Minimum-norm step feasible |
|---|---|---|---|---|
| Globe + field | No | 2.095 | 0.732 | No |
| Globe + field | Yes | 0.073 | 0.073 | Yes |
| Globe, zero field | No | 0.466 | 0.097 | Yes |
| Globe, zero field | Yes | 0.147 | 0.147 | Yes |
| Stationary disc | No | 0.00123 | 0.00119 | Yes |
| Stationary disc | Yes | 1.765 | 0.197 | Yes |
Lengths use normalized calibration coordinates. The audit used the nominal-clock zero start under each model and processing case. It completed 12 linear solves on saved arrays, with no motion integration, new fit or aircraft observation.
A minimum-norm correction reduces some large steps, but not all, and one leaves the allowed bounds. It is therefore not a demonstrated general repair. These linear calculations do not establish how a changed solver behaves on the nonlinear trajectories. Earlier failed labels remain unchanged.
The next method check should also distinguish an already near-zero objective from a nonzero optimum. A squared-residual objective cannot be less than zero; an independently repeated near-zero value can provide a useful numerical stopping criterion without identifying every calibration parameter. Such a rule needs its own fixed tolerance and controls, preserves the ordinary stopping checks elsewhere, and cannot establish scientific significance.
Fixed audit question · Complete control results · Prediction method