What the fitted-prediction scores show
The saved-data calculation is complete. It preserves all 24 generated starts: 20 pass their original numerical checks, and four remain failed. Ten of the twelve training-selected comparisons are available; two have no qualified start.
Among the available primaries, the largest later coordinate error is 5.29–6.25 times the stated conditional noise standard deviation. The score checks the complete later block, with latitude, longitude and height in consistent units. It does not select a favorable time or a better-looking start.
That ratio is not a confidence level or the probability that a model is wrong. It takes a maximum over correlated measurements, and the noise allowance treats estimated calibration as known. The calculation therefore cannot stand in for uncertainty from the complete fitting procedure.
Each case was generated under its named model and fitted using that same model. These are software checks, not another comparison of Earth shapes. They also share one noise seed, so they cannot establish a tail quantile or coverage rate.
The next step uses fresh realizations of the same declared noise laws, with calibration estimated anew for each one. The draw-separation plan keeps covariance assumptions fixed while changing seeds. Independent calibration and validation draws require their own finite freeze before new fits.
No aircraft findings, physical accuracy limits or reserved ESA groups change.