What the nonlinear ray test establishes

An independent saved-data audit confirms that all twelve scaled calibrations stay within the original parameter box and satisfy the ±1-metre constraint at every one of the 292 earlier epochs. It recomputes the differences from the sealed positions using exact rational arithmetic; it does not reuse the study’s feasibility calculation. No trajectories are rerun.

The zero-removal north spread exceeds the frozen 20-metre width target. An interval containing all of these training-compatible numerical trajectories cannot be 20 metres wide. Thus conservative bounding is not the only obstacle in this generated coordinate-box fixture. A more accurate interval algorithm cannot erase the demonstrated separation.

This is not a result for the separate running study’s perturbed observations, nor a universal failure of withheld prediction. An observation model with additional justified joint constraints is a different problem and must be tested as such. The smaller sampled spreads in the other coordinates are not upper bounds and do not establish unique prediction.

The accepted points have at least 0.144184 m of margin inside the training limit. The recorded native integrator produced the trajectories; exact auditing of its outputs does not itself prove its floating-point execution error. There is no aircraft fit, detection significance or change to the 53-stretch findings.

The original limits remain unchanged. Further work should distinguish what the training data can constrain from the overestimation introduced by an enclosure, and justify the observation-error model before new aircraft prediction fits.

All numerical outcomes · Frozen scope · Prediction requirements.