Longer generated predictions remain bounded

All 48 sectors pass, completing all six model/processing comparisons at 7.2 seconds. Complete prediction widths range from 0.620 to 2.973 m across axes and cases. Every original orientation sector and all 27 physical calibration directions remain; 15 shared-error coordinates retain dependence between training and future epochs.

The earlier complete comparison covered 1.8 seconds. Quadrupling the time increases the widths by 8.27–12.47 times. The common-acceleration construction therefore remains correct at this longer horizon, but uncertainty still grows faster than elapsed time. These generated ranges are not measured aircraft errors, calibrated intervals or a new globe/flat result.

The run used 144 native evaluations and 384 bounds over 46.3 minutes, within its frozen 90-minute allowance. Four workers each used one numerical thread. The original short study, measurements, physical bounds and failures remain unchanged. No protected ESA confirmation data were used.

The next step is a saved-evidence audit of training constraints and the affine, position-propagation and lever contributions. A larger linear optimization or longer recording run should follow only if that audit justifies it. Recording-scale usefulness, whole-program numerical error and actual observation-error adequacy remain required before new aircraft prediction comparisons.

Complete results · Frozen scope · Short comparison · Roadmap.