Smaller prediction ranges, with an unmet target
All 48 generated cases pass the numerical checks. Keeping one calibration transform shared across changing accelerometer samples reduces complete prediction widths by 40–66% against the stronger preserved baseline. The ranges now span about 0.22–1.72 metres at 7.2 seconds.
The study required at least a twofold reduction on every axis in every case. Three of six model/processing groups meet that target; three do not. Both specified-field globe groups miss it, as does the profiled-removal zero-field globe group. This is useful progress, but not a passed long-record qualification. The target and the original failures remain unchanged.
All eight orientation sectors, 27 physical coordinates and 24 shared-error coordinates remain. The same generated inputs, physical bounds, training times and future epoch were used. The run completed in about 50 minutes, spending 144 native calls and 384 bounds. No aircraft measurements were fitted.
The independent saved-evidence audit passes all 48 cases and 288 exact support checks. It reconciles source bindings, coefficient blocks and outward prediction endpoints. Unshared allowances make up 28–37% of individual widths; at least 86% of that allowance follows position propagation. All 24 separate auxiliary coordinate ranges remain unchanged, although joint constraints still retain correlations.
The next component tests necessary relationships between errors that the current calculation bounds separately. It must preserve physically possible calibrations before any saved-case narrowing comparison. Broader recordings, a new aircraft prediction run and protected ESA evaluation do not follow from this result.
These ranges describe a conservative calculation on generated inputs. They are neither measured aircraft errors nor calibrated scientific intervals. The published 53-stretch comparison is unchanged.