What horizontal-only refitting established

All twelve reconstructions and all 24 individual fits passed. The IMU8 stretch passes all six repeat checks and favors globe under both processing cases. The IMU21 stretch passes three of six repeat checks and remains unresolved in this sensitivity study; its primary and repeat costs both favor globe. The original 53-stretch findings remain unchanged.

IMU	Processing case	Model	Horizontal RMS (m)	Unscored height RMS (m)	Objective
8	bias1_unsubtracted	sphere_rotating	0.3741	2.047e+04	5.21957
8	bias1_unsubtracted	sphere_still	7.612	1.257e+04	1532.18
8	bias1_unsubtracted	flat_still	33.78	1.104e+05	16881.2
8	bias1_profiled_removal	sphere_rotating	0.3671	2.498e+04	5.11189
8	bias1_profiled_removal	sphere_still	7.612	1.257e+04	1532.18
8	bias1_profiled_removal	flat_still	17.46	1.104e+05	6777.58
21	bias1_unsubtracted	sphere_rotating	0.01425	648.2	0.0807978
21	bias1_unsubtracted	sphere_still	0.3009	2.477e+04	5.56418
21	bias1_unsubtracted	flat_still	16.82	1.965e+04	3645.61
21	bias1_profiled_removal	sphere_rotating	0.01425	648.2	0.0807978
21	bias1_profiled_removal	sphere_still	0.3009	2.477e+04	5.56418
21	bias1_profiled_removal	flat_still	9.013	1.968e+04	1774.11

The large unscored height errors are an important consequence of removing the vertical constraint. These solutions do not reconstruct plausible full three-dimensional flight paths. The successful IMU8 comparison shows that the score advantage persists without scoring height; it does not establish altitude or datum independence, isolate the gyroscope, or calibrate statistical significance.

Stretch	Case	Model	Repeat passed	Timing difference (ms)	Relative cost difference
b365bdf39150ee6afc303ac1	bias1_unsubtracted	sphere_rotating	True	3.24973e-05	3.26781e-10
b365bdf39150ee6afc303ac1	bias1_unsubtracted	sphere_still	True	0	4.57425e-12
b365bdf39150ee6afc303ac1	bias1_unsubtracted	flat_still	True	0	2.70588e-12
b365bdf39150ee6afc303ac1	bias1_profiled_removal	sphere_rotating	True	-1.21767e-08	2.39833e-10
b365bdf39150ee6afc303ac1	bias1_profiled_removal	sphere_still	True	0	2.87448e-11
b365bdf39150ee6afc303ac1	bias1_profiled_removal	flat_still	True	0	1.14908e-12
7df21c836bec5751f83c0cce	bias1_unsubtracted	sphere_rotating	True	-5.04531e-05	1.6966e-10
7df21c836bec5751f83c0cce	bias1_unsubtracted	sphere_still	False	-18.9589	0.00159649
7df21c836bec5751f83c0cce	bias1_unsubtracted	flat_still	False	0	1.62327e-09
7df21c836bec5751f83c0cce	bias1_profiled_removal	sphere_rotating	True	-2.50803e-06	4.39009e-10
7df21c836bec5751f83c0cce	bias1_profiled_removal	sphere_still	False	-18.9589	0.00159649
7df21c836bec5751f83c0cce	bias1_profiled_removal	flat_still	True	0	5.70641e-12

The saved audit completed 48 affine solves across all 24 endpoints, with no original-recording reads, native predictions or optimizer starts. Its linear predictions cannot establish actual improvement or relabel any fit.

Frozen audit scope · Detailed interpretation · Metrics CSV · Observed study.