What the new refits show, and what remains unresolved
The two selected tracks were chosen by duration and IMU configuration from the complete saved-artifact set. All twelve full-position reconstruction checks passed before fitting. All 24 horizontal-only fits then passed their individual numerical gates. The study used 1,112 native calls and took 3.64 hours.
The IMU8 comparison is complete
The 9.75-minute recording is ILVIS0_applanix_55124_Ames_510-20oct09_20Oct09.013.
All six independently initialized repeat comparisons pass. Both processing cases
favor the globe family after refitting without height contributing to the score.
The specified-field globe has horizontal RMS error about 0.37 m, compared with 34 m for the unsubtracted flat model and 17 m when the flat model may profile upstream field removal. The zero-field globe gives about 7.6 m. Thus both shape and the specified-field comparison favor the same models as the preserved full-position analysis. No significance level is assigned.
The IMU21 comparison remains unresolved
The 4.07-minute recording is ILVIS0_applanix_55295_20100409_610_LVIS_09Apr10.013.
All twelve individual fits pass, but only three of its six repeat comparisons do.
Its shape/field interpretation is withheld under the frozen repeatability policy.
Both starts nevertheless rank the specified-field globe first: horizontal RMS is about 1.4 cm, compared with 30 cm for zero-field globe, 17 m for unsubtracted flat and 9 m for profiled-removal flat. The failed repeats are not evidence of equal model support; they prevent a completed matched result.
The two zero-field globe cases settle at timing offsets near −84.23 ms and −103.19 ms. Their costs differ by about 0.16%, and their predicted positions by several millimetres. Both stops report no further positive predicted decrease. Neither is especially close to a packet boundary: the nearest distances are about 0.71 and 1.75 ms. Ordinary and range-scaled affine solves give essentially the same tiny predicted improvements. This differs from the earlier incorrect affine-solve failure; rescaling alone does not explain the disagreement.
The unsubtracted flat pair agrees in cost but has north-position RMS difference about 1.21 mm, above the 1 mm limit. Timing is identical at its bound; gravity correction and weak vertical coordinates differ. Its repeat used all 200 evaluations. The other profiled-removal flat pair passes.
Height removal permits implausible vertical paths
Unscored height RMS ranges from roughly 0.65 km to 114 km across the returned fits. These are residuals against receiver height, not measured changes in the aircraft's altitude. Removing the vertical score leaves nuisance directions insufficiently constrained to reconstruct a credible full flight path.
This sensitivity check shows that the completed IMU8 score advantage survives removing height from the objective, even with that additional freedom. It does not establish a physically adequate three-dimensional alternative fit. Heights still enter initialization, covariance and propagation, both IMU channels remain, and conventional receiver-coordinate construction has not been removed.
The original 53 globe / zero flat / zero unresolved result uses the original full-position objective. Its labels are preserved; the unresolved IMU21 sensitivity check does not overwrite them. No dataset expansion or global-optimum claim follows. The completed fixed-point diagnostic isolates coupled timing/calibration and an unobserved vertical direction without changing any acceptance limit.