What each measurement contributes

All 53 selected stretches now have complete numerical comparisons favoring the globe. This next investigation asks which measured combinations create that separation, and how it changes when specific receiver constraints are removed.

An IMU (inertial measurement unit) records turning with gyroscopes and acceleration with accelerometers. In this analysis, recorded gyro increments drive the reconstructed attitude; that attitude turns the measured acceleration into a trajectory. The predicted positions are then compared with the receiver's recorded positions. There is no separate term that simply rewards an expected Earth-rate gyro value. Measurement dependencies explains the full path, including initialization and selection.

First software result

The new receiver-coordinate adapter passes 19 known-answer checks. It can retain full three-dimensional positions, horizontal positions or heights while using each selection's correct marginal covariance. It preserves all nuisance parameter columns, including columns that become unidentifiable after coordinates are removed.

One control explicitly demonstrates why deleting already-whitened residual rows can produce the wrong likelihood. Other controls verify derivatives, coordinate permutations, lost rank and rejection of invalid covariance or saved arrays. These checks use small constructed matrices, with no original-recording reads, native predictions or observed-data fits. They do not add an empirical finding.

First receiver-coordinate result

The saved-parameter comparison completed 324 projections for eighteen stretches, covering all three models and both processing assumptions. At the existing fitted parameters:

These are 36 processing-case comparisons on eighteen stretches, not independent flights. The globe preference persists when height is removed from the score. Height alone gives mixed rankings. The parameters still come from full-position fits, so these scores do not establish the best fit attainable after removing coordinates. Both gyro and acceleration observations remain in use.

The full scores reproduce the saved objectives exactly. Removing coordinates also makes some nuisance directions unidentifiable: fifty of 108 horizontal projections and all 108 height-only projections lose numerical rank. Every parameter column was retained; this rank diagnostic changes no acceptance rule. Interpretation and limits.

Refitting qualification

The refitting qualification has completed its fail-fast batch: 24 controls passed, one failed and eighteen were not run. The six horizontal fits reached so far passed; their second-start checks remain untested. A height-only control exhausted its allowance without meeting stationarity, despite a very small retained-position error. Omitted coordinates moved substantially. The candidate remains unqualified and no observed refit follows from these partial successes. What passed and why advancement is blocked.

The subsequent trust-region matrix ran all 24 generated fits. Sixteen fits and eight paired-start comparisons passed, including every horizontal fit. Its eight failures were height-only stability probes that made large nonlinear moves through weak directions. The subsequent local-probe matrix used the same numerical tolerances within an explicit 0.001 normalized-coordinate neighborhood. It passed all 24 individual fits but only eleven of twelve paired-start comparisons. The larger-budget continuation completed both height-only fits in 1,035 native calls, within its 1,204-call allowance, but did not resolve that pair's objective disagreement. The complete saved review therefore remains unqualified. Both coordinators finished without launching the broader full-calibration matrix; they are not waiting or running. Call-allowance publication correction.

All twelve predeclared horizontal fits and all six horizontal repeat comparisons pass. A separate horizontal-only capability review verifies that entire subset while retaining height-only as unqualified. The current full-calibration qualification tests horizontal scoring with all 26/27 nuisance columns: twelve generated fits, six repeat comparisons and at most 1,944 native calls. It reads no original recordings and starts no observed-data fits. Success would qualify preparation of a small matched horizontal-only recorded-data study, not height-only refitting or calibrated scientific significance.

The saved-primary inventory is complete. Of 318 final primary fits, 164 have ready residual/Jacobian and receiver-time artifacts. Eighteen stretches have all six fits ready for a matched comparison. Other fits lack a declared array, saved time context or a resolvable reference. These reuse gaps do not change the completed 53/0/0 model result.

1. Preserve the completed inventory and trace the twenty time-context gaps and one unresolved artifact reference through saved provenance.
2. Preserve the completed fixed-parameter comparison and its frozen sources; do not repeat projections to regenerate documentation.
3. Complete the separately frozen horizontal-only full-calibration controls, retaining every column and checking unscored-coordinate effects. Height-only qualification remains a separate open problem; preserve the failed repeat check.
4. Only after those controls pass, freeze a small matched horizontal-only observed refit study with finite limits, common initialization rules and independent repeats.
5. Develop gyro/acceleration contribution controls with explicit replacement constraints, then held-out prediction and calibrated uncertainty.

Removing receiver heights from the objective still leaves heights in the current initialization, covariance and trajectory context. That experiment measures sensitivity to a constraint; it does not establish independence from altitude or the coordinate datum. Likewise, coordinate projections retain both IMU channels and do not isolate the gyroscope by themselves. No fused navigation attitude or angular rate may substitute for a removed observation stream.

The existing findings and all frozen studies remain unchanged. No observed ablation run has started. Detailed scope · Current findings · Next analyses.