The noise test converged, but missed its accuracy target

All 24 generated fits pass the numerical checks. None passes the physical prediction-accuracy screen. Later horizontal RMS error is 83.6–88.6 m against the unchanged 10 m limit. Height RMS is 5.5–11.2 m, within its 30 m limit.

Each candidate fits data generated from that same candidate. These are checks of prediction robustness with known truth, not votes between Earth models. No recorded flight measurements were used.

The cases include both processing assumptions, both receiver-error cases and two starts. Sensor noise is shared across matched starts and receiver cases; the 24 outputs are not independent flights or independent noise trials.

The calculation adds declared white gyro/accelerometer noise and slowly changing bias before initialization and fitting. It retains the existing receiver-only fitting weights. Calibration uses 800 earlier positions; the continuous forecast is saved before opening the 400 later truth positions. The noise amplitudes are research stress settings, not measured instrument specifications.

This result separates numerical convergence from predictive adequacy. More optimizer iterations are not the immediate remedy: every fit converged, yet the prediction target was missed. The analytic error budget shows how gyro error alone can accumulate into tens of metres over this horizon. It explains a relevant error scale, without identifying the full cause of these particular forecast errors.

The next method must propagate errors through the moving-platform equations, preserve temporal correlation and account for uncertainty from fitted calibration parameters. The 46-check covariance component provides checked building blocks; its constant-level examples do not qualify the aircraft error law. The current accuracy limits and all failures remain preserved. Broader observed fitting stays gated.

The study used 682 native calls and 136,626,619,200 conservative packet/query work units in 5.26 hours. A saved-data audit verified all 24 outcomes and charges, then checked all 477 publication artifact hashes. No refitting or regeneration was needed for publication.

The original 53-stretch findings and completed observed prediction pilots are unchanged. This test does not calibrate their statistical confidence or open the seven reserved ESA confirmation groups.

Frozen result table · Frozen plan · Earlier observed-error diagnostic · Roadmap.