Matched integration-clock trial

Completed: all six fits converged. This comparison asks whether the completed control's model preference survives reoptimizing under packet-header timing, rather than only changing the clock at its old parameters.

The globe family fits better than the flat disc under both processing cases; within the globe family, rotating globe fits better than still globe. Reoptimizing under packet-header timing restores rotating-globe vertical error to about 0.70 m, compared with 9.36 m when changing timing at the old parameters. Horizontal error is about 0.081 m, versus 29.0 m for still globe and 60.5–63.6 m for flat disc. The six fits used 963 evaluations and about 52 minutes of fitting.

This control's model preference survives the clock change. Z-axis gyro gain and offset still reach their allowed limits, so calibration-boundary diagnosis remains necessary. This trial establishes neither the instrument's actual clock convention nor a calibrated statistical significance level.

An IMU (inertial measurement unit) records turning and acceleration. The residual diagnostic found that packet-header intervals raised the control's vertical error from about 0.70 m to 9.36 m at unchanged parameters. Its original nominal-clock fit remains preserved.

The fixed target is the same 7.30-minute control already selected for the solver trial: 570b37d23af5e245dd461051, from ILVIS0_applanix_57300_610_LVIS_38614965.013. Its median receiver ground speed is 878.1 km/h. No intervals are selected again.

The approved allowance is six new starts, at most 200 evaluations each, no retry reserve: rotating globe, still globe and flat disc under the two existing processing cases. Only the clock changes. Calibration limits, receiver covariance, original initialization, gravity, coordinates and the original frozen optimizer stay the same. The six nominal reference fits are reused, not refitted. This avoids mixing the revised solver with the clock test; the revised solver did not improve the observed trial.

Save convergence/stationarity, costs, calibration estimates and signed boundaries, per-epoch residuals, whitened diagnostics, and the nominal-clock counterfactual at each new parameter point. Report shape before rotation and preserve failures. Relative preferences remain development findings, with no new significance threshold.

At the previous trial's measured 3.60 seconds per evaluation, using the entire 1,200-evaluation allowance takes about 72 minutes. At its slowest measured rate, the estimate is about 87 minutes, plus extraction/packaging. This is an estimate, not a wall-clock guarantee; early stopping may reduce it. Use two numerical threads.

The source/input/environment freeze and tested launcher were prepared before approval. --prepare-only does not extract raw observations or make model evaluations. The detached worker has completed. Preserve the frozen results and do not launch it again. To regenerate the evidence package without fitting:

env PYTHONPATH=analysis ~/venv/bin/python analysis/tests/ilvis0_clock_trial.py --package-only

The worker uses an inherited exclusive lock, durable start charges, hashed atomic results and source/environment checks. An interrupted charged start is not retried. Completed work supports --package-only; packaging never fits.

All completed NASA studies and their unused allowances remain frozen. This trial does not expand to the other 52 stretches or introduce new calibration assumptions.