What changed when the clock was refitted?
All six clock fits converged and preserved globe over flat, followed by rotating over still globe, under both processing cases. This report compares saved parameters from the same 7.30-minute stretch, with median ground speed 878.1 km/h. Its IMU record type is 21. An IMU (inertial measurement unit) measures turning and acceleration. These are repeats of one stretch, not independent flights.
The clock discrepancy is absorbed by calibration readjustment
Changing the clock at the old parameters raised rotating-globe vertical error from about 0.70 m to 9.36 m. Refitting restored it to about 0.70 m. The objective changed by only about 0.12%. Both clocks therefore produce similar optimized track errors here, despite different calibration estimates.
| Processing case | Model | Nominal cost | Header cost | Cost change (%) | Parameters at limits, nominal / header |
|---|---|---|---|---|---|
| bias1_unsubtracted | sphere_rotating | 2.31103 | 2.30817 | -0.12377 | 3 / 2 |
| bias1_unsubtracted | sphere_still | 10747 | 10748.6 | 0.01438 | 14 / 14 |
| bias1_unsubtracted | flat_still | 49304.2 | 49303 | -0.00247 | 15 / 15 |
| bias1_profiled_removal | sphere_rotating | 2.31103 | 2.30814 | -0.12488 | 4 / 3 |
| bias1_profiled_removal | sphere_still | 10747 | 10748.6 | 0.01438 | 15 / 15 |
| bias1_profiled_removal | flat_still | 45776.4 | 45775.7 | -0.00151 | 15 / 15 |
Cost is the saved sum of squared covariance-scaled receiver errors. It is not a detection significance or a hardware calibration measurement.
Which parameters moved?
The table below shows the preferred rotating-globe fits. Gravity correction leaves its lower limit under header timing; accelerometer offset, X/Y gyro gains and timing offset also move. This shows that these estimates can readjust together. A two-clock comparison cannot attribute the change to one parameter or identify the physical cause. No new Jacobian or covariance was calculated.
| Processing case | Parameter | Unit | Nominal value | Header value | Change | Limits, nominal / header |
|---|---|---|---|---|---|---|
| bias1_unsubtracted | gyro_bias_z | deg/hour | 1 | 1 | 0 | upper / upper |
| bias1_unsubtracted | accel_bias_z | m/s2 | -0.00663631 | 0.0310887 | 0.037725 | interior / interior |
| bias1_unsubtracted | gyro_gain_x | % | -0.02325 | 0.410898 | 0.434148 | interior / interior |
| bias1_unsubtracted | gyro_gain_y | % | -0.143765 | 0.264927 | 0.408692 | interior / interior |
| bias1_unsubtracted | gyro_gain_z | % | -2 | -2 | 0 | lower / lower |
| bias1_unsubtracted | gravity_delta | m/s2 | -0.05 | -0.00839876 | 0.0416012 | lower / interior |
| bias1_unsubtracted | time_offset | s | 0.00341055 | 0.00898592 | 0.00557537 | interior / interior |
| bias1_profiled_removal | gyro_bias_z | deg/hour | 1 | 1 | 0 | upper / upper |
| bias1_profiled_removal | accel_bias_z | m/s2 | -0.00663631 | 0.0325003 | 0.0391367 | interior / interior |
| bias1_profiled_removal | gyro_gain_x | % | -0.02325 | 0.397304 | 0.420554 | interior / interior |
| bias1_profiled_removal | gyro_gain_y | % | -0.143765 | 0.248257 | 0.392023 | interior / interior |
| bias1_profiled_removal | gyro_gain_z | % | -2 | -2 | 0 | lower / lower |
| bias1_profiled_removal | gravity_delta | m/s2 | -0.05 | -0.00995147 | 0.0400485 | lower / interior |
| bias1_profiled_removal | time_offset | s | 0.00341055 | 0.00905022 | 0.00563968 | interior / interior |
| bias1_profiled_removal | earth_removal | fraction | 5.55112e-17 | 3.77476e-15 | 3.71925e-15 | lower / lower |
Every matched parameter and fit comparison preserve the complete changes for all three models.
The Z-axis question remains
Rotating-globe Z gain stays at −2% and its constant offset stays at +1°/hour in both clocks and both processing cases. These are permitted correction limits, not measured instrument drift. The model corrects a logged rate as (logged rate − offset) / (1 + gain). At those limits the multiplier is 1.020408 and the additive correction is −1.020408°/hour. Z correction coefficients make the sign convention explicit.
On a stretch with limited rotation-rate variation, gain and offset can compensate for one another. This algebraic possibility does not prove that it explains these fits; the saved singular values alone do not identify which parameters are coupled. Stable ranking across clocks settles this control’s timing sensitivity, but does not settle why Z calibration remains at a limit.
Calibration comparison
The calibration plan tests each Z correction separately by fixing it to zero while leaving the other free. Zero is a diagnostic reference, not an asserted factory calibration. No limits are widened and no time-varying drift is introduced. All three models receive the same changes.
This audit makes zero native predictions and zero optimizer starts. It verifies published evidence and completion hashes, and checks matched identities, physical parameter values, limits and convergence labels. Original results remain frozen. The separate calibration study has completed its 12-fit allowance, with 11 fits converged; this audit adds no fits.
Reproduce
env PYTHONPATH=analysis ~/venv/bin/python docs/ilvis0-clock-calibration-audit-20261011/build_report.py