Checking gravity during climbs and descents
Why this check is next
The ESA prediction pilot fits earlier positions closely for the specified-field globe candidate but misses later height. The saved-data diagnosis finds calibration at bounds and shows that gravity variation over the recorded height range could matter. Neither observation establishes the cause of the errors.
This is a software check on generated measurements. It opens no aircraft file, changes no existing result and cannot establish an Earth-model preference.
The shared gravity extension now passes 19 focused native checks, and the affected native/compensation regression set passes all 45 checks. Independent truth and related generated-control regressions pass all 95 checks. The longer 48-case study has not run; these short checks do not qualify a revised observed-data fit.
Physical alternatives
Preserve the existing constant-gravity calculation and its interface. Add an explicit alternative in the shared predictor and tangent calculation:
g(h) = g_anchor × ((6,371,000 + h_anchor) / (6,371,000 + h))².
Here h is the predicted height at the integration midpoint. The anchor
height is fixed from the initial observation; its magnitude retains the same
constant adjustment and bounds as the original model. No later receiver height
may enter a forecast. This adds no freely varying gravity or drift parameter.
Apply the same declared height dependence to all three geometry candidates. It is a common force-law sensitivity, not independent evidence for curvature. Latitude dependence, gravity anomalies and other missing dynamics remain outside this first check. Keep the initial force-based reference distinct from measured gravity during aircraft motion.
Implementation and independent checks
Extend shared library components rather than copying a study worker or motion kernel. Keep old entry points on their original constant mode. Historical studies continue to use their archived source, never the revised implementation.
Before a numerical study, run focused deterministic checks covering:
- Exact preservation of the constant-mode output and Jacobian.
- Constant-height limits, gravity signs, dimensions and invalid inputs.
- Midpoint gravity evaluated from predicted state, including its derivatives.
- Every original calibration column, including height, gravity and time offset.
- C value predictions agreeing with the C++ derivative calculation, and full central-difference checks on short generated tracks away from query boundaries.
- Known-answer vertical motion and an independent Cartesian aircraft generator. Truth generation must not call the predictor or its gravity helper.
Finite generated study
Freeze source/environment hashes and passing software checks before launch.
Use 48 identities: three candidates × latitude ±55° × removal 0 or 0.35 ×
constant or height-dependent gravity × two motion patterns. Both patterns last
1,200 seconds and span 2 km in height; one climbs and turns, the other descends
and reverses its heading evolution. Use the existing analytic fast body-motion
pattern. Height is 2500 − 1000*cos(pi*t/1200) metres for the climb and
2500 + 1000*cos(pi*t/1200) for the descent, with analytic first/second derivatives.
Mirror the northward coordinate oscillation for the descent while retaining the
eastward course progression. The anchor height is 1,500 m or 3,500 m respectively.
The independent Cartesian generator derives all force/rate channels without
calling the native predictor or its gravity expression.
Each identity uses known, nonzero sensor/antenna corrections and initial state.
Use the existing positive calibration sign in northern cases and its negative
counterpart in southern cases; neither is an instrument specification.
Generate the same physical trajectory at 300 and 600 Hz. Evaluate 39 fixed queries
at 30 + 30*j + 0.001 seconds, for j = 0..38, retaining every increment.
Run the value predictor at both rates and the complete tangent predictor at
300 Hz. That permits 144 native calls, 2,695,680,000 packet/query visits,
zero optimization starts and no retries. Reserve calls and construction work
before execution. Four workers use one numerical thread each, with a cooperative
one-hour ceiling. No unused allowance from earlier work is available.
At both resolutions the generated acceptance screen uses the geographic adapter's limits: horizontal and height RMS at most 1 m, maximum spatial error at most 5 m, plus the previously declared linear-point refinement rule. Value/tangent predictions must agree within 0.001 m. Report failures rather than relaxing these limits.
What this can qualify
Passing qualifies the tested gravity mechanization with known initialization and calibration. It does not qualify a new observed fit. Longer training-only calibration and forecast controls, with both working receiver-error cases and all nuisance columns, remain a separate gate. Only then may a separately frozen comparison reuse the original development interval under unchanged accuracy limits.
All seven ESA confirmation groups and the original 53-stretch result remain protected. No calibrated confidence, field cause or sidereal/solar conclusion follows from these controls.