Independent geographic fitter controls

24/24 controls passed the predefined numerical checks. These are synthetic software controls, not additional NASA findings. They exercise the native geographic predictor, its automatic derivatives, receiver timing and both nonlinear solvers. There were zero observed-data fit starts. The failed observed stopping-rule trial and the 53-stretch first-pass findings remain unchanged.

An IMU (inertial measurement unit) measures turning and acceleration. Here its three-axis increments come from independently specified continuous motion, not the production predictor. The receiver's positions are separately generated at known times. The fitter must reconstruct those observations while estimating calibration, initialization and latency. See the predefined plan.

What was tested

The six physical fixtures cover rotating globe, still globe and stationary disc, each with a known 12-second equatorial eastward track at 220 m/s and 1,000 m altitude. Body pitch and yaw vary smoothly. Eight-point Gaussian quadrature integrates the analytic body angular rate and specific force. Receiver positions follow the appropriate analytic coordinate metric; neither navigation kernel generates the input data. Receiver epochs fall between packets, with independently specified latency.

Each model has constant-period/interior and variable-period/near-boundary cases. Variable periods have a known 50 ppm modulation; this is a different physical clock construction, not applying two clocks to the same observations. A separate test deliberately applies an incorrect clock scale and confirms a position discrepancy. This does not identify the period represented by NASA's logged increments.

Six nuisance parameters are free: initial pitch, initial eastward velocity, Z gyro offset and gain, gravity correction and receiver latency. Other nuisance parameters are fixed to their known synthetic values, including no upstream Earth-rate removal. The same box and observations are used for both solvers. Each fixture is fitted from zero and an opposed starting point. The common synthetic position scale is 1 cm, with no random noise. Prediction tolerance is 2 mm; objective tolerance above the known-parameter reference is 0.01. These are software checks, not significance levels.

Results

The largest returned-track error is 5.52961e-06 m. Finite-interval navigation uses a midpoint approximation, so exact continuous input integration has a small nonzero reference error. All fitted objectives and reference errors are retained in fits.csv, with complete fit diagnostics in the saved report artifacts.

At the known parameters, the largest reference position error is 4.21112e-05 m. The two native implementations agree to 0 m. Automatic derivatives agree with separate central value differences within 8.57e-09 m per normalized parameter unit. All six fixtures have off-packet receiver epochs. Conditioning and singular values are saved; a numerical rank cutoff is not scientific identifiability.

The largest parameter difference from the injected reference is 0.646787 in units of its bound. Calibration recovery must be assessed separately from prediction recovery: discretization error and weakly constrained parameters can shift the best fit. No claim of uniquely recovered instrument calibration or global optimality is made.

The report charged exactly 24 starts, with no retries. There were 1275 native sweeps, including independent fixture checks and final prediction checks, within the declared finite allowance. Ordinary unit-test invocations are separate from this report's ledger. All results, including failures if any, are retained.

What remains

These short six-parameter equatorial controls do not qualify every nuisance direction in the 27-parameter archival fit, long-record conditioning, alternative processing or independent-flight prediction. Nor do they explain the NASA Z gain/offset boundaries. Next test the difficult calibration directions and full initialization at a finite software budget before preparing matched observed gain–bias profiles. Those observed profiles require a separately approved allowance; no corpus rerun is started here. See the focused analysis plan.

The public first-pass result remains 12 globe preferences, no completed flat preference and 41 unresolved stretches. A successful software control does not turn those exploratory comparisons into calibrated scientific detections.

Rebuild documentation without starting fits:

env PYTHONPATH=analysis ~/venv/bin/python docs/ilvis0-geographic-controls-20261011/build_report.py --package-only