Wider arithmetic at the same airborne endpoints

6/6 fixed endpoints completed, using 84 integration evaluations and zero optimizer starts. Findings remain 30 globe preferences, zero flat and 23 unresolved. No fitted result is promoted.

The diagnostic changes internal arithmetic from 53 to 64 mantissa bits, keeping the same physical equations, observations, bounds, nominal packet clock and timestamp branch choices. Inputs and exported coordinates/derivatives remain float64. The generated source and compiler/binary hashes are retained. The finite plan explains all fourteen evaluations per endpoint.

Eleven known-answer software controls passed before archived evaluations: exact straight-disc motion and its analytic velocity derivative, every active derivative under all three models and both processing cases, and finite-call controls. Software controls are not empirical calibration or Earth-model campaigns.

Identical saved parameters

Costs are the same declared whitened sum of squared receiver-coordinate errors. Coordinate shift is the largest axis RMS difference between arithmetic choices. The paired recorder keeps its own original complete stretch; it does not replace the failed recording or share fitted calibration without evidence.

Small cost changes at identical parameters can matter to a numerical stopping limit of 1e-8. They are not changes in the underlying observations or model identity.

Same frozen correction under both arithmetic choices

The following uses the exact same parameter direction, at full size, for both implementations. All five fractions appear in the downloadable CSV. Linear error is the difference between the actual physical prediction change and that predicted by the center's derivative, reported as the largest axis RMS. It includes nonlinear curvature as well as numerical effects; it is not a direct instrument error.

The same full direction passes 2/6 ordinary-arithmetic probes and 5/6 wider-arithmetic probes. These are fixed diagnostics, not counts of recovered fits. A wider kernel's center and correction must be qualified together before a subsequent numerical trial.

Correction calculated from the wider derivative

This table uses one bounded affine correction calculated at the wider center, then evaluated with wider arithmetic. It is separate from the frozen-direction comparison above. Ratios at most one pass the unchanged numerical limits.

5/6 pass these probes. An affine correction is not an observed-data optimization, and a diagnostic pass does not establish a global minimum or replace a primary/repeat comparison. The original failure labels remain.

Interpretation and next qualification

The comparison isolates numerical representation while holding the declared model fixed. Improvement in local prediction agreement supports further precision testing; lack of improvement leaves nonlinear curvature or weak calibration combinations as unresolved explanations. A general change needs known-answer solver checks and a separately frozen paired-recording trial, retaining all models, both processing assumptions, repeated starts and the original tolerances. No extra bias freedom, automatic retries or corpus-wide rerun follows from this diagnostic.

Reproduce the report

~/venv/bin/python docs/ilvis0-shared-leg-precision-20261014/build_report.py

Individual precision probes retain all fractions and both directions. Rendering performs no integration or optimization. Completed native work is frozen.