Native training constraints: generated feasibility test
Test whether necessary constraints from generated training positions can narrow the full declared calibration box without losing the known true parameters. Use every one of the 27 native coordinates under all three candidates and both processing cases. Reuse qualified bound primitives; change no frozen kernel, physical calibration limit or archived result.
Each of six identities uses 400 generated 200-Hz packets. Bounds and Jacobians are evaluated at six training times (0.2, 0.35, 0.6, 0.9, 1.2 and 1.5 seconds), with a 1.8-second future query. The original broad calibration limits, including full yaw uncertainty, remain. No artificial narrow sensitivity box replaces them.
Construct two training cases with declared per-coordinate measurement allowances of 1 mm and 5 m. These are supplied deterministic control assumptions, not estimated receiver precision. Training positions are generated at a fixed nonzero parameter vector. Future truth is evaluated only after each contraction and its new prediction center have been fixed; it is absent from the contractor.
Propagate a global mathematical remainder over each complete input box, including timing and processing parameters. Build necessary affine inequalities, contract conservatively, and recompute a future bound at the contracted box's midpoint. Retain all directions and both noise cases. Report widths before/after even if they do not improve. Require the known truth to survive and its future prediction to remain inside the resulting mathematical envelope.
The numerical comparison allows 1 micrometre in native values and 1e-7 metres per normalized derivative column, matching earlier short controls. A declared 10-micrometre arithmetic allowance in the constructed inequalities covers those empirical comparison scales; it is not a global floating-point certificate. Passing cannot establish observational coverage or authorize aircraft fits.
Finite allowance: six identities, 36 native calls, 54 bound evaluations with 21,600 reserved packet steps, twelve contractions at eight passes each, four total numerical threads and 180 cooperative wall seconds. Reserve calls/steps before evaluation. No retries, signals, completed-study restarts or allowance reuse.
No archive, fused navigation or protected ESA recording is read. Inputs, sources, environment and limits freeze before execution. An overly wide result remains overly wide: do not select favorable cases or alter limits to obtain contraction.