Distinguish loose bounds from real ambiguity
The first four completed cases have smaller bounds, but still fail the original width targets. Preserve the complete six-case comparison before this follow-up.
The next diagnostic will search for two calibrations that fit the original five perturbed observations but predict substantially different later positions. It uses generated data, not aircraft measurements or reserved ESA recordings.
Fixed comparison
- Keep all three physical candidates and both upstream-processing assumptions.
- Keep every original input increment, all 27 calibration coordinates, their original bounds, the ±1-second clock range and the same coordinate frame.
- Retain the five training times, their ±1 m coordinate allowances and the 438-second prediction endpoint. Do not substitute noiseless training positions.
- Start each search at the fixture's saved, nonzero calibration. Search once toward each north-coordinate extreme: twelve fixed identities altogether.
- Use the actual nonlinear native positions and Jacobians in a constrained search. A linear proposal alone cannot qualify a candidate. Leave 0.01 m of training margin during search; this restricts the search without changing the declared physical domain or observation problem.
- Save every new point before its native call, including rejected points and failed outputs. Cache repeated point requests; do not charge a cached lookup as a new native evaluation. All new evaluations are charged before execution.
- Preserve every attempted point. An optimizer's success label does not establish training feasibility, uniqueness, an optimum or useful prediction bounds.
Qualification and limits
First test the search adapter against independent nonlinear known-answer controls, including a boundary optimum, a misleading linear proposal and exhaustion of its call allowance. Check all parameter columns and native batch/counter behavior on a small generated history. Use the common library and study runner.
The proposed full-history allowance is 12 identities, each with at most 32 search batches and one separately charged verification batch. Each batch queries all five training epochs and the prediction epoch. Maximum totals are 396 native batches, 2,376 queried epochs and 228,096,000 conservatively charged packet visits. There are no trajectory-enclosure calls or aircraft fits.
Freeze four one-thread workers, a one-hour cooperative study ceiling and a 30-minute per-identity ceiling before launch. Do not overlap this allocation with the current study. No retries, process signals or allowance expansion.
The exact saved-pair audit will check source/configuration bindings, every physical bound, actual training errors and future separation. Verify artifact hashes and native execution separately. Reevaluate each selected candidate once within its fixed verification allowance.
Interpretation
A separation beyond 20 m north, 20 m east or 50 m height establishes a lower width for the represented native calculation only after both training matches are checked. A physical impossibility claim also needs a qualified numerical error bound with sufficient margin. The saved-pair audit does not supply that bound or an aircraft error law.
Finding no such pair proves neither uniqueness nor useful outer uncertainty. If real ambiguity is established, the next comparison should add predeclared training information rather than tighten calibration assumptions or widen the targets. Neither outcome changes the 53-stretch findings or authorizes aircraft prediction fits.