The remaining allowance grows with position propagation

All 48 duration identities pass, covering every model/processing case and all eight original orientation sectors at 1.8, 3.6 and 7.2 seconds. Each prefix preserves native values and all 27 physical tangent columns, one global bound, and three parameter probes. All checked outputs stay within their factored bounds; native centers satisfy the unchanged one-micrometre limit.

The largest unshared width contributions grow as follows:

Almost all this contribution comes from position propagation. At 7.2 seconds, its largest per-axis radius is about 75 m, whereas the dynamic lever-arm radius is at most 9.3 cm. These are conservative terms in the mathematical calculation, not observed aircraft errors, instrument drift or minimum physical uncertainty. No training-feasible region is computed here, so these are not complete future prediction widths either.

The screen completed in 752 seconds, using 576 charged native calls and 144 bound calls. Its first 400 input packets reproduce the completed short study; models and calibration bounds remain unchanged. Results and source/input hashes are verified. No original recordings, observed fits or protected ESA outcomes were examined.

The next candidate preserves a common reference acceleration instead of recounting its nonlinear calibration error at every integration step. Every departure from that reference, including aircraft motion, changing metric, field and transport terms, must remain. This is an algebraic refinement, not an assumption of constant aircraft acceleration or permission to relax calibration bounds. Component, native, joint-interface and usefulness gates remain separate.

The 53-stretch empirical result is unchanged. Long-record prediction uncertainty, whole-program numerical treatment and actual receiver-error adequacy remain open.

Complete results · Frozen scope · Short improvement · Roadmap.