The tighter range also improves the source error allowance

All six generated-source comparisons pass. Every one of the 27 coefficient columns agrees with an independent rotation calculation. The largest coefficient difference is about 2.5 × 10⁻¹⁵; the largest value difference is about 1.6 × 10⁻¹⁴ m/s².

The earlier value-only test was not enough to establish a usable propagation bound. This comparison takes the actual compiled-source reference value and coefficient row, then calculates a conservative error allowance around that row. The allowance is 0.00093–0.00097 m/s², compared with the original 0.00305 m/s²: a reduction of about 68–69% in all six groups.

The sources preserve their original training-derived information and calibration domains. All clock queries lie about 1.3 × 10⁻¹³ seconds from a packet boundary, so the saved branch convention matters. The native and independent calculations use the same selected packet. A branch-specific reference row does not establish a unique two-sided derivative at the boundary; the affine bound remains valid about the specified row across the supplied interval.

This is an improvement in the calculation, not a new aircraft result. It covers one body-rotation source term, not every force sample, geographic-frame rotation, velocity update or position prediction. Carrying the tighter vertical bound through those operations is the next task. Aircraft prediction fits remain gated, and the existing 53-stretch findings are unchanged.

Results and evidence · Next integration task · Analysis roadmap.