Direct source bound over full generated histories

All six cases pass. Each uses the same complete 1,600-packet history, sector-zero child domain and inherited parent training certificate as its saved predecessor. Predictions, coefficient matrices, shared caps and physical bindings remain unchanged.

Horizontal approximation allowances decrease by approximately 25–50%. For the two globe-with-field cases, north allowances decrease by about 34–35% and east allowances by about 25%. Height changes very little. These are mathematical approximation allowances, not measured aircraft errors or the complete prediction ranges.

The calculations finish in 63.6–113.9 seconds each, within the unchanged 120-second limit. The study uses all six bounds, 86,400 logical steps and 36 parent support checks. Its saved evidence and reservations are verified. There are no native trajectory oracle calls, aircraft reads or optimizer starts.

Next, qualify the complete training/prediction adapter with known answers, then run a separately frozen comparison covering all eight calibration regions in all six model/processing groups. Every region and the original width target must pass. The last complete result remains four of six groups.

Recording-scale ranges, whole-program floating execution bounds and observational confidence remain separate requirements.

Results · Frozen scope · Roadmap.