Five training epochs leave substantial freedom
The two completed long-record cases have two separate limitations:
- Their approximation bounds grow too wide during integration.
- Five training epochs leave considerable calibration freedom in a simplified linear calculation, even before those approximation errors are included.
The saved linear check finds feasible horizontal spreads of 420–1,068 m. The long-record test's fixed horizontal width target is 20 m. Each saved point satisfies the linear training constraints exactly, within the original physical parameter box.
These are generated-data diagnostics, not aircraft errors. The linear points have not been checked in the nonlinear model. They do not prove the physical target impossible, but they show why improving one remainder is not enough evidence to promise useful prediction ranges.
The next design should compare training coverage using the intended earlier record, rather than rely on five sparse epochs. The earlier prediction plan used 292 training epochs. Any new calculation needs its own finite freeze; the running study keeps its original five epochs and targets.