Predicting the later flight

49/72 returned fits passed their numerical stopping checks. The frozen definition contains 72 starts.

Each model calibrates on earlier receiver positions and IMU increments, then predicts later positions from the continuous IMU stream without a GPS reset. Later GPS positions only score the saved forecast. Height counts alongside horizontal error. Start selection uses converged training cost alone.

These are retrospective development results. Accuracy screens are descriptive; neither a small error nor a relative advantage establishes calibrated significance.

The run charged 3,342 native calculations. Unfinished and failed starts remain in the sealed evidence, including alternatives to each displayed primary. Differences between starts describe sensitivity, not a confidence interval. A smaller horizontal error cannot excuse an implausible height reconstruction.

Selection and splits · Scientific plan · Packaging correction · Prediction protocol