A, B, C or D? I trained all four

Keep the encoder, history length, data and training allowance fixed, and change only the objective. A reconstructs sensors, B predicts the state residual a short horizon ahead, C separates flights, and D compresses the latent. Each objective rewards a different kind of information.

Freeze the trained encoders and probe the thrust-to-mass ratio on held-out vehicles; translational motion cannot tell a heavier drone from a weaker motor. Then change the flight trajectory to see whether a model learned dynamics or a trajectory signature. Shuffle latents between flights to test whether the predictor uses the representation at all.

The resulting table must include all four answers and all training seeds. A high B score would support the course hypothesis for this dataset; a strong A or C score deserves investigation. The written companion defines the experiment and does not claim unperformed training results.