Latent terrain explores the space between representations produced by a fixed language model. The source model and its recorded trajectories remain unchanged.
A separate network is intended to predict logits at synthetic points between those trajectories. At recorded points, its output should agree with the original model. Away from them, the work examines how neighboring regions can fit together consistently.
The question is whether this creates a useful representation for exploring or decoding between known states. The specification sets out constraints for that experiment; it does not establish that the interpolated space has a reliable meaning.