We currently only store observations (or inferred physical quantities). As most of the work will happen in the latent space, we should have the ability to introspect the content and analyze latent trajectories. Time-bounded investigation:

• can we store it in the same format as targets and inputs with zarrio? That would allow further analysis with dask using zarrio and the evaluation package
• can this be done with little changes to the code (ex: with a special decoder that performs no conversion)