Design notes

Retina is the perception encoder of a world model. In the standard decomposition (encoder → dynamics → decoder), Retina is the encoder — s = Enc(x): raw real-world signals → state. It produces state, not the dynamics or the policy; we say "encoder," not "world model."

The full rationale — the dual symbolic + latent output, where Retina sits in the world-model stack, the deep-vs-wide L1 axes, what it absorbed from DeepStream / Holoscan / academia, and the hard layer-boundary rules — is the canonical DESIGN.md in the repository, kept there as the single source of truth so this site never drifts.