How models update state over time without loss. We study structure-preserving transformations that replace repeated correction with constrained evolution. The goal is learning that does not degrade with scale.

New runtimes where computation is evolving state, not instruction sequences. Programs as trajectories through configuration space, not lists of operations applied to memory.

Continuous correction embedded in the system rather than applied periodically from outside. Feedback that operates at the same timescale as the process it governs. Stability by construction.

Direct mapping between computation and real-world processes. Simulation that shares structure with what it models. The boundary between digital and physical dissolves when the primitives align.

Many small systems behaving as one coherent system. Agreement without centralization. State consistency maintained through local interactions and shared constraints.