A four-strand composition for cumulative dynamic regret in non-stationary reinforcement learning. Single-author theory paper. Two-Gap diagnostic separating goal-feasibility from policy-quality; point-mass reverse-KL/TV identity strictly improving Pinsker and Bretagnolle–Huber at the deterministic-π* corner; strategic tempo with forgetting prerequisite as structural survival inequality; loop-as-causal-engine for learnable bounds. Cumulative dynamic regret rate
Repository follows the segmented-paper workflow: paper segments live in src/, with one or more OUT.*.md concatenation manifests assembling them into output forms (e.g. OUT.full-paper.md for the unconstrained version, OUT.neurips-2026-paper.md for the 9-page-budget submission). Currently bootstrapping the segmented layout.
Consumed as a submodule by an umbrella workspace.