STATE-AWARE PROSPECTIVE RECOVERY AFTER NEURAL TRAINING DAMAGE

Neural training can enter damaged states with several admissible recovery inter- ventions, yet one action must be chosen before its future consequences are known. The central question is whether current task behavior contains enough information for that choice. In controlled digits AdamW states, 79 of 80 confirmatory matched pairs reverse the preferred retain/reset action after exact parameter restoration, showing that matched current behavior can conceal different recovery-action val- ues. Self-Dynamics Network (SDN) is the branch-supervised recovery-decision framework used to study this problem. On held-out AdamW cohorts, state-aware scorers outperform a loss-only reference in binary retain/reset control, while state- conditioned policies have lower final-cohort cost than the lowest-cost fixed action in a four-action setting. Classical and neural scorers perform nearly identically, so these results do not support neural function-class superiority. A symmetry- compatible control identifies optimizer non-equivariance as one controlled route to divergent futures, while a separate seven-action Transformer teacher extends the decision principle under explicit future rollouts. These results support recovery- relevant training state as a source of decision value in the tested settings. Natural prevalence, large-scale competing actuation, and operational cost benefits remain open.