Dual-Memory Dissociation Under Partial Observability: An Identifiability-Governed Phase Transition

Abstract

AbstractWe study whether two latent memory channels, an E-channel driven by fast, high-contrast binary growth-rate switches and a B-channel driven by slow capacity drift, produce a statistically detectable double dissociation in lifecycle model output, and how this detection degrades as oracle feature quality is reduced. Using a controlled simulation framework with three observation worlds, World C, World Hybrid, and World A2, we show four main findings: (1) asymmetric channel damage under oracle-feature degradation, (2) the necessity of per-channel recoverability diagnostics, (3) a recoverability-governed transition zone under true partial observability with decoder robustness across tested models, and (4) an empirical non-recoverability baseline in World Hybrid under tested level-based decoders. All results are validated by a skeptic suite including label shuffle, permuted retrain, feature permutation, column swap, and exact recompute. This record is a preprint version of a manuscript submitted for peer review. This preprint corresponds to a manuscript submitted for peer review. The public Zenodo version is provided for open access, citation, and timestamping purposes.