Order-Dependent Fragility in Single-Layer AI: Toward a Learning-Acceptance Controller Inspired by Complementary Learning Systems

Purpose. Large language models entangle three functionally distinct kinds of learning — stable factual knowledge, knowledge retrieval/prediction, and short-term dialogue adaptation — within a single weight space. Prior work has shown that training-data order can causally affect catastrophic forgetting, particularly in smaller models, and that learning-rate schedules interact with curriculum ordering in ways not yet fully understood. Architectures separating stable and fast-changing computation have recently shown empirical promise for continual learning generally, but no existing framework explains why single-layer architectures are structurally vulnerable specifically to training-data order, nor proposes a controller abstraction general enough to be argued for on structural grounds rather than validated only empirically for one architecture. Conclusion. We argue that order-dependent fragility is a structural consequence of collapsing knowledge, prediction, and adaptation into one weight space, not an incidental bug. Drawing on complementary learning systems (CLS) theory of hippocampus–cortex interaction, we propose a three-layer architecture — stable knowledge, knowledge prediction, and fast dialogue adaptation — with explicitly graded plasticity, and reinterpret existing findings on model-size-dependent ordering sensitivity as indirect support. Evidence that one sensory modality's maturation gates another's critical period further suggests order effects arise within the knowledge layer, not only between layers. We argue these findings jointly motivate something more fundamental than layer separation: an explicit learning-acceptance controller, logically prior to any content layer, governing whether, where, and at what rate incoming signal may modify the model — without which layer separation only relocates the problem. We close by formalizing a predictive-coding-style objective for the dialogue layer and specifying falsifiable predictions and a minimal experimental design for future validation. This is a companion paper to "From Retrospective Decoding to Prospective Logging: Traceability as a Byproduct of Gated Consolidation" (https://doi.org/10.5281/zenodo.22031993).