Source data, raw run records and code for: Self-organizing memory allocation recovers task structure and replaces a pretrained episode prior in continual vision

Supporting material for the manuscript "Self-organizing memory allocation recovers task structure and replaces a pretrained episode prior in continual vision". The record contains every value plotted in the main and supplementary figures as a machine-readable Source Data workbook; the unaggregated per-run records from which those values were compiled, including complete task-by-task accuracy matrices, configurations, environment and code hashes for every formal run, control and baseline; and a snapshot of the model implementation, the shared metric implementation used to recompute every reported accuracy, forgetting and backward-transfer value, and the scripts that regenerate every figure and table. Statistical units differ by experiment: n = 5 independent seeds for the primary CIFAR-100 and Tiny ImageNet results, n = 3 for most isolated interventions, and n = 10 official class orders at a fixed training seed for the CORe50 transfer. Model checkpoints and the frozen CLIP embedding arrays are not included; the README explains why neither is needed to reproduce any reported number, and how the embedding arrays are regenerated and hash-checked from pinned sources. Raw run directories and configuration keys retain the internal project identifier NCOM. NCOM and SemaSNN-CL denote the same system.