Replication package: Catastrophic Forgetting under Resource Constraints — Task Separation Governs the Reliability, Not Only the Magnitude, of Memory-Based Mitigation

Complete code and raw results for a study of catastrophic forgetting in memory-constrained continual learning. Four strategies — experience replay, Online EWC, a CURATOR-style value-threshold buffer, and unmitigated sequential fine-tuning — are compared on the N-BaIoT intrusion-detection dataset across two task streams that differ only in whether successive tasks share a label space. The archive contains 1,376 run files, one per (scenario × resource level × strategy × seed). Each stores the raw per-task macro-F1 performance matrix as written by the runner before any metric was computed, so every Forgetting, Intransigence and Backward-Transfer figure in the paper can be recomputed and independently verified. Aggregation, statistical-testing and figure-generating code is included, along with 128 unit and integration tests. Figure generation is deterministic and reproduces the published figures byte-identically. The arms supporting the paper's central comparison were run at 46 seeds per condition, a depth fixed by power analysis rather than convenience; the remainder are at 5 seeds. Roughly 100 hours of CPU training are represented, none of which needs repeating to check the reported results. N-BaIoT itself is not redistributed; it remains available from the UCI Machine Learning Repository under its own terms.