Equal-Byte Non-Inferiority Protocol for Storage-Aware Evaluation of Replay-Based Continual Learning: code, data, and results

This is the reproducibility artifact for the paper "An Equal-Byte Non-Inferiority Protocol for Storage-Aware, Reproducible Evaluation of Replay-Based Continual Learning" (Machine Learning and Knowledge Extraction, 2026). Replay-based continual learners are conventionally compared at a fixed number of stored exemplars, but deployment hardware meters memory in bytes, and methods that store logits, feature vectors, or compressed codes occupy different byte footprints at the same item count. The equal-byte protocol holds total stored bytes constant across competitors, selects the strongest of a pre-registered eight-method baseline pool (GDumb, GSS, class-mean herding, reservoir sampling, DER++, ER-ACE, MIR, Rainbow Memory) at each budget, and returns a non-inferiority verdict at a registered margin of 0.05 on Average Accuracy under Holm-Bonferroni correction across budgets. It is demonstrated on Split-CIFAR-100 and an eight-task family-grouped Abstraction and Reasoning Corpus sequence using a two-tier consolidation memory that exposes a byte-budget knob, at ten seeds per cell. Under the pre-registered rule the two-tier scheme does not achieve non-inferiority at any of the four tested byte budgets on either benchmark; the outcome is robust to task ordering (six orderings) and is explained by an interpretability analysis of the compressed tier. The archive contains the cdemat reference implementation (byte accounting, models, training, statistics); the pre-registered experiment configurations; the curated 50-puzzle / 8-family ARC benchmark; the frozen per-seed result CSVs behind every table; the raw per-run logs; the analysis code that regenerates the paper's tables; the byte-accounting and consistency tests; and the two OSF pre-registration documents. Split-CIFAR-100 and Core50 are public datasets and are not redistributed here; fetch scripts and DATA_NOTES.md are included. Pre-registration (immutable OSF registrations): protocol and first amendment at https://doi.org/10.17605/OSF.IO/7XQGK ; second amendment (task-order robustness and interpretability) at https://doi.org/10.17605/OSF.IO/XND3Z. License: CC BY 4.0.