CL-DriftGuard pre-registration bundle (protocol, addenda 1–3, code, data hashes)

This record archives the pre-registration, source code, tests and stored results of a reconstruction and audit of the continual-learning framework described by Muthumanickam et al. (ICICIT 2026, DOI 10.1109/ICICIT69063.2026.11633941). It supports the manuscript "Does Detector-Triggered Experience Replay Help a Streaming Linear Learner Under Drift? A Pre-Registered Reconstruction and Audit of a Continual Learning Framework for IoT" (in preparation). Contents PREREG.md: the original registration (dated 2026-11-02) with decision rules for the research questions, and Addenda 1–3 for the follow-up Studies D (recent-sample replay), E (recurring drift) and F (matched-update control, tuning, oracle trigger, label delay, adaptive-tree baseline, real recurring stream). Each addendum was written before its runs. default.yaml: the frozen configuration (config hash 591c91c4bdc2df1e). src/cldrift/: Python source for the learner, reservoir memory, drift detectors (ADWIN, DDM, EDDM), metrics, evaluation and report generation. scripts/: one runner per study, figure generation and integrity-check scripts. tests/: 40 unit, integration and replication tests. outputs/: per-run results, detector alarm events, run manifests and the generated replication report for Studies A–F. Design.md, README.md, REPLICATION.md, DATA_README.md: design record and instructions. MANIFEST.sha256: SHA-256 checksum of every file in the archive. What is not included The raw public datasets are not redistributed. DATA_README.md gives their sources and file hashes. The manuscript is not part of this record. Timing and limits of the timestamp The pre-registration was written before the main experiments, and each addendum before its own runs, but this deposit was made after all experiments had been run. The deposit date therefore shows that these exact files existed by that date. It does not independently prove the order in which they were written. The registration and addenda were kept as files under the author's control. Reproducing See REPLICATION.md. Software: Python 3.14, river 0.26.1, scikit-learn 1.9.0, NumPy 2.5.0, SciPy 1.18.0. SHA-256 of the archive as built: 0c1083818f99736c3dcd8b3037a37df90158999ea4fcbf756692e0a9a8472cf4. Scope This is an independent reconstruction. The anchor paper does not specify its datasets, detector, replay algorithm or model, so the code follows the paper's description with documented choices. It is not the original authors' system.