Auditable Empirical Evaluation of Machine Unlearning: Immutable Deletion Manifests, Exact-Retraining Ensembles, and a Reproducible Fashion-MNIST Pilot

Pilot and methodology preprint, version 1.1 (archival manuscript). This manuscript describes an auditable machine-unlearning evaluation workflow based on immutable deletion manifests and an ensemble of independently exact-retrained models. It reports a reduced Fashion-MNIST pilot with one frozen 1% IID deletion manifest, one original model, and 15 exact-retraining seeds. Its predefined seed-count diagnostic passed, while retrospective analysis classified the deletion assay as insensitive under the two evaluated output metrics. It does not establish successful or certified unlearning, a confirmatory method ranking, privacy equivalence, cross-dataset generalization, or independent replication. Archival status, October 21, 2026: This deposit contains the preserved 24-page manuscript and two-page supplement. The original R001 evidence archive, original raw records, timestamps, and wall-clock timings have not been recovered. The repository contains later source edits that are not incorporated into these preserved PDFs. Recovery update, October 21, 2026: Two historical processed JSON artifacts were recovered from archived logs. Separately identified reconstruction experiments recreated the missing analyses; 49 checkpoints matched archived hashes. Re-exporting fixed-checkpoint predictions under the original research account and ROCm environment reproduced the historical calibration values and decisions exactly. Reconstructed LaTeX produced identical text on all 24 pages and pixel-identical renders at 100 DPI. These reconstructions are not recovery of the original raw training records or independent replication by another research group. The historical SCRUB checkpoint is archival only: its setup fails the current method-validity check and is excluded from the successful modern method comparison. The reconstruction package is separate from the two PDF files in this deposit. Public repository: https://github.com/theworker02/unlearning-experiment The manuscript remains a preliminary pilot and methodology report, not a peer-reviewed publication or completed confirmatory study. No prospective registration is claimed for the retrospective P001 pilot.