Data Deletion Propagation — References

The obligations that create the requirement, the sanitisation standards that define what removal means, and the index mechanics that make it slow.

Obligations

• Regulation (EU) 2016/679 (GDPR) — Article 17 on erasure, Article 19 on notifying recipients of a rectification or erasure, and Article 12's one-month response period, which together define scope, reach and deadline.
• California Consumer Privacy Act, as amended by the CPRA — the deletion right and its obligation to direct service providers and contractors to delete, which is the derivative-reach requirement in a different jurisdiction.
• ISO/IEC 27001: Information security management systems — Requirements — the control objectives for secure disposal that an evidence record is usually assessed against. Cited by name.

What removal means technically

• NIST SP 800-88 Rev. 1: Guidelines for Media Sanitization — the clear, purge and destroy distinction, and the argument that a logical delete is not sanitisation. The basis for treating crypto-shredding as a defensible option.
• NIST SP 800-57 Part 1: Recommendation for Key Management — key lifecycle and destruction, which is what makes key-destruction-based erasure meaningful rather than nominal.

Why an index does not forget quickly

• Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs (Malkov & Yashunin, 2016), arXiv:1603.09320 — the graph construction that makes node removal non-local, and therefore the reason deletions are marked and deferred to a rebuild.
• Apache Lucene — the segment and merge model, and the documented behaviour that deleted documents persist in segments until a merge reclaims them.

What a vector still contains

• Information Leakage in Embedding Models — Congzheng Song and Ananth Raghunathan, ACM CCS, 2020. Demonstrations that substantial information about the input text can be recovered from its embedding, which is the reason a retained vector is a retained copy rather than a derived statistic. Cited by name.
• Text Embeddings Reveal (Almost) As Much As Text — John X. Morris, Volodymyr Kuleshov, Vitaly Shmatikov and Alexander M. Rush, EMNLP, 2023. Reconstruction of original text from dense embeddings, strengthening the same conclusion. Cited by name.

No verified reference is recorded for derivation ledgers across an ingestion pipeline, for regeneration of aggregates on deletion, or for post-restore reconciliation procedures. The standards define what removal means and leave the enumeration of derivatives entirely to implementers.