Research · August 2026
The Unarchived State
Measuring the preservation gap of Cuba's digital press. Fourteen systems audited; none stores versions of what the Cuban state publishes.
Summary
The obvious objection to this project is that "the Wayback Machine already does this." It does not. We measured its coverage directly against the CDX API rather than trusting the impression, and found a specific gap: not crawl frequency, but version storage.
Three reasons crawling harder won't close the gap
Latency outruns the edit. The first snapshot lands, at median, a day after publication. Same-day silent edits — the interesting case — happen before any copy exists to compare against.
A changed fingerprint is not a changed article. Pages carry rotating sidebars and ad slots: the hash moves without the body moving. Any system comparing whole pages produces almost only false positives. Extracting and comparing the article body alone is the real engineering core.
Deletion is silence, and silence isn't recorded. Web archives store what they managed to fetch; they never assert "this URL now returns 404." Nothing in the pipeline watches known pages for disappearance.
The nearest misses
| System | Versions? | Gap |
|---|---|---|
| Wayback Machine | Incidental (2–13%) | Thin coverage; no body diffs; no deletion detection |
| ArchiveTeam NewsGrabber | No — dedupes by URL | Covered nearly the same outlets; died in 2017 |
| Media Cloud | No — one row per URL | Continuous ingest, but never re-fetches to compare |
| NewsDiffs / diffengine | Yes | The tool exists; nobody aimed it at Cuba |
| UM collection (Archive-It) | No — front page only | One monthly homepage capture; 30-day blind spot |
Conclusion. Everything upstream of versioning has been built by someone: the crawl targets, the ingest pattern, the diff engine, the storage. Nobody has assembled them for the Cuban state press, and the project that came closest died in 2017 and locked its output.
Methodology, full tables and caveats: extended version in preparation. Figures come from own queries to the Wayback CDX API (1 Jul – 13 Aug 2026), deduplicated by URL and content digest.