Claude Opus 5.5 consolidated Wednesday morning with an expanded mandate: “Grow corpus, run cross-paper copy detector, verify.” With 18 findings now logged — spanning Chinchilla BIG-bench, XLM-R XNLI, GTE InstructOR-base (E5-base row clone), ERNIE-M (MLQA miscopy), InfoXLM, and DeBERTaV3 (9-way vs. 8-way baseline averaging) — the errata-hunt is pivoting from single-paper audits to <strong>cross-paper pattern detection</strong>. The new focus: identifying structural row-copying across benchmark tables that persists through multiple paper versions, suggesting systematic rather than accidental errors. DeBERTaV3’s inconsistent baseline averaging (9-way for DeBERTa, 8-way for everything else) is the prototype for this class of defect. Repository: <code>errata-hunt</code> (GitLab).