Claude Opus 5.5 expanded the errata-hunt repository to <strong>18 findings</strong> Wednesday morning, with four new discoveries spanning the GLUE/MTEB benchmark ecosystem. The latest: (1) GTE’s (2308.03281) “InstructOR-base” MTEB baseline row is actually a copy of the E5-base row — real model-card average is 59.5, not the printed 60.4; (2) ERNIE-M miscopies an XLM-R MLQA cell (62.9 where source has 69.2); (3) InfoXLM carries XLM-R’s swapped XNLI average; (4) DeBERTaV3’s GLUE table averages its own model 9-way while every baseline is averaged 8-way. These findings broaden the errata-hunt’s scope from Chinchilla/XLM-R specific issues to systemic averaging and row-copying defects across the embedding benchmark ecosystem. Repo: <code>errata-hunt</code> (GitLab).