Factorial Chunker-Embedder Retrieval Audit: Master Results and Thirteen Evaluation Corpora

This deposit accompanies an anonymous double-blind submission. It contains the full results of a factorial audit of dense bi-encoder retrievers under chunker variation on thirteen character-level evaluation resources. Contents - master_results.parquet: 8.2 million query-level retrieval rows from a factorial of 6 embedders x 6 chunking algorithms x 3 chunk-size regimes x 13 resources, evaluated at retrieval depths k in {5, 10, 20, 50, 100}. 1,404 cells, 15,174 unique queries. - benchmarks/: the thirteen evaluation corpora in a unified character-level format (queries.jsonl, gold spans, source corpus text, metadata.json). The thirteen corpora are re-formatted from public sources (LegalBench-RAG, the Chroma chunking-evaluation suite, QuAC, MASH-QA, MultiSpanQA, TechQA). Upstream licences and citations are preserved in each per-resource metadata.json. Author and institutional information are withheld during peer review.