Making Library Metadata Traversable: A Provenance-Preserving MARC21-to-BIBFRAME Workflow for Multilingual Discovery

Multilingual academic libraries need to prepare metadata for AI-supported discovery while preserving source integrity, institutional control, and professional cataloging judgment. This pilot examines a provenance-preserving workflow using 13,676 MARC21 bibliographic records. The workflow separates unchanged source records from normalized JSONL and BIBFRAME/SKOS linked-data layers, with Wikidata alignments supplying multilingual access points. A deterministic enhancement pass replaces 28 placeholder primary titles using linked MARC 880 fields and exposes Arabic-script title access in 32 records overall, without translating or generating text. An author-conducted audit judged 193 of 200 sampled Wikidata alignments correct (96.5 percent). A 59-query silver diagnostic compared graph-expanded BM25 with lexical BM25. Mean recall, reciprocal rank, and nDCG at ten were higher with graph expansion, while the number of queries returning at least one relevant record in the top ten fell from 58 to 57 of 59. The tested graph did not contain the recovered primary titles, and the diagnostic is not an independent user evaluation. The pilot identifies source-backed script recovery, explicit snapshot provenance, and professional review of alignments as practical requirements for evaluating this metadata-preparation workflow before deploying an AI-supported discovery service.