Benchmark and Audit Artefacts for Auditable Concept Harmonisation in Bibliometric Analysis (Version 1.0.0)

Benchmark and audit artefacts accompanying the manuscript "Auditable Concept Harmonisation in Bibliometric Analysis: Benchmarking an LLM-DAG Workflow". This record contains files that include Scopus-derived author keyword strings and are therefore subject to restricted access under Elsevier Terms of Use. Access requests are reviewed individually. Contents: benchmark/ gold_benchmark.csv — 500-pair adjudicated gold-standard benchmark with labels and strata dev_set.csv — development set (351 pairs, seed=42) test_set.csv — held-out test set (149 pairs) annotation_sheet_annotator1_COMPLETED.csv — completed annotations, Annotator 1 annotation_sheet_annotator2_COMPLETED.csv — completed annotations, Annotator 2 adjudication_sheet.csv — adjudicator decisions with rationale for 57 disagreement pairs agreement_statistics.csv — inter-annotator agreement statistics (Cohen's kappa = 0.81) pilot_annotator1_COMPLETED.csv — pilot round labels, Annotator 1 pilot_annotator2_COMPLETED.csv — pilot round labels, Annotator 2 audit/ test_pair_decisions_full.csv — full LLM-DAG decisions for all 149 test pairs test_pair_justifications.jsonl — verbatim model justifications for all 149 test pairs error_analysis.csv — per-pair error category classification guard_decisions.jsonl — guard layer decision log for benchmark run ablation_A2_raw_outputs.jsonl — raw LLM outputs for A2 ablation (forced binary) ablation_A4_raw_outputs.jsonl — raw LLM outputs for A4 ablation (simplified prompt) b6_test_raw_outputs.jsonl — raw LLM outputs for B6 Naive LLM baseline, test set prompt_hash_registry.csv — SHA-256 hashes of all prompt versions used mapping/ [EMBARGOED — pending data quality fix] cluster_membership.csv — raw keyword to canonical label to cluster ID accepted_match_edges.csv — pairwise decisions with full provenance metadata canonical_label_registry.csv — cluster labels and selection trace data/ author_keyword_frequencies.csv — frequency count per unique author keyword string corpus_snapshot_metadata.json — corpus retrieval metadata (no raw records) Related publication: https://doi.org/[add manuscript DOI when available]Software record: https://doi.org/[add Zenodo-SW DOI after creating Record 2]GitHub repository: https://github.com/MJabdelilah93/llm-dag-keyword-harmonisation