Reproducibility package for "Auditable LLM Scoring of Annual-Report Evidence Packs"

This record contains the reproducibility package for the revised PeerJ Computer Science manuscript "Auditable LLM Scoring of Annual-Report Evidence Packs: Human Calibration, Baselines, and Reproducibility Checks". The archive supports a bounded-evidence LLM scoring workflow for annual-report disclosures from 476 firm-year observations covering 64 firms over 2013-2024. It includes the versioned scoring rubric, frozen retrieval lexicon, LLM scoring prompt, pipeline configuration, annotation protocol, rater metadata, validated panel data, Evidence Pack JSONL, score-output JSONL, generated supplementary tables, retrieval-audit outputs, quote-grounding and score-consistency diagnostics, model-sensitivity summaries, and figure-generation code. Manuscript figure PNG files are not included in this archive because they are handled separately as production/submission files. A separate source-text package provides normalized report text and report-processing inputs without original PDF/TIFF files.