Replication Data for: Prioritizing Human Review in ESG Claim Verification: A Retrospective Multi-Verifier Benchmark

Title: Prioritizing Human Review in ESG Claim Verification: A Retrospective Multi-Verifier Benchmark Authors: Adam Wasilewski, Marek Wasilewski 1. Dataset SummaryThis dataset contains replication files and archived categorical judgments from the study on automated ESG claim verification and human review allocation under limited capacity. The dataset includes 2,765 eligible verification items derived from 553 source claim families across five 2024 corporate sustainability reports: Alibaba, Amazon, Copart, Meituan, and MercadoLibre. Five historical LLM-based verification systems (GPT-5.2, Kimi K-2, Claude Sonnet 4.5, Gemini Pro 3.0, and Gemini Flash) evaluated each item five separate times, yielding up to 25 recorded verdicts per item. 2. Directory Structure and File OrganizationPlaintext├───raw │ ├───claims │ └───raw_data │ ├───Alibaba │ │ ├───claims │ │ └───results │ │ ├───MAJOR │ │ ├───MINOR │ │ ├───PARA │ │ ├───PLAUS │ │ └───SOURCE │ ├───Amazon │ │ ├───... │ ├───Copart │ │ ├───... │ ├───Meituan │ │ ├───... │ └───MercadoLibre │ ├───... └───results ├───boundary_h005 └───hellinger_expanded_grid Raw Data Directory (raw/)raw/claims/: Contains the master claim manifest with metadata for all generated source claims and their variants. raw/raw_data/{Company}/claims/: Stores company-specific source claims along with page numbers and evidence spans from the corresponding PDF reports. raw/raw_data/{Company}/results/: Holds categorical verification outputs grouped by claim variant type: SOURCE/: Verdicts for original numerical claims extracted from the sustainability reports. PARA/: Verdicts for paraphrased claims that preserve the exact original meaning and figures. MINOR/: Verdicts for claims with minor factual changes such as modified values, dates, or units. MAJOR/: Verdicts for claims with contextual reassignments such as incorrect ESG topics or reporting scopes. PLAUS/: Verdicts for plausible numerical claims that lack supporting evidence in the designated report. Results Directory (results/)results/boundary_h005/: Experimental outputs and post-hoc diagnostics using a fixed bandwidth parameter of $h = 0.05$ in the Hellinger neighborhood model. results/hellinger_expanded_grid/: Evaluation results for the Hellinger model across an expanded bandwidth tuning grid of $h \in \{0.025, 0.05, 0.1, 0.2, 0.4, 0.8\}$. (2026-09-29)