Replication materials — "Evaluation Without Ground Truth: Corrigible Bias and Incorrigible Erraticism in LLM-Based Decision Support"

Replication package for a study of five-model LLM panel scoring of the qualitative sections of European banks' Pillar 3 disclosures (EBA Pillar 3 Data Hub, reference date 31 December 2025; 266 institutions, 29 countries, 18 languages; 1,121 bank–section pairs; 5,605 scores). The package contains the scoring prompts with anchored rubrics and their revision-stage perturbations, the extraction and scoring code with per-call telemetry, all scores, the extracted section texts, the repeat and robustness runs, and an independent extraction audit. Source documents are public regulatory filings and are not redistributed; the corpus manifest lists every institution and retrieval channel. See README.md inside the archive for the full structure.