AI-EQRA: Reproducibility package for qualifying AI-derived information in risk assessment

Abstract

This repository contains the reproducibility package associated with the methodological study “Qualifying AI-derived information for risk assessment: A decision-gated method separating predictive, explanatory, and intervention claims”. The study proposes AI Evidence Qualification for Risk Assessment (AI-EQRA), a sequential method for determining which uses of AI-derived outputs are supported by the available validation evidence. The method distinguishes predictive qualification, probability qualification, explanation qualification, falsification and robustness, while keeping intervention claims outside the scope of predictive explainability. The repository includes a unified R reproducibility script, archived outputs from the 58,000-fit Monte Carlo stress test, the prospectively specified criteria used in the dual-use qualification demonstration, the SHA-256 record of that specification, and a consolidated workbook containing the prospective dual-use experiment, positive-control analysis, estimand-separation benchmark, oracle-top stability results and variable-level summaries. All study data are fully synthetic. No human participants, identifiable personal data or real organisational data are included. The package supports both a lightweight verification mode using the archived Monte Carlo outputs and a full-reproduction mode that reruns the complete computational workflow.