A preprint proposes ARISMA, a framework for logging and validating AI in evidence reviews while keeping consequential decisions with human reviewers.