EVIDENCE-Prompt: An Auditable Secondary-Data Framework for Scientific Inquiry with Generative AI in Higher Education

Abstract

Generative artificial intelligence is increasingly used for questioning, data interpretation, experimental planning, and explanation in higher education, but most prompt frameworks optimise fluency rather than the integrity of scientific reasoning. This study develops EVIDENCE-Prompt, an auditable framework that treats a prompt as an information-governance object linking a learner’s question, admissible evidence, model output, and human verification. A secondary-data design re-analysed 25 operational studies retained from a 2025 systematic review and added seven post-seed records, yielding 32 studies across biological, medical, engineering, computing, social-science, language, and cross-disciplinary settings. Each record was coded for eight components: Epistemic goal; Variables and context; Inquiry phase; Data and evidence; Explicit constraints; Neutrality and alternatives; Citation and verification; and Error checking and reflection. Descriptive prevalence estimates used Wilson 95% confidence intervals; cross-domain contrasts used Fisher’s exact tests with Holm adjustment. Epistemic goals and contextual specification appeared in all records. Explicit constraints and reflective checking each appeared in 29/32 studies (90.6%), whereas inquiry-phase alignment and evidence boundaries appeared in 20/32 (62.5%), neutrality and alternatives in 18/32 (56.3%), and citation verification in only 9/32 (28.1%). Median framework coverage was 6 of 8 components (interquartile range 5–7), and only four records operationalised all components. STEM, health, and engineering studies more frequently bounded data and operationalised citation verification than other domains, although adjusted subgroup tests were exploratory and non-significant. EVIDENCE-Prompt converts these gaps into a reproducible prompt schema, scoring rubric, and governance workflow. The accompanying CSV dataset and codebook support independent audit, adaptation, and future model benchmarking.