Embedding assurance within learning: Empirical evidence from the SAGE framework for repositioning take-home assessment in AI-integrated higher education

This study examined whether a mature, empirically validated generative artificial intelligence (GenAI) intervention framework can produce reliable process evidence when deployed in unsupervised take-home assessments. Twenty-five group submissions from two cybersecurity management cohorts, designed under the Structured AI-Guided Education (SAGE) framework, were audited using the five-check SAGE Audit Protocol. Only 3 of 25 submissions (12%) produced evidence chains that were substantially auditable, and full traceability between documented AI outputs and human evaluation claims was not achieved in any submission. Across both cohorts, the remaining submissions showed mismatches between human-authored tables and AI outputs, generic compliance text, workflow-focused reflections, and process records that were often indistinguishable from reconstructed accounts. The paper identifies a compliance gradient in which conscientious students who follow the process in good faith incur a disproportionate documentation burden, while simulated compliance can produce comparable outputs with less effort. It also highlights a marker's dilemma, where auditing AI-supported process evidence approximately doubles marking time and shifts educators from assessing learning to interpreting logs. The paper argues that conventional unsupervised take-home assessment can no longer function as sufficient standalone assurance in the GenAI era. Rather than abandoning AI-integrated learning or defaulting to low-technology examinations, the findings support assessment architectures in which assurance is structurally embedded. The Defend step subsequently added to the SAGE framework operationalises this principle.