Generative AI has made the finished product an unreliable witness to learning, and the sector’s main response, sorting assessment into secured and open lanes, concentrates assurance in invigilation. This poster starts somewhere else. A grade is not a measurement or a verdict on conduct, but a warranted claim about the evidence a task required and received (Kane, 2013; Toulmin, 1958). On this view, most AI misuse is a quality failure: the work does not demonstrate the learning the task was designed to assess, and it is graded accordingly. Misconduct narrows to a small set of cases: fabrication of data, sources, or experiences, and circumventing secured conditions.

The framework redirects the hard-won vigilance educators have built over years of detecting misconduct, from the compromised product towards the evidence a task required and received. A decision procedure separates quality from misconduct at the point of marking, trialled in a first-year critical thinking subject where warrant is read from process evidence (Fawns, Boud & Dawson, 2026). Early experience suggests the misconduct caseload contracts to dishonest cases and administrative effort shifts from investigation towards defensible grading.