Teacher-in-the-Loop AI Assessment for African Classrooms: Paper-Based Grading, Offline Exams, and Code-Switched Oral Verification

AI-assisted assessment in African classrooms must account for handwritten examinations, unreliable connectivity, and students who explain their reasoning across languages. We examine how assessment systems can address these conditions while keeping decisions about student grades under lecturer oversight. We present a teacher-in-the-loop research prototype that combines paper-script grading, local exam generation, and code-switched oral verification. In a shadow-mode evaluation of 60 introductory Python scripts, mean absolute grading error against lecturer-derived reference scores was 4.39 marks out of 60 without lecture context and 5.20 with it. Although error was higher with context in this sample, uncertainty in the difference prevents a firm conclusion about its broader effect. The workflow enforced final score limits and flagged 42 or 43 scripts for lecturer review, depending on the condition, indicating substantial reliance on human oversight. Speech-recognition comparisons on 11 classroom recordings and 30 separate benchmark clips showed that low average transcription error could coexist with missing transcripts, and that provider rankings differed between settings. Exploratory local-execution observations further highlighted the latency and recovery requirements of running assessment tools on teacher hardware. These findings identify practical requirements for teacher-supported assessment rather than establish readiness for autonomous grading. We argue that assessment AI should be evaluated not only by agreement with reference scores, but also by how it applies marking policies, handles transcription failures, supports teacher review, and operates under classroom constraints.