AI Proctoring and Grading Ecosystem Passes Real-World Exam Test

Computer-based testing has quietly become the backbone of modern examination halls, but the shift from paper to screens has exposed a stubborn problem: the tools that deliver exams, watch for cheating, grade answers, and analyse results usually live in separate systems that barely talk to each other. A new study published in Discover Education now reports one of the most complete real-world evaluations of what happens when those functions are fused into a single platform. The system, called Eduvate CBT, was developed and tested by Peace Awofesobi of Osun State University in Nigeria and independent researcher Obiseye Obiyemi, and its four-week deployment across 250 undergraduate examination sessions offers a rare quantitative window into an integrated AI-powered assessment ecosystem.

The motivation is straightforward. Since the COVID-19 pandemic accelerated online and distance learning, institutions have wrestled with examination integrity, candidate authentication, and the credibility of remotely administered tests. Weak oversight creates openings for impersonation, unauthorised assistance, and prohibited devices. Meanwhile, automated grading promises speed and consistency but raises questions about validity, explainability, and where human academic judgement must remain in charge. The authors argue that the educational AI literature has typically treated monitoring, scoring, feedback, and analytics as distinct application areas, and that the real innovation worth testing is not any single component but the coordinated workflow that binds them together.

Eduvate CBT is built on a four-layer architecture: a presentation layer with role-specific dashboards for administrators, instructors, and students; an application layer handling examination scheduling, question management, session control, proctoring, grading, and analytics; a data and intelligence layer combining relational databases with computer vision-based proctoring and language model-based evaluation modules; and an infrastructure layer running on PHP 8.x and MySQL/MariaDB with HTTPS/TLS encryption. The technology choices were deliberately conventional, the authors explain, so that resource-limited institutions could deploy the system without specialised infrastructure. Role-based access control ensures that administrators manage configurations, instructors handle examination content, and students interact only with restricted assessment interfaces.

The proctoring engine is the platform’s most technically ambitious element. It captures real-time webcam video and applies computer vision to monitor facial presence, head movement, and eye gaze direction. Environmental signals such as background noise and system activity patterns are analysed for external interference, while mouth-movement detection flags possible unauthorised communication. Across the 250 sessions, the subsystem logged 1,847 behavioural events, with tab switching accounting for 33.1 percent, window-focus loss 26.4 percent, camera-frame absence 20.6 percent, and gaze deviation 14.5 percent. Crucially, the researchers validated performance at the session level rather than treating each event as an independent classification case, a distinction that matters enormously for interpreting the statistics honestly.

Three instructors independently reviewed recorded session evidence and reached consensus on whether each session contained a violation, providing the ground truth for the confusion matrix. The results were striking: session-level accuracy of 94.8 percent (95 percent CI 91.3 to 96.9), precision of 94.6 percent, recall of 96.6 percent, specificity of 92.3 percent, and an F1-score of 95.6 percent. Only 13 of 250 sessions were misclassified, comprising 8 false positives and 5 false negatives. The high recall is particularly important in examination contexts, where a missed violation undermines assessment credibility. Yet the authors are careful to note that the validation dataset did not preserve which behavioural event categories drove the errors, and that automated flags must be interpreted within transparent, fair, and privacy-conscious review processes because normal behaviour can sometimes be misclassified as suspicious.

Automated grading delivered equally impressive numbers. The platform processed 7,500 responses across the cohort, and in a randomly selected validation sample of 750 responses, 742 automated grades agreed exactly with instructor-consensus scores, corresponding to 98.9 percent agreement (95 percent CI 97.9 to 99.5). Eight discrepancies were identified but not categorised by cause, and the authors treat them simply as cases requiring human review. The efficiency gains were dramatic: the system completed grading in approximately 9 seconds, compared with an estimated 23.1 hours of manual grading staff time for the same 250 examinations, an operational reduction of roughly 9,240-fold. The authors stress that this is a staff-time benchmark rather than a monetary saving, and that actual savings will vary with institutional workflows and assessment formats.

Beyond raw performance, the analytics dashboard demonstrated a closed-loop function that connects assessment evidence to instructional action. The system identified six examination items with difficulty indices below 0.30, concentrated in algorithmic complexity notation within a Data Structures and Algorithms course. Rather than limiting outcomes to score reporting, the platform enabled instructors to pinpoint specific knowledge areas where learners struggled. The authors frame this within Assessment for Learning principles: the dashboard surfaces patterns, but instructors must determine whether the cause is conceptual difficulty, ambiguous wording, pacing, or a mismatch between instruction and assessment before choosing an intervention such as reteaching or item revision.

Usability results were equally strong. Fifteen instructors and five administrators completed the System Usability Scale after extensive interaction with the platform during examination setup, administration, monitoring, grading, and result management. The aggregate SUS score was 87.4 (SD 5.8, 95 percent CI 84.7 to 90.1), placing it within the commonly interpreted excellent range. The relatively narrow confidence interval indicates the high mean was not driven by extreme ratings. However, the study reports only the aggregate result rather than item-level response profiles, so it cannot identify which individual usability statements were rated highest or lowest, a gap the authors acknowledge for future refinement.

The security architecture is documented as a defence-in-depth design rather than a certified guarantee. Cross-site scripting prevention and Content Security Policy headers protect the browser layer, PDO prepared statements with parameter binding mitigate SQL injection, session-based CSRF tokens validate state-changing requests, and IP whitelisting restricts access during high-stakes examinations. Examination-specific safeguards include exam window enforcement, session integrity validation, and question randomization. The authors explicitly note that no independent penetration test or external security certification was performed, and that the controls should be understood as an architectural description rather than proof of resistance to all possible attacks.

The study’s limitations are candidly acknowledged. The evaluation was conducted at a single tertiary institution within computing-related courses, creating possible sampling and context bias. No competing proctoring or grading platform was included as a control condition, so the results demonstrate within-system performance rather than comparative superiority. Individual rater labels were not retained after consensus was established, precluding formal inter-rater reliability analysis for both proctoring and grading validation. The SUS result relied on self-report from only 20 staff users. The authors also emphasise that institutional deployment requires explicit governance for privacy notice, lawful basis for processing, retention periods, access to evidence, deletion procedures, appeals, and reasonable accommodations, particularly under frameworks such as the Nigeria Data Protection Act 2023. Future work should prioritise multi-institution replication, comparative evaluation, formal rater-reliability analysis, event-specific error analysis, independent security testing, and longitudinal educational-outcome analysis before broader claims of generalisability can be made.

Subject of Research: Educational technology

Article Title: Development and evaluation of eduvate CBT as an integrated artificial intelligence powered assessment ecosystem

Article References: Awofesobi, P., & Obiyemi, O. (2026). Development and evaluation of eduvate CBT as an integrated artificial intelligence powered assessment ecosystem. Discover Education, 5(1), Article 1197. https://doi.org/10.1007/s44217-026-02317-9

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02317-9

Keywords: computer-based testing, AI-assisted proctoring, automated grading, examination integrity, learning analytics, educational technology, assessment ecosystem, System Usability Scale, role-based access control, responsible AI, Development, evaluation